A Review of Image-Based and AI-Driven Water Level Monitoring: Methods, Datasets, and Deployment Considerations

A Review of Image-Based and AI-Driven Water
Level Monitoring: Methods, Datasets, and
Deployment Considerations
Atif Latif 
Utah State University
Masoumeh Hashemi 
Utah State University
Matt Yost 
Utah State University
Safran Khan 
Utah State University
Systematic Review
Keywords: Water Level, Waterline detection, Computer vision, Deep learning, Classical machine learning,
image processing
Posted Date: November 12th, 2026
DOI: https://doi.org/10.21203/rs.3.rs-11693497/v1
License:   This work is licensed under a Creative Commons Attribution 4.0 International License.  
Read Full License
Additional Declarations: The authors declare no competing interests.

A Review of Image-Based and AI-Driven Water
Level Monitoring: Methods, Datasets, and
Deployment Considerations
Atif Latif 1, Masoumeh Hashemi 2*, Matt Yost 2, Safran Khan 3
1Department of Computer Science, Utah State University,Old Main
Hill,Logan, 84322, Utah, USA.
2*Plants, Soils and Climate Department, Utah State University, Old
Main Hill, Logan, 84322, Utah, USA.
3Department of Civil and Environmental Engineering and Utah Water
Research Laboratory, Utah State University, Old Main Hill, Logan,
84322, Utah, USA.
*Corresponding author(s). E-mail(s): masoumeh.hashemi@usu.edu;
Contributing authors: atif.latif@usu.edu; matt.yost@usu.edu;
safran.khan@usu.edu;
Abstract
Accurate and continuous water-level monitoring is essential fo r agriculture, flood
management, and water resources planning. However, many rivers a nd agri-
cultural canals remain ungauged because conventional sensin g infrastructure is
costly to install and maintain, limiting basin-scale water accounting and con-
servation policies such as water banking and leasing. Comput er vision enables
water-level estimation from ground-based cameras, unmanned aeri al vehicles
(UA Vs), and satellites, offering a low-cost, non-contact alt ernative to conven-
tional sensors. This review synthesizes 93 studies selected fro m 168 articles
retrieved from five databases, following PRISMA 2020 guidelines. We summa-
rize contact-based and non-contact sensing methods and exam ine image-based
approaches, including traditional image processing, classic al machine learn-
ing (ML), and deep learning (DL), alongside their imaging moda lities and
datasets. DL methods, particularly object detection, semantic segmentation,
and transformer-based models, are more robust than earlier approache s under
low light, reflections, and cluttered backgrounds, with several studies reporting
field errors of approximately 1–3 cm. However, these results are not d irectly
1

comparable because studies differ in datasets, reference measu rements, and eval-
uation metrics. Three major research gaps emerge: (1) fragmented dat asets and
inconsistent evaluation protocols; (2) limited cross-site v alidation and model
generalization; and (3) insufficient long-term field evaluatio n and robustness
to illumination changes, reflections, and camera displacemen t. Future research
should prioritize standardized, publicly available multi-si te datasets, common
evaluation protocols, cross-site and seasonal testing, and a daptive models that
maintain reliability under changing field conditions.
Keywords: Water level, Waterline detection, Computer vision, Deep le arning,
Classical machine learning, image processing
1 Introduction
Reliable water level estimation is fundamental to hydrological monitori ng and effective
water resources management. Key quantities such as stream discharge, r eservoir stor-
age, flood extent, and water allocation are rarely observed directly; th ey are instead
derived from water level through established physical and empirical relationships [ 1].
As a result, uncertainties in water level measurements directly translate into errors
in discharge estimation, water allocation, and flood prediction, partic ularly under
dynamic flow conditions [ 2].
Although the terms water level , water stage , and water surface elevation are often
used interchangeably in hydrological studies, they have distinct d efinitions based on
the reference datum used. Water surface elevation (WSE) refers to the height of the
water surface relative to a global or geodetic datum, enabling comparison ac ross loca-
tions [ 3]. Water stage refers to the height of the water surface relative to a lo cally
defined reference, such as the gauge datum at a specific gauging station [ 4]. The term
water level is used more broadly to describe the vertical position of t he water surface
and, in practice, is often used interchangeably with stage when a consi stent reference
is implied [ 4]. Although WSE, water stage, and water level differ in their reference def-
initions, they describe the same physical quantity and can be relate d through constant
offsets when the datums are properly defined.
The measurement of water level has evolved from manual observations to in creas-
ingly automated sensing systems. Early monitoring relied on people vi sually reading
staff gauges, where an observer records the position of the water surface again st a grad-
uated scale at specific time intervals [ 5]. While simple and widely used, this approach
is non-continuous, labor-intensive, and difficult to sustain in remot e or hazardous
environments [ 5]. The need for continuous and automated measurements led to the
development of contact-based sensing technologies grounded in physic al measurement
principles. One of the most widely adopted approaches is the use of pre ssure trans-
ducers, which infer water level from the hydrostatic pressure ex erted by the overlying
water column, based on the relationship between pressure and fluid de pth [6]. Another
common method is the float-based system, in which a buoyant element r ises and falls
2

with the water surface. Its vertical displacement is converted in to a continuous mea-
surement signal, often within a stilling well designed to reduce the influence of surface
fluctuations. In addition, bubbler gauges estimate water level indire ctly by measur-
ing the air pressure required to force gas through a submerged outlet at a known
elevation, where the back-pressure reflects the height of the water column above the
outlet [ 7]. Although these methods provide continuous measurements with re latively
high accuracy, their reliance on direct contact with water introduce s significant oper-
ational challenges. Sensors are continuously exposed to sediment trans port, floating
debris, biofouling, corrosion, and ice formation, all of which can degrade pe rformance
and increase maintenance requirements. These limitations become p articularly crit-
ical during high-flow and flood conditions, where both measurement re liability and
physical durability are most challenged [ 8, 9].
To address the limitations of contact-based sensors, non-contact techn iques were
developed to measure water level without physical interaction with the water surface.
Ultrasonic sensors estimate distance by emitting acoustic pulses and calculating the
time delay of the reflected signal, assuming a known speed of sound in air [ 10]. A
related approach uses radar sensors, which rely on electromagnetic wave s for ranging
and are generally less sensitive to variations in temperature, humid ity, and atmospheric
conditions. By avoiding direct exposure to water, these systems i mprove durability
and operational safety, particularly in harsh or inaccessible environme nts. However,
their performance can still be affected by surface roughness, wave act ivity, signal
scattering, and installation geometry. In addition, radar-based system s, while more
robust, are often more expensive to procure and deploy [ 11]. More broadly, both
ultrasonic and radar sensors provide measurements at fixed locations, l imiting their
ability to capture spatial variability across larger or complex environme nts [ 12, 13].
Such variations can result from changes in channel geometry and slope, flow conditions,
sediment accumulation, vegetation, gates, and other hydraulic structu res, and inflows
or diversions along the channel. Therefore, monitoring water-level v ariability across
large or hydraulically complex water networks requires many sensor ins tallations, with
correspondingly high infrastructure and maintenance costs.
Computer vision offers a more recent, non-contact alternative that addre sses sev-
eral of these limitations. Instead of placing sensors in the water, image -based methods
estimate water level from images or video captured by cameras, using th e position of
the water surface relative to reference features such as staff gauges, c alibration tar-
gets, or fixed structures [ 5, 14]. Because cameras can be mounted away from the water,
they avoid fouling and flood damage, provide a visual record that can be ch ecked by
an operator, and can make use of existing surveillance and webcam infrastr ucture.
Beyond fixed ground-based cameras, imagery from unmanned aerial vehicles (UA Vs)
and satellites has extended visual monitoring to sites without fixed infrastructure,
although at coarser spatial and temporal resolution [ 15, 16].
Early image-based approaches relied on classical image processing, in whi ch the
waterline is located with handcrafted rules such as intensity thr esholding, edge detec-
tion, or line detection, and then converted to a physical water level using the gauge
scale or a geometric calibration [ 17, 18]. These methods are computationally light and
easy to interpret, and under controlled conditions they can achieve m illimeter-level
3

accuracy [ 19]. However, because they depend on manually tuned features and thres h-
olds, their performance degrades under changing illumination, reflec tions, shadows,
occlusion, turbid water, and cluttered backgrounds, and they usuall y require site-
specific adjustment [ 18]. These limitations restrict their robustness in long-term field
deployment.
Recent advances in artificial intelligence (AI) have significantly en hanced the capa-
bility of image-based water level estimation. Classical machine learni ng and deep
learning approaches shift the problem from rule-based detection to d ata-driven infer-
ence, enabling models to learn robust representations directly f rom visual data.
Convolutional neural networks (CNNs), in particular, have been widely adopted for
detecting and segmenting the water surface under complex condition s [ 20, 21], while
temporal models leverage continuity in image sequences to improve s tability and
reduce noise [ 22]. More recently, transformer-based architectures have been appl ied to
water level estimation; unlike CNNs, which analyze small local regions of an image,
transformers use a self-attention mechanism that relates every part of the image to
every other part, allowing them to capture the global context of the image , such as
the full extent of the water surface and its relation to the gauge or river bank. Building
on this architecture, vision foundation models such as the Segment Any thing Model
(SAM) [23] are pretrained on very large and diverse image collections and can segment
objects, including water, in new scenes with little or no addition al training, which
reduces the need for large site-specific labeled datasets and improv es generalization
across sites and imaging conditions [ 14, 24, 25]. Long-term field deployments have also
begun to test these systems under operational conditions [ 26]. Collectively, these devel-
opments have transformed image-based approaches into scalable and low-m aintenance
alternatives to traditional sensing methods.
Research on image-based water level monitoring spans several discipli nes, includ-
ing hydrology, computer vision, and environmental monitoring. Although nu merous
studies have proposed and evaluated individual methods, a comprehe nsive understand-
ing of their comparative performance, data requirements, and practic al limitations
remains limited. Previous reviews have covered parts of this land scape. Wu et al. [ 27]
reviewed non-contact water level measurement based on computer visi on and radar,
Manfreda et al. [ 28] surveyed image-based river monitoring with an emphasis on sur-
face velocity and discharge, Kossieris et al. [ 29] reviewed satellite-based inland water
level monitoring, and Adikari et al. [ 15] focused on computer vision for flash flood
forecasting. None of these reviews jointly examines image processing, c lassical ML,
and DL methods for water level estimation alongside the imaging modaliti es, datasets,
and deployment constraints that determine whether such methods can be used oper-
ationally, and none covers the rapid adoption of transformer and foundation-m odel
approaches since 2023. This paper addresses this gap through a structured r eview of
93 studies selected following the PRISMA 2020 guidelines [ 30]. The aim is to provide
a unified perspective on how water level can be estimated from visual data, link-
ing methodological developments with data sources, sensing platform s, and practical
deployment considerations. In doing so, the review seeks to clarif y the strengths and
limitations of existing approaches and to identify directions for fut ure research.
4

