论文信息
论文解读
AI 辅助整理- 方法
- 综述高光谱反射数据结合植被指数、特征选择与机器学习/深度学习分类。
- 发现
- 高光谱可无损高精度识别病害,但成本高、数据冗余、标注稀缺限制泛化。
- 机会点
- 可探索多传感器融合与跨作物跨环境迁移学习,缓解标注稀缺与模型泛化难题。
论文摘要
植物病害是农业生产中不可避免的挑战,但及时的干预和有效的管理措施能够显著防止作物损失,维持产量和品质。传统诊断方法依赖视觉评估和劳动密集、耗时的物理样本采集,难以满足大规模作物监测的需求,且往往仅在可见症状出现后才能检测到病害,严重制约了早期病害检测的潜力。这些局限性迫切需要无偏倚的自动化病害监测与诊断方法。高光谱遥感通过测量可见光、近红外(NIR)和短波红外(SWIR)波段内数百个连续波长的光谱反射率,代表了一种革命性技术,能够实现非侵入式、空间全面的作物健康评估。病害相关的生理和代谢变化,如叶绿素损失、水分胁迫和色素改变,会产生独特的光谱特征,可被高分辨率高光谱传感器检测到。高光谱数据分析采用光谱植被指数(VIs)、特征选择以及机器学习/深度学习模型,实现自动化、高精度的病害分类,较传统田间巡查显著提高了检测效率。然而,该技术的实际部署受到传感器成本高、数据量大、光谱冗余、对光照和校准变异性敏感以及大规模标记数据集稀缺等因素的制约,限制了模型在不同作物和环境间的泛化能力。通过多传感器集成将高光谱数据与其他互补遥感技术相结合,可提高植物病害检测的精度和空间分辨率。本文综述了用于植物病害检测的高光谱数据采集、预处理和分析技术、当前方法、应用及研究空白,将高光谱遥感定位为下一代可持续农业的基础性技术。
全文
Plant Science TodayPLANT SCIENCE TODAY
Vol 13(sp9): 01-12
https://doi.org/10.14719/pst.17031
eISSN 2348-1900
Review Article | Agriculture 2026, Volume 13, Issue sp9, pp. 01-31
Plant disease detection and recognition using
hyperspectral remotely sensed data: A review
Vishalini Sivasubramani1, Kumaraperumal Ramalingam1*, Jagadeeswaran Ramasamy1, Sivasubramanian
Karuppusamy2, Kavitha Shanmugam3, Nivas Raj Moorthi1 & Sai Divya Sri Jagarlamudi1
1Department of Remote Sensing and Geographic Information System, Tamil Nadu Agricultural University, Coimbatore 641 003,
Tamil Nadu, India
2Department of Agronomy, Tamil Nadu Agricultural University, Coimbatore 641 003, Tamil Nadu, India
3Centre for Students Welfare, Tamil Nadu Agricultural University, Coimbatore 641 003, Tamil Nadu, India
*Correspondence: Kumaraperumal Ramalingam, kumaraperumal.r@tnau.ac.in
Abstract
Plant diseases are an inevitable challenge in agricultural production, yet timely intervention and effective management
practices can significantly prevent crop losses and sustain yield and quality. Conventional diagnostic approaches depend on
visual evaluation and labour-intensive, time-consuming physical sample collection, making them inadequate for large-scale
crop monitoring and often detecting diseases only after visible symptoms appear, considerably hampering the potential for
early disease detection. These limitations necessitate unbiased and automated approaches to disease monitoring and
diagnosis. By measuring spectral reflectance across hundreds of continuous wavelength bands in the visible, near-infrared
(NIR) and shortwave infrared (SWIR) domains, hyperspectral remote sensing represents a revolutionary technique facilitating
non-invasive, spatially comprehensive crop health assessment. Disease-related physiological and metabolic changes such
as chlorophyll loss, water stress and pigment alteration generate distinctive spectral signatures detectable by hyperspectral
sensors with high resolution. Hyperspectral data analysis employs spectral vegetation indices (VIs), feature selection and
machine/deep learning models to enable automated, high-accuracy disease classification, substantially improving detection
efficiency over conventional field scouting. However, the technology’s practical deployment is constrained by high sensor
costs, large data volumes, spectral redundancy, sensitivity to illumination and calibration variability and the scarcity of large,
labelled datasets, which limit model generalisability across crops and environments. Combining hyperspectral data with other
complementary remote sensing techniques through multisensor integration enhances the accuracy and spatial resolution of
plant disease detection. This review elucidates hyperspectral data acquisition, preprocessing and analytical techniques for
plant disease detection, current methodologies, applications and research gaps, positioning hyperspectral remote sensing as
a foundational technology for next-generation sustainable agriculture.