The review investigates several important aspects of image-based water level
monitoring, as outlined below:
• What image-based and AI-driven approaches have been used for water level
estimation, and what are the principles and architectures underlyi ng each approach?
• What measurement accuracy has been reported for image-based water level moni-
toring methods, and what challenges and limitations have hindered th eir widespread
adoption?
• What are the different imaging modalities and devices used for water le vel
monitoring?
• What datasets are available in the literature, and what are their characte ristics and
limitations for enabling robust and reproducible evaluation, and what ac curacy has
been reported for different methods in previously published stu dies?
• What are the key challenges and considerations that determine the appli cability of
these methods across diverse real-world environments?
2 Methodology and Literature Selection
This review follows the Preferred Reporting Items for Systemati c Reviews (PRISMA
2020) reporting guidelines [ 30]. PRISMA is a widely used, evidence-based standard
for reporting literature reviews transparently, so that readers can see how studies were
found, which were excluded and why, and how the final set of studies w as reached,
and can reproduce the selection process. Its central element is a flow diagram that
tracks the number of articles through four successive stages: (1) iden tification, in
which articles are retrieved from the selected databases and duplic ates are removed;
(2) screening, in which the titles and abstracts of the remaining arti cles are checked
against the eligibility criteria; (3) eligibility, in which the ful l texts of the retained
articles are assessed in detail; and (4) inclusion, which gives the fin al number of studies
analyzed in the review. At each stage, the number of excluded article s and the reasons
for their exclusion are reported. The PRISMA flow diagram for this rev iew is shown
in Figure 1.
2.1 Search Strategy and Databases
Research articles were retrieved from five databases: Google Scholar, S emantic Scholar,
IEEE Xplore, ScienceDirect, and Scopus. These databases were chosen because
together they cover the hydrology, remote sensing, and computer vision literature
relevant to image-based water level monitoring.
The first phase of the literature selection process involved ident ifying relevant
keywords derived from the research objectives and questions: water level monitoring ,
water level estimation , computer vision , deep learning , machine learning , and image
processing. Search strings were then constructed by combining each water leve l term
with each method term using the Boolean operator AND. Examples include :
• “water level monitoring” AND “computer vision”
• “water level monitoring” AND “deep learning”
• “water level monitoring” AND “machine learning”
5

Fig. 1 PRISMA flow diagram of the literature identification, screen ing, eligibility assessment, and
inclusion process.
• “water level monitoring” AND “image processing”
• “water level estimation” AND “computer vision”
• “water level estimation” AND “deep learning”
• “water level estimation” AND “machine learning”
• “water level estimation” AND “image processing”
Each search string was adapted to the syntax of the respective database. The
articles retrieved by all search strings were combined, and duplic ates were removed
before screening.
2.2 Eligibility Criteria
Studies were included if they met all of the following criteria: ( 1) they were original
research articles or full conference papers presenting a method an d its experimental
6

validation; (2) they were published in English; (3) they estimated w ater level, or
detected the water surface for the purpose of water level estimation, from images;
(4) they developed or evaluated image processing, machine learning, or de ep learning
methods for water segmentation or water level detection, computer vis ion systems for
real-time water level monitoring, or different imaging modalities and d evices for water
level estimation.
References cited in Sections 1 and 3 to describe measurement principles, sensing
technologies, satellite missions, and previous reviews provide bac kground context are
not part of the review.
2.3 Screening and Selection
The database searches returned 168 research articles: 65 from Google Schol ar, 55 from
Scopus, 25 from Semantic Scholar, 14 from ScienceDirect, and 9 from IEEE Xplore.
After duplicates and records that clearly did not meet the eligibilit y criteria were
removed (n = 45), the titles of the remaining 123 articles were screen ed, and 18 articles
were excluded. The abstracts of the remaining 105 articles were then screened, and 7
articles were excluded because they primarily addressed water qu ality monitoring or
other water-related applications rather than water level. The full te xts of the remaining
98 articles were assessed against the eligibility criteria, and 5 artic les were excluded
because they did not provide sufficient information on the application of AI techniques
to water level measurement. In total, 93 studies were included in t he review (Figure 1).
For each included study, we recorded the publication year, study lo cation, water
body type, imaging platform and modality, method category (image proces sing,
machine learning, or deep learning), model architecture, dataset char acteristics, and
reported accuracy metrics. These data form the basis of Tables 1–4.
2.4 Bibliometric Overview
To characterize the thematic structure and evolution of the literatu re, a keyword
co-occurrence network was constructed from the cited literature u sing VOSviewer
(Figure 2). Title keywords were grouped into three methodological approaches (c las-
sical image processing, classical machine learning, and deep learning), two detection
strategies (gauge-based and gauge-free), and two application settings (irr igation canals
and other applications). Circle size indicates the number of titles in each category,
lines indicate co-occurrence within the same title, and color and ve rtical position show
the average publication year of the titles in each category. The network s hows that
deep learning is now the most frequently used approach ( ≈56 titles) and that the field
has shifted from classical image processing through classical machine l earning to deep
learning, alongside a move from gauge-based to gauge-free detection. This p rogres-
sion indicates the structure of the review, which moves from class ic image processing
(Section 4), classic machine learning (Section 5), and deep learning (Section 6).
7

Fig. 2 Keyword co-occurrence network derived from the cited liter ature.
3 Water Level Monitoring Methods
As defined in Section 1, water level describes the vertical position of the water sur-
face of a water body, such as a canal, river, reservoir, or lake, relative to a reference
datum. Water level measurement approaches are primarily divided int o two broad
groups: (a) contact-based methods, in which the sensing element is placed in the
water, and (b) non-contact methods, in which the water surface is obse rved remotely.
Figure 3 presents the overall taxonomy of water level monitoring techniques, showing
the transition from contact-based methods to non-contact methods base d on sensors
and computer vision.
8

Fig. 3 Classification of water level monitoring techniques into co ntact-based and non-contact
approaches.
3.1 Contact-based Methods
Contact-based methods use sensors or reference structures placed in the water to
measure water level. These sensors range in complexity and are often de signed for
specific applications. Figure 4 shows the common contact-based methods, and their
working principles are described below.
9

Fig. 4 Common contact-based water-level measurement methods: (a ) crest-stage gauge, (b) staff
gauge, (c) pressure transducer, (d) float gauge, and (e) bubb ler. Source: U.S. Geological Survey
(USGS) Technical Memorandum 3-A7.
3.1.1 Crest Stage Gauges
A crest-stage gauge is one of the simplest ways to measure water level [ 31]. The gauge
consists of a vertical pipe containing a wooden staff and cork. Water ent ers through
intake holes at the bottom of the pipe and rises inside it, and the cork float s on the
rising water. When the water recedes, cork adheres to the staff at the highest water
level reached, allowing the peak stage of an event to be read. The main di sadvantages
are that the intake holes can become clogged, insects can nest in the pipe , and cork
can be washed out through the vent hole. In addition, the gauge records only t he peak
level, not a continuous time series.
3.1.2 Staff Gauges
A staff gauge is a graduated scale installed in the water that provides a dir ect visual
indication of water level [ 4]. Staff gauges are among the most widely used reference
devices and are routinely used to check and calibrate electronic lev el sensors. They
10

can be installed in almost any setting, including canals, lakes, and riv ers. Their main
limitations are that readings require an observer, the scale can corrode or become
fouled, and the gauge cannot be monitored remotely without an additional devi ce.
Although the gauge itself is in contact with the water, reading it with a c amera, as
described in Sections 4–6, provides a hybrid measurement approach that combines a
contact-based reference with non-contact remote observation.
3.1.3 Pressure Transducers
Pressure transducers measure water level from the pressure exer ted on them by the
overlying water column [32]. As the water level above the sensor increases, the pressure
increases. The measured pressure is then converted to a water dep th in meters or
feet. Atmospheric pressure must be accounted for absolute (non-ven ted) transducers
require separate barometric compensation, whereas differential (ven ted) transducers
compensate for it directly through a vent tube.
3.1.4 Float Gauges
A float gauge consists of a tape or cable running over a pulley [ 33]. One end of the
tape or cable is attached to a float in a stilling well while the other end is attached
to a counter weight. The float rises and falls with the water level, and t he water level
can be read by means of an index and graduated tape or the pulley can be connect ed
to a water-stage recorder or shaft encoder for continuous recording.
3.1.5 Bubblers
Bubblers [ 34] are hydrostatic water level measurement sensors that estimate water
level by measuring the air pressure required to continuously forc e bubbles through the
open end of a submerged dip tube. The measured back pressure equals t he hydrostatic
pressure exerted by the water column above the tube outlet. The wat er depth above
the outlet is then obtained from the hydrostatic relationship P = ρgh, where P is the
gauge pressure (Pa), ρ is the water density (kg m −3; approximately 1000 kg m −3 for
fresh water), g is the gravitational acceleration (9.81 m s −2), and h is the water depth
above the sensor (m).
3.2 Non-contact Methods
Non-contact methods measure water level without placing the sensor in the water.
They include ultrasonic, radar, laser, and satellite-based methods , as well as the image-
based and AI methods that are the focus of this review. Non-contact sensor s are
advantageous in hazardous or harsh environments where direct contact with water
could damage equipment or create safety risks. However, they remain su sceptible to
equipment failure, power outages, wind, and extreme weather, and th ey require careful
calibration for accurate measurement.
11

3.2.1 Sensor-based Methods
Ultrasonic Sensors
Ultrasonic sensors were the first widely deployed non-contact techn ology in operational
hydrometry [ 35]. These sensors estimate water level by transmitting acoustic pul ses
toward the water surface and measuring the time required for the refl ected signal to
return to the sensor. The distance to the water surface is then calc ulated using the
known speed of sound in air. Because no components need to be submerge d, ultrasonic
sensors are relatively simple to install and maintain compared to contac t-based gauges
[36]. However, their performance is strongly influenced by environmen tal conditions.
Variations in air temperature, humidity, wind, rainfall, and surface turbulence can
affect sound propagation and introduce measurement errors. Consequentl y, ultrasonic
sensors generally perform best in controlled environments such as re servoirs, culverts,
and hydraulic structures, while their accuracy may deteriorate un der highly turbulent
or rapidly changing flow conditions [ 37].
Radar Sensors
Radar sensors were introduced to overcome several limitations assoc iated with ultra-
sonic measurements. Similar to ultrasonic sensors, radar systems es timate the distance
to the water surface using a time-of-flight principle, but instead of acoustic pulses
they transmit electromagnetic waves. Since electromagnetic signals are largely unaf-
fected by atmospheric conditions such as temperature and humidity, radar sensors
typically provide more stable and reliable measurements across a wid er range of envi-
ronmental conditions [ 38]. Modern systems commonly employ frequency-modulated
continuous-wave (FMCW) radar technology, which enables highly accurat e water level
measurements while maintaining reliable operation during rainfall, fog, and nighttime
conditions. These advantages have made radar one of the most widely used n on-contact
technologies in modern hydrometric networks [ 39]. Nevertheless, measurement qual-
ity can still be affected by factors such as excessive surface roughnes s, strong wave
activity, signal scattering, and unfavorable installation geometry, part icularly during
extreme flow events.
Laser Sensors
Laser sensors, including terrestrial laser rangefinders and LiDAR (Light Detection and
Ranging) systems, provide another non-contact approach for water level mon itoring.
Similar to ultrasonic and radar sensors, they estimate the distance to the water surface
by measuring the travel time of reflected near-infrared laser puls es [ 40]. Their narrow
beam width enables highly precise measurements and detailed mappin g of water sur-
face elevations. However, their use in routine hydrometric monitori ng remains limited.
Measurement accuracy can be affected by weak reflections from rough, turb ulent, or
highly turbid water surfaces, while eye-safety requirements con strain laser power and
operating range. As a result, laser-based systems are used primarily i n research and
specialized surveying applications rather than in permanent monitor ing networks [ 41].
12

Satellite Altimetry
Satellite remote sensing has greatly expanded the ability to monitor w ater levels in
regions where ground-based measurements are unavailable or difficult to maintain.
This capability is particularly valuable given that a large portion of the w orld’s rivers,
lakes, and reservoirs remain ungauged or poorly monitored. Both laser altim eters (light
detection and ranging, LiDAR) and radar altimeters (radio detection and rangi ng,
RADAR) have been widely used to monitor water levels in inland lakes [ 29, 42]. Early
radar altimeters operated primarily in low-resolution mode (LRM), wh ich is suited to
large water bodies. CryoSat-2 introduced the ability to switch bet ween three modes:
LRM, synthetic aperture radar (SAR), and SAR interferometry (SARIn), which allow
for measurements with smaller footprints and the monitoring of smalle r water bodies
[43–46]. Sentinel-3 extended global observations in SAR mode, and Sentinel -6 fur-
ther reduced measurement noise through a higher pulse repetition f requency [ 47, 48].
Several satellite altimetry-derived products for time-series lake-level data have been
developed, such as DAHITI [ 49] and Hydroweb [ 3], which integrate data from multi-
ple radar altimeters. Despite these advances, radar altimetry remain s limited by large
footprints and by the performance of retracking algorithms [ 46].
In contrast, laser altimetry, with its smaller footprint and higher al ong-track sam-
pling density, is well suited to observing small lakes and reserv oirs [50]. The Ice, Cloud,
and Land Elevation Satellite (ICESat), the first laser altimeter, prov ided highly accu-
rate lake level data for 100 lakes in the Tibetan Plateau (TP) region from 2003 to
2009, with an accuracy of approximately 10 cm for Qinghai Lake during the ice-fr ee
season [ 51, 52]. ICESat-2, launched in September 2018, carries the Advanced Topo-
graphic Laser Altimeter System (ATLAS) and provides substantially high er spatial
resolution and denser sampling than the original ICESat [ 53]. Validation results of the
ICESat-2 Along-Track Inland Surface Water Data (ATL13) product based on in- situ
measurements indicate that ICESat-2 has significantly improved the ability of satellite
altimetry to capture lake water levels, with lower uncertainty and bias than previous
satellite methods [ 54]. The Global Ecosystem Dynamics Investigation (GEDI) lidar
was designed to monitor forest structure and the global carbon cycle, but its eleva-
tion data have also been applied to inland waters. Comparative assessme nts show
that GEDI elevations are less accurate than those of ICESat-2 [ 55]. A major recent
advancement is the Surface Water and Ocean Topography (SWOT) mission, wh ich
employs Ka-band radar interferometry to generate two-dimensional meas urements of
water surface elevation, extent, and slope [ 56, 57]. Unlike conventional nadir altime-
ters, which observe only along the satellite ground track, SWOT provi des spatially
distributed observations across wide swaths, substantially improv ing the monitoring
of rivers, lakes, reservoirs, and floodplains over large regions [ 56].
Despite their ability to provide observations over vast and often in accessible areas,
satellite-based methods have several limitations. Revisit inte rvals are often too long for
continuous monitoring or rapid flood assessment, while spatial resolut ion and footprint
size can restrict observations of narrow channels and small water bodie s. Consequently,
satellite observations are best viewed as a valuable complement to groun d-based
monitoring systems rather than a complete replacement. In contrast, ground camera-
based methods offer a further non-contact alternative, in which camer as capture the
13

water surface images, together with a staff gauge or without staff-gauge, to esti mate
the water level from the images. These techniques include image-pr ocessing, classical
machine-learning, and deep-learning methods. They allow remote me asurement with-
out installing sensors in the water and can provide continuous measure ments at high
temporal resolution, and record visual information on site. Image-based te chniques
are the main focus of this review and are discussed in detail in the fol lowing section.
4 Visual Data Acquisition and Image Processing
Methods
4.1 Computer Vision for Water Level Monitoring
Computer vision, a subfield of artificial intelligence (AI), enables c omputers to process,
analyze, and interpret visual information from digital images and videos. T hrough
techniques such as image processing, feature extraction, and pattern recognition,
computer vision converts visual information into meaningful numeric al or symbolic
representations that can support classification, detection, and quanti tative estimation
[58]. Practical applications of computer vision include autonomous driving [59], flood
detection [ 60], facial recognition [ 61], and, in hydrology, staff gauge detection [ 62].
These techniques enable computers to learn and recognize patterns i n images and
videos, as illustrated in Figure 5. In human vision, the eye captures the scene of the
staff gauge and the brain interprets this visual information to determin e the water
level. Computer vision follows the same two-stage process: a camera ac ts as the sens-
ing device that captures the image, and a computer acts as the interpre ting device
that analyzes the image to estimate the water level, in the same way that the eye and
brain work together in human perception. However, unlike the human br ain, which
uses previous knowledge and context to adjust to changing conditions, a computer
vision approaches relies on patterns learned from its training data and m ay therefore
be less accurate under conditions it has not seen before.
Fig. 5 Analogy between human vision and computer vision for readin g water level from a staff gauge.
14

Computer vision techniques have been applied to water level monit oring since
1998, when Takagi et al. [ 63] developed a non-contact water level measuring system
based on image processing. These techniques have since been used t o measure water
levels in various water bodies, including rivers, streams, ditc hes, and canals, using
images captured with or without staff gauges [ 62, 64]. Image-based monitoring systems
typically use visible-spectrum (RGB) cameras and, less commonly, near-infrared or
thermal cameras [ 65–67].
Computer vision approaches for water level monitoring can generally be cat egorized
into two main groups (1) image processing methods and (2) machine learni ng-based
methods [ 17, 18, 68]. Image processing methods rely on techniques such as edge
detection, thresholding, image segmentation, and feature extraction to identify water
boundaries and estimate water levels. In contrast, machine learning-b ased approaches
use algorithms that learn patterns and relationships directly from image data. These
approaches can be further divided into classical machine learning me thods, such as
random forests, support vector machines, and multilayer perceptron s [ 69, 70], and
deep learning methods, such as convolutional neural networks (CNNs)[ 71–74].
The following sections review the most commonly applied classical im age process-
ing, traditional machine learning, and deep learning techniques for im age-based water
level monitoring.
• Thresholding: One of the simplest and most common methods is thresholding,
which converts grayscale images into binary (black and white) images [ 75]. A region
of interest (ROI), such as a staff gauge in a river or canal, can be different iated from
background pixels by applying a threshold value. Choosing an appropriat e threshold
ensures that relevant pixels are correctly included or excluded [ 76].
• Edge detection: These techniques detect the boundaries of a staff gauge within
an image. The Canny and Sobel operators are the most common. They identify th e
staff gauge by detecting sharp changes in pixel intensity, such as the m arkings on a
gauge or the edges of the water body [ 27].
• Convolutional neural networks (CNNs): Within deep learning, CNNs have
transformed vision-based water level monitoring. These networks autom atically
extract features and learn from image data, capturing variations in water su rface
patterns, reflections, and staff gauge markings. By learning spatial and cont extual
features from camera imagery, CNNs have become the most widely used appr oach
in recent studies (Section 6).
4.2 Traditional Image Processing Methods
Traditional image processing methods use predefined mathematical ope rations and
rule-based algorithms to process digital images and extract relevant vi sual information,
such as edges, shapes, colors, and intensity variations. These method s typically ana-
lyze images individually, using pixel-level characteristics to identify features of interest.
Image processing-based water level methods depend on handcrafted al gorithms, so
their accuracy is highly sensitive to environmental conditions, im age quality, and the
effectiveness of the chosen features [ 77]. They typically follow a standard workflow of
15

image acquisition, region of interest (ROI) identification, preproce ssing, feature extrac-
tion, waterline detection, and mapping of image coordinates to physical water levels
(Table 1). The ROI is usually defined around the staff gauge or calibration region and
contains both water and non-water pixels, as illustrated in Figure 6 [78]. Within this
region, various feature extraction schemes have been developed, wh ich can be broadly
classified into thresholding-based methods, multi-frame analysis methods, edge-based
detection methods, and gauge-reading methods that recognize the marki ngs on the
staff gauge.
Fig. 6 Example of a region of interest (ROI) for water-level estima tion. Source:
https://doi.org/10.3390/w15183233, reproduced under th e Creative Commons Attribution 4.0 Inter-
national License (CC BY 4.0).
Thresholding-based methods estimate water levels by exploiting d ifferences in pixel
intensity or color between water and non-water regions. They generally t ransform
images into binary masks by selecting an optimal threshold, with Otsu ’s maximum
inter-class variance method being one of the most widely used [ 79–81]. Although
thresholding methods offer computationally efficient and relatively s imple implementa-
tion, their accuracy degrades under varying lighting conditions becau se of reflections,
shadows, glare, and uneven intensity between the gauge and the water su rface. For
instance, [ 71] converted images from the RGB to the Hue-Saturation-Value (HSV)
color space to enhance the separation between water and gauge regions. Simil arly, [82]
applied HSV-domain enhancement to increase the contrast between the w ater surface
and the staff gauge, whereas [ 83] combined k-means clustering with region-growing
techniques to improve water body segmentation. To reduce sensit ivity to illumination
changes at the acquisition stage, [ 66] used near-infrared (NIR) imaging to enhance
the visibility of the staff gauge before detecting the waterline, alth ough the approach
16