Keywords: disease classification; dimensionality reduction; feature extraction; machine learning; spectral reflectance; vegetation
indices
Received: 26 July 2026. Accepted: 16 September 2026. Available online: 16 November 2026.
Cite this article: Vishalini S, Kumaraperumal R, Jagadeeswaran R, Sivasubramanian K, Kavitha S, Nivas R M, Sai Divya SJ.
Plant disease detection and recognition using hyperspectral remotely sensed data: A review . Plant Science Today.
2026;13(sp9):01-31. https://doi.org/10.14719/pst.17031
Introduction
Plant diseases pose an alarming threat to worldwide
agricultural production and food security. The
distribution, host range and consequences of plant
diseases are increasing due to global climate change
and the rapid expansion of international trade.
Diseases that have already been controlled may re-
emerge (1). To fulfil the increasing demand for both the
quantity and quality of food, plant protection is
necessary, particularly for preventing crops from
diseases.
Nearly 10–40 % of global production is lost
due to plant diseases. Detection and diagnosis of
diseases have been accomplished using conventional
techniques such as microscopy, culturing,
immunoassays and molecular analyses, which are
complicated, time-consuming and constrained to
centralised laboratories. Most of these techniques
suffer from low sensitivity and specificity and thus fail
to combat emerging plant diseases. The existing
techniques used for detection normally require the crop
agronomist to conduct a visual inspection for
symptoms that have appeared by physically examining
the crop. Given the fact that many commercially grown
crops cover a vast area, monitoring crop growth using
this technique is highly labour-intensive. The manual
detection of any diseases requires visible symptoms,
which only appear when the disease is at an advanced
stage. Highly sensitive, rapid and economical methods
Plant Science Today | 2026 | Volume 13(sp9) | Page 2
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
of detecting and diagnosing plant diseases have
become an urgent need. Recent developments in
computer science and biotechnology have contributed
to developing new technologies for disease detection.
Among these, biological sensing and remote imaging
methods appear to be the most promising ones (2, 3).
Plant disease detection has increasingly relied
on non-invasive optical sensors such as hyperspectral
imaging. The use of such technology is more accurate
than traditional methods in both precision farming and
resistance breeding and its applications range from
laboratory testing to field studies. Specifically,
hyperspectral imaging is important for the diagnosis of
diseases as it offers advantages at various levels,
starting from individual plant testing to canopy-level
assessment (4).
Hyperspectral imaging (HSI) has become an
important technique for precision agriculture
applications compared to standard imaging methods
that employ multispectral or visible spectrum (RGB)
data. Hyperspectral imaging can identify plant stress,
disease and nutritional deficiencies by examining each
crop's distinct spectral signature. This information can
be utilised to maximise crop productivity and lower
input costs (5). By utilising machine learning (ML)
methods and algorithms combined with hyperspectral
images, diseases can be identified. A broad range of
wavelengths can be observed in hyperspectral images.
Cereals, oilseeds, fruits, vegetables and other
agricultural crops can be monitored using these
wavelengths, among other applications. Crop
parameters such as nutritional status and maturity
index, as well as the early identification of some
diseases that result in crop losses, can be examined
and diagnosed within these wavelength ranges (6).
Plant disease detection starts with image
acquisition and is accompanied by segmentation-
based pre-processing. This is further done in
association with various methods of feature extraction
as well as classification (7). Machine learning and
deep learning algorithms are advancing the process of
image classification-based disease identification using
artificial intelligence (AI) (8). The aim of this review is to
provide a comprehensive overview of the application of
hyperspectral remotely sensed data for plant disease
detection, covering several aspects of hyperspectral
remote sensing such as sensing principles, sensors,
spectral features, pre-processing methods, feature
extraction techniques, vegetation indices (VIs) and ML
methods. The following sections discuss these aspects
in detail.
Materials and Methods
Plant diseases pose an alarming threat to worldwide
agricultural production and food security. The
distribution, host range and consequences of plant
diseases are increasing due to global climate change
and the rapid expansion of international trade. Diseases
that have already been controlled may re-emerge (1).