still relies on precise gauge detection and calibration. These modific ations demon-
strate that combining additional image characteristics can enhance thre shold-based
techniques, although they remain sensitive to site-specific cond itions [ 18].
Multi-frame aggregation approaches mitigate some limitations of single-fram e tech-
niques by leveraging temporal information from consecutive images. Ins tead of relying
solely on spatial differences, these techniques analyze variations cau sed by water move-
ment while assuming that the water surface changes more than fixed objec ts such as
staff gauges. Cumulative histograms and temporal intensity variations are pr ocessed
to emphasize water-related features and produce projection curve s from which the
waterline is identified [ 84, 85]. Similarly, image differencing methods identify pixel
intensity changes across consecutive frames that correspond to water s urface move-
ment, often enhanced through edge detection and morphological operations [ 86], while
[87] estimated water level from temporal intensity variations using a hi stogram-based
approach. These techniques perform best under high-flow or turbule nt conditions,
because surface movement provides additional information for separati ng water from
its surroundings. However, their performance can degrade under low- flow conditions
where insufficient temporal variation exists, as well as under camera m ovement or
rapidly changing environmental conditions.
Edge-based techniques identify the waterline by detecting shar p changes in pixel
intensity associated with the boundary between water and its surrou ndings. Because
the boundary separating the water surface and staff gauge produces a strong edge
feature, edge detection has been widely investigated for water level estimation. For
instance, [ 88] proposed a water level detection system by applying a horizontal edge
detector combined with a finite impulse response (FIR) filter to e xtract features under
different lighting conditions. Similarly, [ 89] employed the Marr-Hildreth operator to
detect the edge pixels of the ruler within ROI. However, distingu ishing the water
line from other edges, such as scale markings and structural boundaries, remains a
major challenge. To address this issue, [ 18] proposed the maximum mean difference
(MMD) method, which exploits the fact that the periodic pattern of a bicolor water
gauge is visible only above the water surface. It computes the row-wis e mean of the
grayscale and edge images and, using a coarse-to-fine search, locates the w ater line
as the position with the largest mean difference between the regions abo ve and below
it. Other studies have combined several detectors to compensate f or the weaknesses
of each. [ 90] localized the region of interest using edge maps from a pretrained ed ge
detector and template matching, and then estimated the waterline thr ough two parallel
strategies: fitting a line to the detected edge points using linear regression, and a
sliding-window search based on pixel intensity differences. The se methods improve
detection accuracy by combining multiple image characteristics but require carefully
designed algorithms, careful integration, and parameter selection.
Gauge-reading methods measure water level by recognizing the numer ical markings
on the staff gauge and infer the water level from the known layout of the gauge s cale.
[62] integrated template matching with sequence verification to ident ify the digits on
a gauge, exploiting the fixed order of the markings to correct detection errors caused
by noise, missing characters, or partial occlusion. Compared with tem plate matching
alone, these sequence constraints improve reliability, although pe rformance remains
17

dependent on clear visibility of the gauge markings, and recognition can de grade under
poor lighting, blurred imagery, or partially submerged markings.
Beyond staff-gauge-based monitoring, classical image processing methods have also
been used to extract water bodies directly from camera and remote se nsing imagery.
Unlike the gauge-based methods described above, which analyze a small region around
a staff gauge to locate a single waterline position, these methods classi fy water pixels
across the entire scene. For instance, [ 91] developed an automated threshold-based
method for extracting rivers and lakes from Landsat imagery, producing a water extent
map rather than a water level. For camera imagery, GRIME2 [ 19] offers an open-source
tool that measures water levels without a staff gauge, using calibration tar gets and
geometric transformations between image coordinates and real-world measu rements.
The workflow includes image preprocessing, calibration marker recogni tion, waterline
extraction, and conversion of pixel-based measurements into physic al water levels.
Figure 7 shows a typical image processing pipeline for water level estimati on, consisting
of preprocessing, feature extraction (e.g., edge detection and thres holding), and feature
projection.
Fig. 7 Workflow of a classical image processing approach for water l evel detection. Source:
https://doi.org/10.3390/w15183233, reproduced under th e Creative Commons Attribution 4.0 Inter-
national License (CC BY 4.0).
Overall, conventional image processing techniques have demonstrat ed the feasi-
bility of automated water level monitoring and remain attractive because of their
simplicity, transparency, and minimal computational requirements . However, their
reliance on manually selected features and predefined thresholds li mits their adapt-
ability to diverse environmental conditions. Factors such as changing illumination,
water reflections, camera displacement, complex backgrounds, and vege tation can sig-
nificantly affect performance. These limitations have motivated the u se of machine
18

learning methods, which learn decision rules from labelled data rat her than relying
solely on fixed thresholds and hand-tuned parameters.
5 Machine Learning-Based Techniques for Water
Level Monitoring
Image-based water level detection has evolved from approaches relying p rimarily on
classical image processing to those incorporating artificial intellige nce (AI), particu-
larly classical machine learning (ML) and deep learning (DL). While cl assical image
processing applies predefined rules, such as edge detection and th resholding, classi-
cal ML methods learn a decision model, such as a random forest or suppor t vector
machine, from labelled training data, typically using handcrafted fe atures as input.
DL methods, in contrast, learn the features themselves directly from raw imagery and
are discussed in Section 6. These categories are not mutually exclusive, as many ML-
and DL-based frameworks integrate classical image processing algorithms f or prepro-
cessing, feature extraction, or postprocessing. More broadly, AI metho ds have been
applied in hydrology to flood prediction [ 92, 93], water level and stage estimation [ 94],
and assessing water availability at ungauged sites [ 95]. Camera-based water level esti-
mation pipelines, both with and without a staff gauge, are illustrated in F igure 8, and
the main image processing and classical ML methods, together with the ir reported
performance and limitations, are summarized in Table 1.
Fig. 8 Camera-based water level estimation pipelines (a) with a st aff gauge (SG) and (b) without
a staff gauge, using image-to-elevation calibration.
19

Compared with image processing and deep learning, relatively few stu dies have
applied classical ML to camera-based water level monitoring. In these s tudies, trained
models have mainly been used to segment the staff gauge or water region, r ecognize
gauge markings, or locate the gauge before the waterline is extracted. [ 69] segmented
the staff gauge using a two-class random forest classifier trained on texton f eatures,
applied Gaussian mixture model segmentation to separate bars and numer als, and rec-
ognized the gauge digits using an optical character recognition (OCR) modul e based
on a random forest classifier. Similarly, [ 96] applied supervised random forest classifi-
cation to smartphone images to highlight flooded regions before detectin g the flooding
line with edge-based methods and converting it to water level usin g photogrammetry.
[64] combined threshold-based image processing with a trained Aggregated Channel
Features (ACF) object detector, which located the staff gauge in night i mages before
rule-based waterline extraction.
Another learning-based alternative is dictionary learning, which repr esents image
patches as sparse combinations of learned basis functions. [ 97] used this approach to
distinguish staff gauge from water regions, and compared it with a frame-diff erence
method and a convolutional neural network. Compared with handcrafted ru les, dictio-
nary learning adapts better to diverse conditions while remaining l ess computationally
expensive than deep learning, although it still requires represen tative training data
and careful parameter tuning.
Overall, classical ML methods improve adaptability compared with im age process-
ing algorithms, but they still depend on handcrafted features and sit e-specific training
data, which limits their generalization across monitoring locations. Th ese shortcomings
have motivated the adoption of deep learning methods, which automatical ly learn hier-
archical feature representations directly from raw imagery and often ac hieve improved
robustness under complex real-world conditions.
20

T able 1 Summary of classical image processing and classical machin e learning techniques for
water level monitoring.
Method Reference Key T echnique Metric Notes / Limitations
Traditional image processing
Thresholding
methods
[79] Maximum inter-class variance (Otsu)
binarization on stage-camera images of
a reference pole
MAE 0.27–2.20 cm; RE
3.42–19.95%
Low-cost wildlife camera; sensitive to
noise, reflections, shadows, and flares
[80] Maximum inter-class variance (Otsu)
binarization for real-time water level
recognition
– Also estimates surface velocity; sensitive to
illumination changes
[81] Maximum inter-class variance (Otsu)
binarization for reservoir water level
monitoring
– Common baseline; sensitive to reflections
and shadows
[82] HSV color-space enhancement + recog-
nition of ruler readings
Avg. measurement error
±0.90 cm
Improves contrast between gauge and
water
[83] K-means clustering (RGB space) +
region-growing algorithm
Event accuracy 89.29%;
miss rate 11.76%
Unsupervised clustering used as a segmen-
tation aid; selects water body after seg-
mentation
NIR imaging [ 66] Near-infrared video imaging, ROI
extraction, and NIR intensity response
analysis
Resolution 1 mm Reduces sensitivity to illumination
changes; still relies on precise gauge detec-
tion and calibration
Multi-frame
methods
[84] Accumulated histogram with band-pass
filter across multiple frames
– Relies on water surface movement being
more prominent than gauge changes
[85] Multi-frame binary images; reference
level at the sharpest drop in the his-
togram; embedded real-time system
– Effective under high-flow/fluctuating con-
ditions; tested on a real river
[86] Image differencing + edge detection +
morphological operations
∼1 cm Detects water surface movement via tem-
poral differences
[87] Histogram-based temporal analysis of
intensity changes between consecutive
frames
Average normalized cross-
correlation (NCC) = 0.87
Degrades under camera movement or low
surface motion
Edge-detection
methods
[88] Horizontal edge detector + FIR filter – Extracts features under varyin g lighting
conditions
[89] Marr-Hildreth operator – Detects edge pixels of ruler within ROI
[18] Combines grayscale and edge features;
coarse-to-fine detection via maximum
mean difference (MMD)
Accuracy 1 cm; resolution
1 mm
Addresses similarity between water line
and scale line edges
[90] Pretrained edge maps + template
matching for ROI; ensemble of line fit-
ting (linear regression) on edge points
and sliding-window intensity search
MAE 4.8; MAPE 3.1%;
R2 0.92
Combines two branches for robustness;
increases computational complexity
Gauge reading [ 62] Digit template matching + sequence
verification of gauge markings
R2 0.943 Robust to noise, scale, orientation, and
partial occlusion; requires visible markings
Gauge-free calibra-
tion
[19] GRIME2: calibration target-based geo-
metric mapping from pixel to real-world
units
±3 mm (lab); ±5 mm at
90% CI (field)
Open-source framework; includes prepro-
cessing, waterline, and calibration marker
detection
Classical machine learning
Random forest [ 69] Random forest gauge segmentation on
texton features + GMM segmentation

• random forest-based OCR of gauge
  digits; quadratic water level function
  – Portable mobile-phone cameras; requires
  the gauge to be distinguishable from the
  background
  [96] Random forest classification of flooded
  regions + Hough transform/Canny line
  detection + photogrammetric conver-
  sion with UA V-derived DEM
  Classification accuracy
  88.05%; water level differ-
  ence <3 cm on average
  Uses crowdsourced smartphone images;
  requires a high-resolution elevation model
  Boosted object
  detector
  [64] Aggregated Channel Features (ACF)
  detector for gauge localization at night
• multi-threshold frame classification
• threshold- and morphology-based
  waterline extraction
  Night ∼98.7%, day
  ∼84.8% of readings within
  ±3 cm
  Edge-based frame quality check; day
  frames less reliable due to variable illumi-
  nation
  Dictionary learning [ 97] Sparse representation via dictionary
  learning on image patches; reconstruc-
  tion error analysis
  – Separates water/non-water regions; com-
  pared with frame differencing and a CNN
  6 Deep Learning-Based Techniques for Water Level
  Monitoring
  Deep learning (DL), a subset of ML, has emerged as the dominant approach for i mage-
  based water level monitoring owing to its ability to automatically ext ract hierarchical
  feature representations from raw data. Unlike classical image processi ng and ML tech-
  niques that rely on handcrafted features, DL models learn both low-l evel and high-level
  21