To fulfil the increasing demand for both the quantity and
quality of food, plant protection is necessary, particularly
for preventing crops from diseases.
Nearly 10–40 % of global production is lost due
to plant diseases. Detection and diagnosis of diseases
have been accomplished using conventional techniques
such as microscopy, culturing, immunoassays and
molecular analyses, which are complicated, time-
consuming and constrained to centralised laboratories.
Most of these techniques suffer from low sensitivity and
specificity and thus fail to combat emerging plant
diseases. The existing techniques used for detection
normally require the crop agronomist to conduct a visual
inspection for symptoms that have appeared by
physically examining the crop. Given the fact that many
commercially grown crops cover a vast area, monitoring
crop growth using this technique is highly labour-
intensive. The manual detection of any diseases
requires visible symptoms, which only appear when the
disease is at an advanced stage. Highly sensitive, rapid
and economical methods of detecting and diagnosing
plant diseases have become an urgent need. Recent
developments in computer science and biotechnology
have contributed to developing new technologies for
disease detection. Among these, biological sensing and
remote imaging methods appear to be the most
promising ones (2, 3).
Plant disease detection has increasingly relied
on non-invasive optical sensors such as hyperspectral
imaging. The use of such technology is more accurate
than traditional methods in both precision farming and
resistance breeding and its applications range from
laboratory testing to field studies. Specifically,
hyperspectral imaging is important for the diagnosis of
diseases as it offers advantages at various levels,
starting from individual plant testing to canopy-level
assessment (4).
Hyperspectral imaging (HSI) has become an
important technique for precision agriculture applications
compared to standard imaging methods that employ
multispectral or visible spectrum (RGB) data.
Hyperspectral imaging can identify plant stress, disease
and nutritional deficiencies by examining each crop's
distinct spectral signature. This information can be
utilised to maximise crop productivity and lower input
costs (5). By utilising ML methods and algorithms
combined with hyperspectral images, diseases can be
identified. A broad range of wavelengths can be
observed in hyperspectral images. Cereals, oilseeds,
fruits, vegetables and other agricultural crops can be
monitored using these wavelengths, among other
applications. Crop parameters such as nutritional status
and maturity index, as well as the early identification of
some diseases that result in crop losses, can be
examined and diagnosed within these wavelength
ranges (6).
Plant Science Today | 2026 | Volume 13(sp9) | Page 3
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
Plant disease detection starts with image
acquisition and is accompanied by segmentation-based
pre-processing. This is further done in association with
various methods of feature extraction as well as
classification (7). Machine learning and deep learning
algorithms are advancing the process of image
classification-based disease identification using artificial
intelligence (AI) (8). The aim of this review is to provide
a comprehensive overview of the application of
hyperspectral remotely sensed data for plant disease
detection, covering several aspects of hyperspectral
remote sensing such as sensing principles, sensors,
spectral features, pre-processing methods, feature
extraction techniques, Vls and ML methods. The
following sections discuss these aspects in detail.
Scope and methodology of this review
Literature search strategy
Taking into account the multi-stage pipeline of
hyperspectral remote sensing for plant disease
identification, the review has been structured around
eight thematic strands spanning sensing, preprocessing,
feature extraction, VIs, machine learning, deep learning,
sensor integration and future prospects. The
identification and organisation of these themes were
based on the literature assessed through the various
database searches conducted for this review.
Google Scholar, Scopus, Web of Science, Elicit
and Institute of Electrical and Electronics Engineers
(IEEE) Xplore were employed to obtain research on
hyperspectral remote sensing applications for plant
disease detection and recognition, spanning studies
published from 2015–26. Google Scholar was used for
broad multidisciplinary coverage across agriculture and
remote sensing journals; Scopus was used for its
extensive indexing of peer-reviewed journal articles
across agricultural, remote sensing and computer
science domains; Web of Science was used to ensure
coverage of high-impact, citation-indexed research and
to support backward citation tracing of foundational
studies in the field; Elicit was used for AI-assisted
semantic search and rapid screening of relevant studies
across published research; and IEEE Xplore was used
to capture engineering and sensor-focused literature on
hyperspectral instrumentation and data-driven
classification methods.
Search query
The literature search utilised keyword-based queries
across the title, abstract and keywords fields, structured
around the eight thematic pillars of this review:
Google scholar and elicit search queries
“Hyperspectral disease detection”; “preprocessing of
hyperspectral data”; “feature extraction hyperspectral”;
“VIs for disease detection”; “ML disease detection
hyperspectral”; “deep learning disease detection
hyperspectral”; “integration of remote sensing sensors
plant disease”; “future scope hyperspectral plant
disease detection”.
Institute of Electrical and Electronics Engineers Xplore
search query
(“hyperspectral imaging” OR “hyperspectral remote
sensing”) AND (“plant disease” OR “crop disease” OR
“plant stress detection”) AND (“machine learning” OR
“deep learning” OR “feature extraction” OR
“preprocessing” OR “sensor fusion”).
Each keyword group was searched
independently and, where relevant, combined (e.g.,
“hyperspectral” + “deep learning” + “disease detection”)
to capture literature specific to each subsection of the
review.
Study selection and screening
The preliminary search resulted in 163 papers, which
were screened for relevance using the following
criteria:
Inclusion criteria
● Research on hyperspectral (imaging or non-imaging)
sensing for plant/crop disease or stress detection.
● Studies of spectral data employing machine learning,
deep learning, feature extraction, or preprocessing
techniques; empirical or review articles, peer-
reviewed, published in English, from 2015 until 2026.
Exclusion criteria
● Duplicate or irrelevant studies.
● Studies based on only RGB or multispectral data
without a hyperspectral component.
● Application studies for non-agricultural/plant
diseases.
After screening, 100 articles were selected for in-depth
review. Among them, 11 articles concern sensor types
and data acquisition, including instruments for imaging
and non-imaging hyperspectral collection; 16 articles
address preprocessing and feature extraction
methods; 18 articles address VIs and spectral
signatures for disease detection; 30 articles address
ML and deep learning classification/regression
approaches; and 25 articles emphasise sensor fusion,
integration with other remote sensing modalities and
future prospects in precision agriculture. The article
selection process followed the preferred reporting
items for systematic reviews and meta-analyses
(PRISMA) framework, as shown in Fig. 1.
Hyperspectral remote sensing
Remote sensing technologies fall into two categories:
- passive (optical) and active (LiDAR and radar). The
spectral resolution of the sensors used determines
whether passive optical remote sensing is typically
separated into two categories: multispectral and
hyperspectral remote sensing. Among the methods of
passive remote sensing that quantify reflected solar
radiation or an artificial light source, hyperspectral
sensing shows significant promise as a non-
destructive, non-invasive method of monitoring biotic
and abiotic plant stress (9).
Plant Science Today | 2026 | Volume 13(sp9) | Page 4
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
Fig. 1. PRISMA flowchart depicting the article selection process used in this review.
Plant Science Today | 2026 | Volume 13(sp9) | Page 5
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
By capturing hundreds of continuous spectral
bands, hyperspectral remote sensing, also referred to
as imaging spectroscopy, provides comprehensive
spectral reflectance data of imaged targets.
Thousands of bands that represent various
types of light energy make up the electromagnetic
spectrum. Imaging spectrometers (instruments that
capture hyperspectral data) divide the electromagnetic
spectrum into groups of bands that assist in the
classification of objects by their spectral qualities on the
target’s surface. Farmers can identify plant health issues
and take preventative action before they become
serious issues due to their capacity to record many
wavelength bands and high-resolution images that
provide adequate crop information (6).
Hyperspectral remote sensing employs a wide
spectrum of wavelengths, spanning from the visible
near-infrared (VNIR) to the shortwave infrared (SWIR)
range of 400–2500 nm, facilitating a non-invasive
approach to identify molecule reflection and absorption
(10).
Hyperspectral images, comprising numerous
narrow spectral bands, provide an accurate spectral
reflectance curve capable of detecting minor changes in
soil, canopies or individual leaves, facilitating the early
identification of plant diseases through slight variations
in spectral reflectance (9). In order to accurately and
promptly determine the physiological status of
agricultural crops, a broader range of problems can be
solved using hyperspectral images. Timely detection of
disease outbreaks and their spread may avoid
significant agricultural losses, reduce pesticide
application and mitigate their harmful impacts on the
ecosystem and public health.
Hyperspectral versus multispectral imaging
Researchers have been using multispectral imagery
(images including 3–20 discrete bands that are not
contiguous) to identify diseases in crops for years.
Because hyperspectral imaging can capture continuous
spectral information spanning the visible, red edge and
near-infrared (NIR) plateau regions, it has proven to be
clearly superior to multispectral systems for the
detection of crop diseases like wheat leaf rust.