features through multiple hidden layers using backpropagation. This capability enables
the extraction of spatial, contextual, and semantic information from compl ex scenes,
making DL approaches more robust to variations in illumination, weather c onditions,
and water surface appearance. Figure 9 shows a general DL framework for water level
monitoring, including dataset preparation, data splitting, model trai ning, evaluation,
and water level estimation for systems with and without staff gauges.
Fig. 9 Deep learning (DL) framework for water level estimation, fr om dataset preparation to water
level estimation with and without a staff gauge.
Current DL-based approaches can be broadly categorized into object detec tion,
semantic segmentation, regression or classification, and hybrid detect ion–segmentation
networks. Although these frameworks differ in architecture, they ty pically follow
a common pipeline in which an ROI is identified, relevant water or gauge f ea-
tures are extracted, and the detected features are mapped into phys ical water level
measurements.
Object detection networks are among the most extensively deployed approaches
for staff gauge monitoring because they automatically locate the gauge before me a-
suring the water level. The YOLO (You Only Look Once) family of detector s [ 98] has
attracted particular attention because of its high detection accuracy and real-time
performance [ 68, 99–101]. Most YOLO-based frameworks first detect the staff-gauge
or scale markings and then estimate the waterline relative to the dete cted region.
Recent studies have further improved detection accuracy by inc orporating attention
mechanisms. For instance, [102] incorporated a Convolutional Block Attention Module
(CBAM) into YOLOv7 to enhance feature representation and focus the ne twork on the
most informative regions of the gauge. Other studies have combined CNNs wi th image
22

enhancement, such as vertical shearing, gamma correction, and morphologic al filtering,
to improve robustness under varying environmental conditions [ 94]. These advance-
ments demonstrate that enhancing feature extraction before water le vel estimation
improves detection accuracy under varying field conditions. Figure 10 illustrates the
workflow of YOLO-based water level monitoring.
Fig. 10 Workflow of YOLO (You Only Look Once) object detection for wat er level monitoring.
CNN-based methods are among the most extensively studied because the y effec-
tively extract and learn hierarchical features directly from visual data. Residual
networks, including ResNet-50 and variants such as ResNeXt and ResNetV2, h ave
been widely used for both continuous water level regression and flood c lassification
[103–105]. Compared with standard CNNs, residual architectures enable deeper net-
works through skip connections, improving feature learning while r educing training
difficulties. CNNs have also been deployed to satellite imagery, web -based images,
and crowd-sourced data. For example, DeepWaterMap [ 106] segments water bodies
from surrounding land, snow, and cloud cover in multispectral Landsat imagery, and
[107] applied a deep CNN with image segmentation to web camera images for flood
detection and water level estimation. Other studies combine CNNs wit h ranking or
classification schemes to estimate water levels or classify flood cond itions from social
media and crowd-sourced photographs [ 108, 109]. Hybrid frameworks that combine
conventional image processing with deep learning have also emerged. For instance,
[71] combined HSV-based segmentation, CNN-based digit recognition, and pixel- to-
metric calibration to measure water levels, while [ 105] coupled classical preprocessing
with CNN-based regression models for sensor-free monitoring. These st udies show that
classical image processing can complement deep learning by improvi ng feature quality
before prediction.
Transformer-based models have recently drawn attention because of their ability
to model long-range spatial dependencies that are difficult for conventi onal CNNs to
capture. Rather than depending solely on convolutional operations, trans former archi-
tectures leverage self-attention mechanisms to learn relationship s across the entire
23

image. [ 110] used transformer-based segmentation models, namely the Segment Any -
thing Model (SAM) [ 23], SegFormer [ 111], and MobileSAM [ 112], to extract water
regions under diverse environmental conditions, and then applied ran dom forest regres-
sion to estimate water level and discharge from the segmented water area. Compared
with CNN-based segmentation models, transformers typically capture gl obal con-
text better, but they often require larger training datasets and more computational
resources. Vision foundation models, which are pretrained on very large and diverse
image collections, extend this trend. Because they can segment water with little or no
site-specific training, they have been used to reduce dependen ce on site-specific labeled
data [ 24], to measure water level with minimal retraining [ 25], and to refine water-
line detection by coupling SAM with conventional water-line detect ion techniques [ 14].
Temporal information has also been combined with semantic segmentation t o stabilize
water level recognition across image sequences [ 22], and long-term field applications
have begun to assess the robustness of AI image-based monitoring under op erational
conditions [ 26].
Table 2 summarizes representative DL methods for water level monitoring, t heir
applications, and performance metrics. In terms of accuracy, most DL-bas ed methods
report water level errors in the range of approximately 1–3 cm, the highes t accura-
cies, with errors below 1 cm, have been reported by methods that c ombine precise
localization of the gauge or water region with accurate pixel-to-metric conv ersion. For
example, [ 101] reported a maximum deviation of 0.8 cm from manual readings, [ 113]
and [ 114] achieved errors below 1 cm by combining segmentation with a virtual gauge
or with detection-based waterline modules, and [ 99] reported a systematic error of
7.7 mm, with 95% of daytime images within 1 cm. Other approaches achieved slightly
lower but still competitive accuracy, such as [ 105] (MAE = 1.14 cm), [ 100] (±1.2 cm),
and [ 110] (MAE = 1.03–2.33 cm). In addition, high classification or segmentation
scores, such as accuracies above 98% [ 115] or IoU values above 0.97 [ 110, 113], do
not necessarily translate into equally accurate water level estimate s, since the final
accuracy also depends on calibration and the conversion from image coordinat es to
physical units.
Overall, DL has significantly advanced automated water level monitoring b y elimi-
nating the need for handcrafted feature extraction and facilitating en d-to-end learning
directly from imagery. Although object detection and CNN-based methods r emain the
most widely adopted because of their efficiency and maturity, transform er-based and
hybrid detection–segmentation approaches increasingly perform bette r under complex
environmental conditions. Nonetheless, challenges related to dataset availability, com-
putational requirements, and model generalization continue to motivat e research into
more resilient and transferable deep learning solutions.
24

T able 2 Summary of deep learning-based techniques for water level m onitoring.
Method Reference Key T echnique W ater Level
Accuracy
Segmentation /
Detection Metric
Notes / Limitations
YOLO-based
detection
[99] YOLOv5s detects the staff
gauge and scale characters;
waterline located by image pro-
cessing
Systematic error
= 7.7 mm; 95%
of daytime images
within 1 cm
– Field station on a river in Bei-
jing; accuracy drops for trans-
parent water
[68] Improved YOLOv5m with
attention for gauge segmenta-
tion + skew correction + linear
regression of water level
Average error
<2 cm
– Tested under daytime, night-
time, occlusion, and lighting
variations
[101] Temporal super-resolution +
lightweight improved YOLOv5
for water level line detection
Mean relative
accuracy = 98.5%;
maximum devia-
tion from manual
readings = 0.8 cm
– Indoor hydropower sump wells
with poor lighting and water
vapor
[100] DSC-YOLOv8n-seg for joint
detection of water level line and
gauge characters
Average error =
±1.2 cm
Precision = 93.1%;
recall = 94.5%;
mAP50:95 = 93.9%
Also estimates surface flow
velocity; robust in low-light
and nighttime scenes
[102] Improved YOLOv7 with
CBAM
– mAP = 87.26%; F1
= 75.6%
CBAM between backbone and
head focuses the network on
gauge regions
Object detec-
tion (SSD)
[116] SSD trained on physical pool
experiments to detect staff
gauge regions
NSE = 0.98; R2 =
0.99
– Uses photographic principles
and geometric mapping to
compute water level
CNN with
image
enhancement
[94] Vertical shearing, gamma cor-
rection, and morphological
opening filters + CNN
∆E = 3.3 cm;
RMSE = 1.2
– Improves robustness against
image noise
CNN-based
regression/
classification
[104] CBAM-improved ResNet50 for
classifying flood depth levels
– Accuracy = 92.45% Classifies urban waterlogging
depth from images
[103] ResNet-50 virtual water gauge
for real-time river water level
monitoring
– Accuracy = 83.6–
96%
Learns hierarchical fea-
tures from river surveillance
imagery
[105] Conventional preprocessing

• residual CNN with multi-
  branch feature aggregation
  MAE = 1.14 cm – Sensor-free monitoring; fine-
  tuned for regression and clas-
  sification
  Web-image
  CNN segmen-
  tation
  [107] Deep CNN + image segmenta-
  tion on web camera images
  R2 = 0.84–0.87;
  bias = 2.44–
  3.04 cm
  Flood detection:
  accuracy >90%; IoU

80%; BF1 >70%
Water level estimation and
flood detection under varying
illumination and weather
Hybrid two-
step approach
[71] HSV-based segmentation +
CNN digit recognition +
quadratic pixel-to-metric map-
ping
– Accuracy = 98.25% Combines classical and deep
learning techniques
IoT +
instance seg-
mentation
[77] IoT-enabled cameras + Mask
R-CNN segmentation
– Accuracy = 0.86; F1
= 0.97
Addresses environmental dis-
tortion, poor lighting, and
nighttime visibility; low-cost
real-time flood monitoring
Semantic
segmentation
[65] SegNet, U-Net, and FCN water
segmentation compared with
continuous image subtraction
RMSE = 0.013–
0.066 m (SegNet
vs. gauge station)
Accuracy: SegNet =
97%, U-Net = 94%,
FCN = 98%
SegNet outperformed the con-
ventional image subtraction
method
[113] Improved SegFormer-UNet
water segmentation + virtual
water gauge
Error <1 cm in
field tests
Pixel accuracy =
99.10%; IoU =
98.34%
Virtual gauge removes the
need for a physical staff gauge
[115] Canny edge detection +
VGG16 and 2D CNN
– Accuracy ≥98.3%;
CNN = 99.99%
Enhances water boundary vis-
ibility for pixel-level segmen-
tation
[114] Enhanced YOLOv4 (detection)

• DeepLabv3+ (segmentation)
• waterline detection modules
  <1 cm – Addresses inaccurate edge
  localization and low measure-
  ment precision
  Transformer-
  based
  segmentation
  [110] Fixed ROI + SegFormer [ 111],
  SAM [ 23], MobileSAM [ 112] +
  random forest regression
  R2 = 0.93–0.98;
  MAE = 1.03–
  2.33 cm
  IoU ≈ 0.98 (Mobile-
  SAM), 0.97 (SAM),
  0.90–0.93 (Seg-
  Former)
  Maps water pixels within the
  ROI to water level and dis-
  charge under diverse condi-
  tions
  25