Hyperspectral systems may identify minute biochemical
alterations in infected plant tissue before obvious
symptoms appear, in contrast to multispectral sensors
that are restricted to a few distinct wavebands (11).
By employing ML and deep learning
approaches to process these images, they have
discovered new findings that lead to more efficient
analysis methodologies and save time (12). Despite this
advantage, hyperspectral imaging is not always the best
choice. Multispectral systems remain sufficient when the
target disease produces distinct, well-characterised
spectral signatures detectable within a few known
bands, providing useful benefits in terms of cost, data
volume, processing speed and compatibility with current
satellite platforms (e.g., Sentinel-2, Landsat).
Multispectral imaging is preferred for large-scale, time-
sensitive, orcost-constrained monitoring where
symptoms are already spectrally distinguishable,
whereas hyperspectral imaging is most justified for
early, asymptomatic infection stages or diseases with
subtle biochemical signatures requiring fine spectral
resolution (often <10 nm bandwidth).
Hyperspectral sensors and data acquisition
The initial and most critical step in analysing plant
diseases is acquiring high-quality hyperspectral images.
Achieving accurate results depends on selecting
appropriate sensors and platforms, properly configuring
spatial and spectral resolution, designing an effective
lighting setup and correctly adjusting scan rate, frame
rate and exposure time (13).
Three bands of the visual portion of the
electromagnetic spectrum-red, green and blue light-are
measured by conventional RGB cameras, while
multispectral sensors measure the spectral information
of objects across several discrete bands spanning the
visible, NIR, red-edge and SWIR regions. Hyperspectral
sensors capture several hundred bands in the
electromagnetic spectrum (spectral range of 400–2500
nm and with a narrow spectral resolution of less than 1
nm). Each of these spectral bands only captures a few
nanometres of the electromagnetic spectrum, so the
hyperspectral sensor has a high spectral resolution.
There are two types of sensors: imaging sensors and
non-imaging sensors (14, 15).
Non-imaging sensor
The average reflectance spectrum across a specified
surface area is determined by the non-imaging sensor
without retaining any spatial information, where
variables like focal length, field of view and distance to
the object define the size of the measured region. These
sensors are generally portable, lightweight and do not
need complex mounting platforms, making them
practical for field use. These sensors have a high
spectral resolution of 1–3 nm and a spectral range of
300–2500 nm.
Spectrometers like ASD FieldSpec, which are
used to identify Fusarium head blight in wheat, are
among the most widely used non-imaging tools (16),
spectral vegetation characteristics (SVC), used in
detecting potyvirus potato virus Y (PVY) in potato plants
(17), ImSpector, used in detecting wheat rust (18) and
STS-VIS spectrometer suitable for use on unmanned
aerial vehicles (UAVs) (19). These spectrometers have
been widely used in labs, greenhouses and in the field
for assessing plant health and diseases. Compared to
imaging sensors at similar distances, non-imaging
sensors have a poorer spatial resolution since their
spectral data is averaged over a certain region inside
their range of view. As a result, despite the limitation of
using non-imaging spectrometers in detecting diseases,
they can potentially detect pre-symptomatic
physiological changes if the disease-induced spectral
changes are detectable within their measured field of
Plant Science Today | 2026 | Volume 13(sp9) | Page 6
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
view, since the output signal from the spectrometer
reflects the average spectrum of healthy and diseased
tissues in the sample area (14). However, these sensors
are suitable for rapid evaluation of plants in fields and
greenhouses (15).
Non-imaging hyperspectral signatures acquired
using VIS-NIR spectrometer effectively differentiated
healthy from diseased wheat leaves and ears (Septoria
tritici blotch and Stagonospora nodorum blotch) by
analysing spectral reflectance patterns, particularly in
the red edge region where quantile spread increased
systematically with disease severity (20).
Imaging sensors
Pixel-wise spectral identification is made possible by a
transition from a non-imaging spectroradiometer to
hyperspectral imaging, making it possible to detect
subtle modifications characteristic of early-stage plant
diseases. Spectral signature changes in host tissue can
be identified not only at primary symptom sites and
across subsequent disease stages, but also during pre-
symptomatic phases and for localised tissue effects,
thereby providing valuable insights into the underlying
process of plant biology and fungal pathogenesis (21).