7 Imaging Modalities for Water Level Monitoring
Several imaging modalities, including RGB, near-infrared (NIR), an d thermal infrared
(TIR) imaging, have been used to estimate water levels under differ ent environmen-
tal conditions. In addition, UA V, synthetic aperture radar (SAR), multi spectral, and
hyperspectral imagery are commonly used to map water extent. Table 3 lists repre-
sentative studies, the water bodies they addressed, and the imager y and sensors they
used.
7.1 RGB Images from Cameras
Among the different imaging modalities, RGB imagery is the most common for vision-
based water level monitoring, used in the majority of the studies sum marized in
Table 3, due to its low cost, wide availability, and compatibility with both classi-
cal image-processing and deep-learning algorithms. Most current stu dies collect RGB
images using fixed surveillance cameras, closed-circuit televisi on (CCTV) systems,
internet protocol (IP) cameras, or smartphones for continuous monitorin g of rivers,
canals, reservoirs, and urban waterways. Early vision-based monitorin g systems relied
on RGB images of staff gauges, in which the water level was estimated by dete cting the
intersection between the water surface and the graduated ruler usin g image-processing
techniques [ 71, 88, 89].
More recent studies have used higher resolution RGB cameras to enhan ce the visi-
bility of staff-gauge markings and numerical digits and thereby improve m easurement
accuracy. For instance, [ 117] used deep learning for automatic staff-gauge reading
with a fixed RGB camera; [ 65] used RGB images acquired by surveillance cameras to
identify the waterline on staff gauges under different illumination con ditions. High-
resolution RGB imagery supports the extraction of fine-scale features an d can be
used for object detection, semantic segmentation, and optical character recognition
for water-level estimation.
RGB imagery is also used for water level estimation where staff gauges are not
present [ 103, 104, 110]. These methods do not rely on a physical gauge, but estimate
the water level by detecting natural shoreline boundaries or permane nt structures,
such as bridge piers, retaining walls or canal banks, that act as referenc e objects. For
example, [ 73] estimated river levels by segmenting the water surface in river camera
images and relating it to surveyed landmarks, and [ 94] used the surrounding infrastruc-
ture as reference features in live camera images. More recent deep l earning methods
extract water boundaries with semantic segmentation and map them to a vi rtual
gauge, allowing continuous monitoring without installing physical gauges [65, 113].
7.2 Hyperspectral and Multispectral Imagery
Hyperspectral and multispectral imaging techniques have gained inc reasing atten-
tion for waterbody mapping and monitoring [ 118, 119], particularly in regions where
ground-based observations are limited [ 120]. The primary difference between these two
modalities lies in spectral resolution: hyperspectral images captu re data across hun-
dreds of narrow spectral bands, while multispectral images consist of a smaller number
of broader bands.
26

These imaging techniques enable detailed spectral reflectance analy sis, allowing
differentiation between water, vegetation, and soil surfaces [ 121, 122]. Satellite multi-
spectral missions, in particular, provide broad spatial coverage, regu lar revisits, and
long historical archives, making them valuable for large-scale monitorin g of surface
water extent [ 123].
However, these systems have limitations. Hyperspectral sensors i n particular are
expensive and generate large volumes of partly redundant data that requi re careful
calibration. Because both modalities are passive optical systems, the y are also lim-
ited by cloud cover, haze, and the lack of nighttime acquisition [ 123]. Hyperspectral
imagery captures finer spectral variations than multispectral imagery an d can there-
fore discriminate surface materials more precisely, but it does not overcome these
atmospheric limitations [ 124]. In addition, these modalities map water extent rather
than measuring water level directly; level must be inferred by c ombining the mapped
shoreline with topographic data.
7.3 Synthetic Aperture Radar (SAR) Imagery
Synthetic Aperture Radar (SAR) is an active remote sensing technology t hat operates
in the microwave domain and provides high-resolution, all-weather i magery [125, 126].
Unlike optical systems, SAR can acquire data regardless of lighting condi tions or
cloud cover, making it particularly suitable for flood monitoring, wet land analysis,
and inland waterbody mapping [ 127].
SAR imagery is widely used for flood and water extent mapping, especiall y where
optical imagery is limited by cloud cover or low light [ 128]. Data from missions such
as Sentinel-1 are commonly used for water body extraction and flood monit oring
[127, 128], and super-resolution processing of Sentinel-1 data has recentl y been used to
estimate water levels directly [ 16]. However, SAR imagery has several limitations for
water monitoring. Speckle noise reduces the accuracy of water extracti on [ 129], and
wind-roughened water can appear bright while dry sand can appear dark, so a single
global threshold may misclassify water and land [ 130, 131]. In addition, radar shadow
and layover cause errors in urban and vegetated areas [ 132, 133], C-band signals have
limited ability to detect water beneath vegetation [ 134], and narrow rivers may fall
below the sensor’s spatial resolution.
7.4 Near-Infrared and Thermal Infrared Imagery
Infrared imaging is used where visible light imaging becomes unreli able, such as at
night or in poor weather. Two distinct modalities should be distingu ished. Near-
infrared (NIR) imaging (approximately 0.75–1.4 µm) records reflected radiation,
usually with active infrared illumination, and is widely used in cam era-based water
level systems because water strongly absorbs NIR and therefore contrast s with gauges
and banks [ 66, 67, 79]. Thermal infrared (TIR) cameras, in contrast, record radiation
emitted by objects in the 8–14 µm range, representing surface temperature rather
than reflected light [ 135].
Both modalities are deployed in automated monitoring stations and on UA V pl at-
forms. However, thermal imagery generally has lower spatial resolution than RGB or
27

NIR imagery, limiting its ability to resolve fine details such as staff gau ge markings
or small shoreline features, and thermal cameras remain considerably mor e expensive
than conventional RGB cameras.
28

T able 3 Water level and surface water monitoring methods, water bod ies studied, and
sensor/imaging types.
S.No W ater Body References Location Method /
Approach
Image Type Sensor Type
1 River [ 65] Yufeng Creek, Taiwan Deep learning
(SegNet & U-
Net CNNs)
RGB RGB camera
2 River / urban
stream
[94] Buffalo Bayou, USA Fully Convolu-
tional Network
(FCN)
Video Live camera
3 River [ 136] Wesenitz River, Germany CNN (Seg-
Net/FCN) with
SfM photogram-
metry
RGB RGB camera +
SfM
4 Reservoir tank
(testbed)
[137] Japan YOLOv5 object
detection (CNN)
RGB RGB camera
5 River
(complex envi-
ronments)
[68] China Improved
YOLOv5m
(CNN)
RGB RGB camera
6 River [ 138] Missouri River, USA Multi-task CNN
regression
RGB Webcam
7 Reservoir [ 20] Myeonggyeong Reservoir,
South Korea
U-Net seg-
mentation +
CNN/ResNet/EfficientNet
classification
Video CCTV
8 Watershed [ 139] Germany Smartphone
image-based
photogrammet-
ric measurement
RGB Smartphone
camera
9 River [ 140] Multiple rivers, UK and
Ireland
Deep learning
using CNN
RGB RGB camera
10 Reservoir [ 141] Japan WSN-integrated
computer vision
RGB Vision + wireless
sensor network
11 Lakes, rivers,
coastal water
[121] Multiple countries Automated
Water Extrac-
tion Index
(A WEI)
Multispectral Landsat 5 TM
12 Coastal lagoon [ 122] Venice Lagoon, Italy Modified NDWI
(MNDWI) with
sharpened SWIR
band at 10 m
Multispectral Sentinel-2 MSI
13 Global surface
water
[123] Global Expert-system
classification on
Google Earth
Engine
Multispectral Landsat 5/7/8
14 River flood-
plain
[125] River Dee, UK Direct SAR flood
extent mapping
for flood risk
SAR Spaceborne
radar
15 Urban floods [ 126] River Severn, UK SAR segmen-
tation with
LiDAR-based
shadow/layover
simulation
SAR (X-band) TerraSAR-X +
airborne LiDAR
16 River floods [ 127] Evros River, Greece and
Turkey
Automatic
thresholding,
fuzzy-logic
refinement,
HAND masking
SAR (C-band) Sentinel-1
17 Floods [ 128] Italy, Malawi, and Uganda Normalized Dif-
ference Flood
Index (NDFI) on
SAR time series
SAR (C-band) Sentinel-1
18 Global surface
water
[142] Global MNDWI-based
change detection
on Google Earth
Engine
Multispectral Landsat 4/5/7/8
19 Urban stream [ 67] Portugal Image process-
ing for water
level under bad
weather
RGB / NIR Camera with
infrared capabil-
ity
29

8 Datasets for Water Surface and Water Level
Detection
Image-based datasets consist of digital images collected for training, test ing, and eval-
uating deep learning and computer vision models. These datasets typ ically capture
water bodies under varying conditions, often including visible r eferences such as staff
gauges, physical markers, or virtual stations to correlate images with groun d-truth
water level measurements. Table 4 summarizes the image-based datasets reported in
the literature, their key features, and their applications.
Despite their importance, building such datasets remains challen ging. One major
limitation is the annotation process, which is labor-intensive and ti me-consuming,
particularly for pixel-level segmentation or precise waterline ext raction. Another chal-
lenge is ensuring sufficient diversity in the dataset, including variations in location,
water body types, camera viewpoints, and environmental conditions. The lack of
such diversity often leads to poor generalization and limits real-worl d deployment.
Data augmentation, including generative adversarial network (GAN)-based synthesis
of gauge images, has been proposed to partly offset limited training data [ 143]. The
main datasets reported in the literature are described below.
8.1 ATLANTIS Texture (ATeX) Dataset
ATLANTIS (ArTificiaL And Natural waTer-bodIes dataSet) [ 144] is a semantic seg-
mentation benchmark of 5,195 pixel-wise annotated images covering a wide range of
natural and artificial water bodies, collected from online sources throu gh the Flickr
API under permissive licenses and refined through multi-stage filte ring. The derived
ATLANTIS Texture (ATeX) dataset represents the texture appearance of w ater in
15 types of water bodies, such as lakes, rivers, and swamps. It consist s of 12,503
patches of 32 ×32 pixels, split into 8,753 training, 1,252 validation, and 2,498 test
patches. Because ATeX targets water texture classification, it suppor ts water surface
recognition rather than direct water level estimation.
8.2 Eye of Horus Dataset
The Eye of Horus dataset [ 145] contains RGB river images paired with ground-truth
water level measurements. The dataset was collected using a Raspber ry Pi Zero W
and an Arduino Nano system. A Raspberry Pi Camera Module 2 (Sony IMX219, 8
MP) captured images at a resolution of 1920 ×1440 pixels, while water levels were
simultaneously measured using an HC-SR04 ultrasonic sensor. The datase t consists of
1,172 images, split into 812 training, 124 validation, and 236 testing samples.
8.3 Ningbo Staff-Gauge Dataset
The dataset of [ 68] consists of 2,556 images of a staff gauge captured at a single site in
Ningbo, China, at 5-minute intervals over 10 days, of which 1,776 images were used for
training and 780 for testing. The images cover a wide range of conditions, in cluding
sunny (17%), cloudy (16%), rainy (1.4%), night-time (37.9%), soiled gauge (26%) , and
other special situations (1.7%), such as rapid changes in water level and m ud blockage.
30