The imaging hyperspectral sensors identify the
spectral bands for each pixel of the acquired image by
combining spectral and spatial resolutions. The
reflectance values for each spectral band that the
hyperspectral sensor measured make up each pixel's
own spectral signature in the image. Consequently, the
final image is a hyperspectral data cube with two spatial
dimensions and one spectral dimension (15).
Imaging hyperspectral sensors are preferable
over non-imaging hyperspectral sensors as their spatial
dimension gives additional details on gradient, colour
and shape, so imaging hyperspectral sensors can be
applied for site-specific and targeted field management,
abiotic and biotic stress screening and thus support
sustainable agriculture (16).
Depending on the sensors' configuration, a
hypercube can be acquired by four different methods,
namely whisk-broom, push-broom, spectral scanner and
hyperspectral snapshot (13). Hyperspectral imaging
sensors often only cover a narrow spectral range, like
SWIR (1000–2500 nm) or VNIR (300–1000 nm). The
spectral resolution of these sensors ranges from 1–7
nm. The distance between the object and the sensor
determines the spatial resolution and ranges from
micrometres to centimetres.
Whisk-broom and push-broom cameras
represent the most frequently used point scanning and
linear scanning sensors in hyperspectral imaging
research. These scanning sensors require either the
object or the camera to be in motion during the
acquisition of images or data from scenes. Some
examples of popular scanning cameras commonly
employed in scientific research are Specim, Headwall,
Photonfocus and Pika L and these devices are usually
mounted on specific moving platforms for linear motion
and stabilisation in laboratory environments. Based on
the size of the measured area, these scanning sensors
have a considerable image acquisition time, which
makes it difficult to capture moving objects (14).
Liquid crystal tunable filter (LCTF) pass filters,
which only allow specific wavelengths to pass through,
are employed by spectral scanning sensors to obtain
two dimensional (2D) spatial images for each
wavelength without the need for a moving object or
sensor in order to obtain a hypercube. Hence,
acquisition is faster than other scanning methods like
point and line scan, but due to the difference in timing of
capturing the bands, inconsistency may arise when the
object is in motion (22).
Snapshot sensors operate using a mosaic
approach that is similar to RGB camera sensors but
capture the entire hypercube instantly instead of using a
scanning process, resulting in high recording rates but
with low spatial resolution. It can be used in UAV due to
their small size and shorter acquisition time. Some
snapshot sensors used are Rikola Fabry-Perot
interferometer (FPI)-based snapshot hyperspectral
camera mounted on a UAV for potato late blight
detection (23). Cubert UHD 185 FireFly snapshot
hyperspectral camera mounted on a UAV for vegetation
monitoring in a barley experiment (24).
Hyperspectral sensor setup
A hyperspectral sensor, an artificial or natural light
source and a control device for measuring and storing
hyperspectral images make up a typical hyperspectral
sensor configuration. In addition to being handheld,
these parts can be installed on a variety of carrier
platforms, including automobiles, tractors, UAVs, robots,
drones, aeroplanes and satellites. The appropriate
wavelength range and object of interest are taken into
consideration when choosing an imaging detector.
Charge-coupled devices built of silicon and
complementary metal-oxide semiconductors are the
most widely used detectors for visible and near-infrared
wavelengths (25). These are used in several sensing
types, including full-frame cameras, push-broom
cameras and whisk-broom scanners. One dimension of
spectral information and two dimensions of spatial
information are provided by the hyperspectral data cube
that contains the data acquired through HSI (16).
In fact, the "Hypercart," a ground-based
proximal sensing platform, was created to obtain high-
resolution hyperspectral images of potato crops in field
settings. The Hypercart setup included an aluminium
framework with a motor-driven translational stage
equipped with hyperspectral sensors and 18 halogen
spotlights to ensure controlled and uniform illumination.
Crop measurements were collected at a height of 0.60
m from the crop surface, yielding a spatial resolution of
0.30 mm per pixel. The hyperspectral imaging system
was based on a push-broom system integrated with
prism-grating-prism (PGP) spectroscopy system
(Imspector V9, Specim) covering a wavelength range
from 430–900 nm at a 7 nm spectral resolution together
Plant Science Today | 2026 | Volume 13(sp9) | Page 7
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
with a high-speed camera (Pike F421B). For capturing
images, the Hypercart was wrapped in black cloth in
order to prevent exposure to external illumination. This
was done in an effort to reduce the impact of different
lighting conditions on the surroundings while collecting
data. With this configuration, Alternaria solani lesions in
potato crops could be accurately detected and mapped
using high-resolution hyperspectral imaging (26).