Images were annotated with separate daytime and night-time gauge classes. Because
all images come from a single site, the dataset is better suited to test ing robustness
under changing illumination and weather than to evaluating cross-site generalization,
and it does not appear to be publicly released.
8.4 Barcode-Patterned Panel Dataset
This dataset [ 146] consists of RGB images of a stainless-steel barcode panel (155 ×
40 cm, with seven black bars) used as a physical marker for water level e stimation,
where the number of bars visible above the water surface indicates t he water level. It
comprises 214 images captured in a pool, where water levels were varied by moving the
panel and artificial rain was simulated in some captures, and 59 images captur ed in
the Mau´ es-A¸ cu River, Brazil. In both settings, images were taken during the daytime
from different positions and angles, and each image is annotated with the num ber of
visible bars and a bounding box around the panel. In the original work, im ages were
resized to 224 × 224 pixels for training.
8.5 River-Camera Dataset for Flood Event Monitoring
This dataset [ 73] was collected using four fixed cameras located along the rivers Sev-
ern and Avon in the United Kingdom, specifically at Diglis Lock (DIGL), Eve sham
(EVES), Strensham Lock (STRE), and Tewkesbury Marina (TEWK). The images
capture a major flood event between November and December 2012. The datas et was
acquired using Mobotix M24 HD cameras at a resolution of 2048 ×1536 pixels and
contains 557 RGB images in total. Each image is annotated with landmarks to enabl e
flood analysis and water level estimation.
8.6 River Segmentation Dataset
The River segmentation dataset [ 147] consists of RGB images with binary water
segmentation masks. It contains 300 manually collected images from riverban k envi-
ronments, with 225 images used for training. This dataset is particularly useful for
direct river segmentation tasks. The image sizes vary, ranging from 118 × 158 to
2448 × 3264 pixels with a mean of approximately 550 × 826 pixels [ 147].
8.7 W ATERDB Dataset
W ATERDB [73] is a large-scale dataset containing RGB images with pixel-wise anno-
tations for water and non-water regions. It is constructed by combining i mages from
the ADE20K dataset [ 148] (water, sea, river, waterfall classes) and the COCO-Stuff
dataset [ 149] (river, sea, and water-other classes). The dataset contains a total of
12,684 training images. As the images are taken from these two source datasets w ith-
out being resampled to a common size, W ATERDB does not have a fixed sp atial
resolution, images from COCO-Stuff are around 640 × 480 pixels, whereas ADE20K
images vary in both size and aspect ratio. Therefore, images are usually res caled during
preprocessing before the segmentation.
31

8.8 Stage-Camera Ephemeral Streams Dataset
This dataset [ 79] consists of images captured using a consumer-grade wildlife camera
(Bushwhacker D3 trail camera) focused on a white-painted steel pole used as a ref-
erence object. Images were collected in time-lapse mode at interval s ranging from 3
seconds to 24 hours. The dataset includes 16-megapixel RGB images and also i ncor-
porates nighttime near-infrared (NIR) imagery (850–940 nm). Images were captu red
every 30 minutes, enabling approximately two weeks of continuous mon itoring before
SD card storage limitations were reached.
8.9 Drone/UA V-Based Image Datasets
8.9.1 Aerial Drone Image Dataset for Flood Monitoring
This dataset [ 150] contains RGB images for object detection and water level estima-
tion. It includes 47 flood images and 5 normal images, along with annotated obje cts
such as 420 houses and 313 cars using bounding boxes and segmentation masks. A
total of 627 object-level samples are used for water level estimation. Wate r level anno-
tations range from 0 to 3 m based on object visibility and inundation depth , and
include data collected from public sources and drone-based aerial image ry to improve
generalization.
8.9.2 UA V Photogrammetry Dataset
This dataset [ 151] consists of UA V-based photogrammetry data collected using drone-
mounted sensors, producing dense point clouds of river surfaces for water level
estimation. The data is processed using Random Sample Consensus (RANS AC) to
fit planar surfaces corresponding to water levels. Ground-truth annot ations include
reference altitude measurements, enabling accurate estimation of w ater surface ele-
vation. This dataset provides a cost-effective alternative to in sit u gauging for flood
monitoring and hydrological analysis.
32

T able 4 Summary of image datasets reported in the literature for wat er surface and water level
detection.
Dataset Reference Size / Split Resolution /
Sensors
Description / Key F eatures
ATLANTIS Tex-
ture (ATeX)
[144] 12,503 patches
(8,753 train /
1,252 val / 2,498
test)
32×32 px Derived from ATLANTIS (5,195 pixel-wise
annotated images of natural/artificial water-
bodies); covers 15 waterbodies (lake, river,
swamp); collected via Flickr API under per-
missive license with multi-stage filtering;
focused on water texture appearance
Eye of Horus [ 145] 1,172 images
(812 train / 124
val / 236 test)
1920×1440 px;
Raspberry Pi
Camera Module
2 (Sony IMX219,
8MP) + HC-SR04
ultrasonic sensor
RGB river images paired with ground-truth
water level measurements; collected using
Raspberry Pi Zero W and Arduino Nano sys-
tem
Ningbo Staff-
Gauge
[68] 2,556 images
(1,776 train /
780 test); single
site
Fixed surveillance
camera
Staff-gauge images captured every 5 min over
10 days in Ningbo, China; sunny, cloudy,
rainy, night-time (37.9%), and soiled-gauge
(26%) conditions; separate day/night annota-
tion classes; not publicly released
Barcode-
Patterned Panel
[152] Not specified RGB camera Images of stainless-steel barcode-patterne d
panels as physical water-level markers; cap-
tured in controlled pool and real river settings
under varying environmental noise; annota-
tions link barcode position to water level
River-Camera
(Flood Event
Monitoring)
[73] 557 images total 2048 ×1536 px;
Mobotix M24 HD
cameras
Collected from four fixed cameras on rivers
Severn and Avon, UK (DIGL, EVES, STRE,
TEWK); captures major flood event (Nov–
Dec 2012); landmark-annotated for flood anal-
ysis and water level estimation
River Segmenta-
tion
[147] 300 images (225
train)
RGB camera Manually collected riverbank images with
binary water segmentation masks; useful for
direct river segmentation tasks
W ATERDB [ 73] 12,684 images RGB camera Large-scale dataset with pixel-wise
water/non-water annotations; combines
ADE20K [ 148] (water, sea, river, waterfall
classes) and COCO-Stuff [ 149] (river, sea,
water-other classes)
Stage-Camera
Ephemeral
Streams
[79] Continuous
time-lapse (˜2
weeks)
16 MP RGB +
NIR (850–940 nm)
Captured using consumer-grade wildlife cam-
era (Bushwhacker D3 trail camera) on a white-
painted steel reference pole; time-lapse inter-
vals from 3 seconds to 24 hours; images every
30 minutes; includes nighttime NIR imagery
Aerial Drone
Image Dataset
(Flood Monitor-
ing)
[150] 52 images
(47 flood, 5
normal); 627
object-level
samples
RGB (drone-
based)
Object detection and water level estimation;
420 houses and 313 cars annotated (bound-
ing boxes + segmentation masks); water level
annotations 0–3 m based on object visibil-
ity/inundation depth; combines public and
drone-collected sources
UA V Photogram-
metry
[151] Not specified
(point cloud
data)
UA V-mounted
sensors
Dense point clouds of river surfaces processed
via RANSAC to fit planar water-level sur-
faces; ground-truth reference altitude mea-
surements; cost-effective alternative to in-situ
gauging
33

9 Discussion
9.1 Methodological Trends
This review has examined how water level can be measured from visual d ata. Water
level monitoring has developed from contact-based sensors, such as cre st-stage gauges,
staff gauges, pressure transducers, and bubblers, toward non-contact m ethods, and
within non-contact methods from ultrasonic, radar, and laser ranging tow ard AI-driven
image analysis. Within computer vision-based studies, a further sh ift is evident from
handcrafted image processing to feature-based machine learning and t hen to deep
learning (Figure 2). These developments share a common motivation: reducing reliance
on physical contact with the water body and on manually engineered featur es, in favor
of approaches that are more durable, scalable, and able to generalize across si tes and
conditions.
The review also clarifies the trade-off between methodological complex ity, robust-
ness, and deployment cost. Classical image processing techniques , including thresh-
olding, multi-frame aggregation, and edge-based detection, remain attracti ve for their
low computational cost and ease of implementation, but their dependenc e on manu-
ally tuned parameters makes them highly sensitive to illumination c hanges, reflections,
and site-specific conditions. Machine learning techniques, such as ensemble feature
extraction, template matching, optical character recognition, and dicti onary learning,
improve robustness over purely rule-based methods by learning relationships from
extracted features, but they still depend on careful preprocess ing. Deep learning meth-
ods, particularly YOLO-based object detection, residual CNNs, and trans former-based
segmentation models such as SAM and SegFormer, largely remove the need f or man-
ual feature engineering by learning hierarchical representations d irectly from imagery.
This has improved both accuracy and adaptability, especially under com plex real-world
conditions involving variable lighting and cluttered backgrounds.
9.2 Performance and Comparability
The reported accuracies of the reviewed methods overlap considerab ly (Tables 1
and 2). Calibration-target and NIR-based image processing systems report mi llimeter-
level resolution and errors of a few millimeters under controlled con ditions [ 19, 66],
whereas threshold-based systems in the field report errors from bel ow 1 cm to more
than 2 cm depending on conditions [ 79]. Deep learning systems typically report errors
of about 1–2 cm in the field [ 68, 105, 110], often under conditions in which image pro-
cessing methods fail, such as low light, strong reflections, or clut tered backgrounds.
Direct comparison of these figures is, however, unreliable. The stu dies use differ-
ent sites, water body types, camera geometries, reference measur ements, and metrics
(MAE, RMSE, R2, NSE, or classification accuracy), and many detection and segmen-
tation studies report only mAP, IoU, or pixel accuracy, which do not trans late directly
into water level error. Very few studies evaluate several method f amilies on the same
dataset. The conclusion that DL outperforms earlier approaches under com plex con-
ditions is therefore supported by the qualitative pattern of result s rather than by
34

controlled head-to-head comparison, and establishing common evaluation pr otocols is
a priority for the field.
9.3 Datasets and Generalization
The review highlights a gap between algorithmic advancement and data inf rastructure.
The datasets described in Section 8, including ATLANTIS and ATeX, Eye of Horus,
GaugeCam, the barcode-patterned panel dataset, the river-camera floo d dataset, the
river segmentation dataset, W ATERDB, the stage-camera ephemeral str eams dataset,
and UA V-based collections, are individually valuable but collectivelyfragmented. They
differ in imaging modality, site characteristics, annotation type, and scale, and most
are limited to a small number of rivers, canals, or reservoirs. This f ragmentation limits
direct comparison across studies, prevents benchmarking under con sistent conditions,
and, most importantly, restricts the generalization of models trained at a single site
when they are deployed elsewhere. Because illumination, shadows , reflections, and
local occlusion vary substantially across locations and seasons, models that perform
well in one environment frequently underperform in another. Datase t diversity and
standardization are therefore a central bottleneck for the field.
9.4 Deployment Considerations
Beyond data availability, several practical factors determine wheth er image-based
methods can be used operationally.
Calibration and geometry . Converting a detected waterline into a physical
water level requires a stable relationship between image coordinate s and elevation.
Gauge-based systems obtain this from the scale markings, whereas gauge-fr ee systems
rely on calibration targets, surveyed ground control points, or homography [ 19, 136].
Camera movement caused by wind, vibration, or maintenance invalidate s this calibra-
tion, so operational systems need either rigid mounting and periodic re-surveying or
automatic detection and correction of camera displacement [ 86].
Environmental conditions. Camera orientation, temporary obstructions, water
turbidity, surface turbulence, rain, fog, and image quality can all degr ade waterline
detection regardless of the underlying algorithm. Night-time operation is a particular
challenge; NIR illumination provides one practical solution, and some gauge -reading
systems perform better at night under infrared illumination than du ring the day [ 64,
79].
Computing and communication. Real-time monitoring at remote sites requires
either transmitting images to a server or processing them on site. Edge computing
and IoT architectures reduce bandwidth needs and latency [ 77, 110], but they con-
strain model size, which favors lightweight models such as MobileS AM or small YOLO
variants over large foundation models.
Reliability and redundancy . Like other non-contact technologies, camera sys-
tems remain susceptible to power interruptions and equipment f ailure. Combining
camera-based estimates with a conventional sensor provides redundan cy and allows
each to check the other [ 152]. Long-term field evaluations remain rare [ 26], and the
35