The push-broom sensors are highly spectrally
accurate, but they are susceptible to any movement on
the side and lighting variations. The snapshot sensors
are preferred for real-time or handheld field applications
but are more practical in terms of speed and
robustness. The Hypercart/ground-based platforms are
the most controllable and have the best resolutions but
are not practical for large-scale monitoring.
Although higher spectral resolution increases
biological sensitivity, it frequently comes at the expense
of processing time, data volume and spatial coverage-a
limitation for large-scale or real-time deployment. On the
other hand, systems with coarser spectral but finer
spatial resolution prioritise field usability and scalability
over detection sensitivity, suggesting that there isn't a
single configuration that works well in every deployment
scenario.
Spectral characteristics of healthy and diseased
plants
Light striking plant tissue is reflected, scattered,
absorbed, or transmitted depending on the wavelength.
In healthy plants, this spectral reflectance pattern is
shaped by biochemical composition and its internal leaf
structure (27). Photosynthetic pigments (carotenoids,
anthocyanins and chlorophylls) absorb the majority of
light in the visible wavelength range (400–700 nm),
producing low reflectance except for the 550 nm green
reflection peak in healthy plants. The leaf's cellular
structure determines reflection in the NIF (700–1100
nm) range and scattering by air pockets between plant
cells will result in increased reflectance. In the SWIR
(1100–2500 nm), reflectance mainly depends on water
content, proteins and carbonaceous materials in the
plant (3).
Physiological stress from pests, pathogens,
drought and nutrient deficiency disrupts the plant’s
internal chemistry, altering leaf coloration, leaf size, leaf
surface water evaporation and canopy architecture.
These changes shift reflectance and absorption
patterns, producing measurable spectral differences
between normal and stressed plants (28).
Plant-pathogen interactions are characterised
by distinct spatial and temporal dynamics, each
influencing different portions of the electromagnetic
spectrum in specific ways. Comprehending the cause-
and-effect relationships between plant physiological
changes and associated spectral responses offers a
useful foundation for examining plant biochemistry and
creating focused experimental studies (1).
Hyperspectral imaging proves to be very
successful in detecting such minor spectral reflectance
differences and the resulting data has enabled several
ML algorithms for automated disease classification. This
involves feature extraction from the spectral data,
training the classification algorithm using samples of
healthy and infected plants and then applying the
trained model to the new data for disease prediction.
One common feature extraction approach relies on
calculating VIs tied to specific plant characteristics;
however, most such indices are not specifically
designed to distinguish between healthy and diseased
states of plants (29).
Pre-processing of hyperspectral data
Enhancing contrast, removing noise, lowering data
dimensionality and computational complexity and
ensuring data accuracy are the primary goals of data
pre-processing. Numerous elements, such as sensor
noise and error, platform stability, environmental factors,
light source and background conditions, affect the
quality of hyperspectral data. Preprocessing is used to
eliminate distortions such as curvature effects, sensor
noise and other instrumental errors that may arise
during data acquisition (14). Hyperspectral reflectance
signals of non-imaging sensors can be easily pre-
processed. The direct deletion technique is used in
numerous studies to eliminate abnormal wavebands.
Spectral smoothing and filtering are also used in pre-
processing. Geometric correction, atmospheric
calibration and spectral calibration are pre-processing
techniques used in hyperspectral data imaging (1, 30).
Atmospheric correction
Atmospheric calibration is mostly used in satellite-scale
data analysis due to the surface reflectivity inaccuracy
caused by air scattering and absorption (1). This step is
used for imaging hyperspectral data.
Geometric correction
For field, low-altitude, aerial and satellite-scale data
processing, geometric correction is mostly necessary
due to the geometric distortion caused by platform tilt,
terrain changes, etc. (1). This step is carried out for
imaging hyperspectral data. It is used to correct image
distortion by transforming each spectral channel to a
standard reference axis. This process aligns the image
with real-world coordinates, such as map coordinates.
The most common method is ground control point
(GCP)-based correction. Ground control points are
common points found in both the distorted image and
the reference map. During correction, control points are
selected and interpolation is performed to align the
image correctly. Resampling methods are used along
with geometric correction to assign new pixel values.
Common resampling methods include:
● Nearest neighbour: determines the value of the
closest pixel.