reviewed studies indicate that fully automatic adaptation to changing fi eld conditions,
without site-specific recalibration, has not yet been achieved.
Application settings. Most reviewed studies address rivers, urban streams, and
reservoirs (Table 3). Agricultural irrigation canals, where low-cost level monitoring
could directly improve water allocation, are rarely studied, even t hough their regular
geometry and fixed structures make them well suited to camera-based m easurement.
This gap between controlled experimental performance and sustained fi eld reli-
ability is one of the clearest barriers separating current research p rototypes from
operational hydrological monitoring networks.
9.5 Future Research Directions
Several research directions follow from these findings. First, the development of large,
openly accessible, multi-site datasets spanning diverse hydraul ic structures, seasons,
and illumination conditions, together with common evaluation protocols, w ould allow
meaningful cross-study comparison and directly address the generaliz ation limita-
tions identified above. Second, vision foundation models, transforme r architectures,
and self-supervised or transfer learning strategies offer a promisin g route to reducing
dependence on large volumes of manually labeled, site-specific data wh ile improv-
ing transferability across monitoring locations [ 24, 25]. Third, multi-sensor fusion, for
example combining camera imagery with pressure transducers at the s ite scale [ 152]
or UA V optical imagery with satellite SAR at the regional scale [ 13], can improve
reliability and extend spatial coverage. Fourth, edge and cloud computi ng architec-
tures could enable real-time, low-latency processing for autonomou s monitoring in
resource-limited or remote settings. In the longer term, research s hould extend single-
purpose water level models toward multi-task computer vision syst ems that jointly
estimate water level, surface velocity, discharge, and debris tr ansport from a single
image stream [ 28, 110], providing a richer basis for flood early warning and water
resources management than water level alone.
10 Conclusions
This review consolidates a rapidly evolving and previously dispers ed body of literature
spanning hydrology, computer vision, and environmental monitoring int o a unified
account of image-based and AI-driven water level estimation. It shows th at the field
has advanced from classical image processing to classical machine learni ng and, most
recently, to deep learning and transformer-based architectures c apable of accurate,
non-contact measurement. At the same time, it identifies dataset fragm entation, the
lack of standardized evaluation, limited cross-site generalization, and sensitivity to
environmental variability as the main obstacles to operational deploym ent. Addressing
these challenges through open and diverse datasets, foundation-model- based learning,
multi-sensor fusion, and edge-enabled real-time processing would move the field from
largely site-specific research toward robust and scalable monitoring s ystems.
36

Funding
This review article is supported by the Utah Agricultural Experimen t Station and the
Utah State University Extension Water Initiative Grant Program under pr oject no.
EXT00185.
References
[1] Le Coz, J., Hauet, A., Pierrefeu, G., Dramais, G., Camenen, B.: Perf ormance of
image-based velocimetry (lspiv) applied to flash-flood discharge me asurements
in mediterranean rivers. Journal of hydrology 394(1-2), 42–52 (2010)
[2] Di Baldassarre, G., Montanari, A.: Uncertainty in river discharge obse rvations:
a quantitative analysis. Hydrology and Earth System Sciences 13(6), 913–921
(2009)
[3] Cr´ etaux, J.-F., Arsen, A., Calmant, S., Kouraev, A., Vuglinski, V., Berg´ e-
Nguyen, M., Gennero, M.-C., Nino, F., Del Rio, R.A., Cazenave, A., et al. : Sols:
A lake database to monitor in the near real time water level and storage vari-
ations from remote sensing data. Advances in space research 47(9), 1497–1507
(2011)
[4] Sauer, V.B., Turnipseed, D.P.: Stage Measurement at Gaging Stations . Tech-
niques and Methods, vol. Book 3. U.S. Geological Survey, Reston, Virginia
(2010). https://doi.org/10.3133/tm3A7
[5] Dou, G., Chen, R., Han, C., Liu, Z., Liu, J.: Research on water-level recognition
method based on image processing and convolutional neural networks. Wate r
14(12), 1890 (2022)
[6] Dedrick, A., Clemmens, A.: Double-bubblers coupled with pres sure transducers
for water level sensing. Transactions of the ASAE 27(3), 779–0783 (1984)
[7] Styles, S., Herman, S., Yasutake, M., Keezer, C.: Water level se nsor testing. Tech-
nical report, Irrigation Training and Research Center, California Poly technic
State University, San Luis Obispo, California (2003)
[8] Delgado, A., Briciu-Burghina, C., Regan, F.: Antifouling strategies f or sensors
used in water monitoring: review and future perspectives. Sensors 21(2), 389
(2021)
[9] Xiscatti, L., Dziedzic, M.: Comparing methods to improve reliab le sensor deploy-
ment time in continuous water quality monitoring. Water supply 20(1), 307–318
(2020)
[10] Mohindru, P.: Development of liquid level measurement technol ogy: A review.
Flow Measurement and Instrumentation 89, 102295 (2023)
37

[11] Matos, T.: Sensor technologies for water velocity, flow, and wave moti on mea-
surement in marine environments: A comprehensive review. Journ al of Marine
Science and Engineering 14(4), 365 (2026)
[12] Darji, K., Patel, D., Dubey, A.K., Gupta, P., Singh, R.: An approach of satellite
and uas based mosaicked dem for hydrodynamic modelling—a case of flood
assessment of dhanera city, gujarat, india. J geomatics, 247–257 (2021)
[13] Fawakherji, M., Hashemi-Beni, L.: Flood detection and mapping th rough multi-
resolution sensor fusion: integrating uav optical imagery and satellite sar data.
Geomatics, Natural Hazards and Risk 16(1), 2493225 (2025)
[14] Fang, C., Duan, K., Lv, Z., Huang, J., Zhong, Q., Chen, J., Long, D.: Impro ving
image-based water-level monitoring by coupling water-line detecti on techniques
and the segment anything model. Environmental Modelling & Software 192,
106566 (2025)
[15] Adikari, H., O’Leary, C., Harrington, J., Lynch, C.: Computer vision in flash
flood forecasting: A narrative review of applications, integration pathw ays, and
future directions. Environmental Modelling & Software, 106910 (2026)
[16] Jia, Y., Liu, Q., He, H., Jin, S., Song, C., Gao, K., Wu, Z.: High accuracy wat er
level estimation using super-resolution sentinel-1 data. Intern ational Journal of
Applied Earth Observation and Geoinformation 146, 105156 (2026)
[17] Owen, R., Wijaya, V., Wardhana, A., Asy’ari, M., Suharjanto, G.: Accur acy
of computer vision hough line transform for water level measurement. In : IOP
Conference Series: Earth and Environmental Science, vol. 1169, p. 012004 (2023).
IOP Publishing
[18] Zhang, Z., Zhou, Y., Liu, H., Zhang, L., Wang, H.: Visual measurement of water
level under complex illumination conditions. Sensors 19(19), 4141 (2019)
[19] Chapman, K.W., Gilmore, T.E., Chapman, C.D., Birgand, F., Mittel stet, A.R.,
Harner, M.J., Mehrubeoglu, M., Stranzl Jr, J.E.: Open-source softw are for water-
level measurement in images with a calibration target. Water Resources Research
58(8) (2022). Article e2022WR033203
[20] Jung, W., Kim, J., Jo, H., Lee, S., Kim, B.: Video-based deep learnin g approach
for water level monitoring in reservoirs. Water 17(17), 2525 (2025)
[21] Muhadi, N.A., Abdullah, A.F., Bejo, S.K., Mahadi, M.R., Mijic, A.: De ep learn-
ing semantic segmentation for water level estimation using surveill ance camera.
Applied Sciences 11(20), 9691 (2021)
[22] Zhou, N., Chen, H., Liu, B., Yi, R., Wei, M., Xu, C.: Intelligent wate r level
recognition method integrating semantic segmentation with spatiotemp oral
38

information. Journal of Hohai University (Natural Sciences) 54(1), 18–27 (2026)
[23] Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T.,
Whitehead, S., Berg, A.C., Lo, W.-Y., et al. : Segment anything. In: Proceedings
of the IEEE/CVF International Conference on Computer Vision, pp. 4015–4026
(2023)
[24] Hu, S., Wang, Z., Fan, X., Zhang, C.: Mitigating site-specific data dep endency
in image-based water level estimation using vision foundation models. Water
Resources Research 62(5) (2026). Article e2025WR042686
[25] Wu, Z., Wen, Y., Huang, K., Zhou, N., Chen, H.: Intelligent water leve l
measurement based on visual foundation models. Measurement, 118502 (2025)
[26] Blanch, X., Grundmann, J., Hedel, R., Eltner, A.: Ai image-based met hod for a
robust automatic real-time water level monitoring: a long-term applicati on case.
Hydrology and earth system sciences 30(3), 797–824 (2026)
[27] Wu, Z., Huang, Y., Huang, K., Yan, K., Chen, H.: A review of non-contact
water level measurement based on computer vision and radar technology. Wat er
15(18), 3233 (2023)
[28] Manfreda, S., Miglino, D., Saddi, K.C., Jomaa, S., Eltner, A., Perks , M., Pe˜ na-
Haro, S., Bogaard, T., Van Emmerik, T.H., Mariani, S., et al. : Advancing
river monitoring using image-based techniques: challenges and oppor tunities.
Hydrological Sciences Journal 69(6), 657–677 (2024)
[29] Kossieris, S., Tsiakos, V., Tsimiklis, G., Amditis, A.: Inland wat er level monitor-
ing from satellite observations: A scoping review of current advances and future
opportunities. Remote Sensing 16(7), 1181 (2024)
[30] Page, M.J., McKenzie, J.E., Bossuyt, P.M., Boutron, I., Hoffmann, T .C., Mul-
row, C.D., Shamseer, L., Tetzlaff, J.M., Akl, E.A., Brennan, S.E., e t al.: The
PRISMA 2020 statement: an updated guideline for reporting systematic r eviews.
BMJ 372 (2021) https://doi.org/10.1136/bmj.n71 . Article n71
[31] Friday, J.: The operation and maintenance of a crest-stage gaging station .
Technical report, US Geological Survey (1965)
[32] Keeland, B., Dowd, J., Hardegree, W.: Use of inexpensive pressur e transducers
for measuring water levels in wells. Wetlands Ecology and Management 5(2),
121–129 (1997)
[33] Buchanan, T.J., Somers, W.P.: Stage measurement