Plant Science Today | 2026 | Volume 13(sp9) | Page 8
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
● Bilinear interpolation – calculates pixel values using
nearby pixels.
● Bicubic interpolation – uses surrounding pixels to
produce smoother results.
Geometric correction improves the spatial
accuracy of hyperspectral images. It helps reduce errors
caused by distortion, sensor movement and scanning
conditions (31).
Spectral correction
Spectral calibration is necessary for hyperspectral data,
both imaging and non-imaging and is a process which
involves complex algorithms (16). Methods such as
radiation transfer models, spectral interpolation and
spectral normalisation are employed (1). Image pre-
processing typically consists of image masking, contrast
enhancement and edge detection. For plant disease
identification and image analysis, image enhancement
may be the most important stage. Eliminating data noise
and emphasising the qualities of the desired attributes
are its main goals. Numerous methods, including
spectral smoothing (moving average smoothing,
minimum noise fraction rotation, Savitzky–Golay
smoothing, etc.), data filtering (Gaussian filtering, linear
filtering, etc.) and image enhancement (histogram
equalisation, homomorphic filtering, etc.), can be
employed to accomplish these goals (1). Other methods
like Fourier transform and wavelet transform are also
used to reduce noise. Baseline shifts in the spectrum
are corrected using the first and second derivatives. To
reduce spectral variation brought on by scattering,
standard normal variate (SNV) and multiplicative scatter
correction (MSC) are employed (14).
Savitzky-Golay filter or digital smoothing
polynomial filter (DISPO) is a digital filter extensively
used for data smoothing and differentiation purposes. It
smooths digital data points by fitting segments of
adjacent data points into a polynomial function of low
degree using the linear least squares convolution,
improving accuracy without distorting the tendency of
the signal. This algorithm entails establishing a
constant-sized and odd-numbered window, choosing
the degree of the polynomial to be fitted, applying least
squares fit locally within the window and calculating
values of the polynomial at the middle of each window,
thus generating the output. Savitzky-Golay filters are
ideal for reducing least squares error, so efficiently
conserving high-frequency signal components, in
contrast to traditional finite impulse response (FIR) filters
that average and eliminate high-frequency material
along with noise. However, this filter provides poor noise
reduction performance above the cutoff frequency and
the filtering close to the data boundaries is prone to
artifacts when computing derivatives (32).
Image pre-processing for the extraction of
Fusarium head blight-infected kernels using
hyperspectral microscopes, white reference and dark-
current reference images were used to adjust the raw
images in order to get the calibrated reflectance images.
Since wheat spikelets have curved surfaces, uneven
light exposure occurred in the hyperspectral images. To
minimise this effect, relative spectral images were
obtained by normalising each pixel with its maximum
reflectance value for further analysis. In order to
distinguish wheat spikelets from the background and
reduce background interference during classification,
threshold segmentation was also used (28). When
choosing a pre-processing method, the image's type
and quality are crucial (1).
Dimensionality reduction
Principal component analysis (PCA) or multivariate
curve resolution (MCR) are two methods used to reduce
the dimensionality of the initial image data. To do this,
the unfolded three-dimensional array is divided into a
set of two sub-matrices (the loading and score
matrices). The primary and relevant information is kept
in the score images (which include the spatial
information and loadings with the spectral information),
while the noisy information is eliminated in the residuals
(33). Other dimensionality reduction methods used to
analyse the whole spectrum include the minimum noise
fraction algorithm, linear discriminant analysis, stepwise
discriminant analysis and partial least square
discriminant analysis (14, 31, 34).
Calibration
The hyperspectral image's spectral and spatial data are
standardised through the calibration procedure. This
procedure will guarantee the precision and coherence of
data gathered under various circumstances. Moreover,
the calibration process eliminates image curvature
effects and instrumentation errors. The most popular
form of calibration is reflection calibration. There are two
forms of reference images used in this type of
calibration: The black reference image was obtained by
completely covering the camera lens with an opaque
cap. A whiteboard, like Teflon, which reflects almost all
light, can be used to create a white reference image.
The following formula is used to adjust the raw
hyperspectral image using these reference images:
where,
R = corrected hyperspectral image
Is = raw hyperspectral image
ID = dark image
IW = white reference image
To remove surface curvature effects, methods such as
spectral image normalisation, adaptive spherical
transform and Lambert transform are used (14).
Image mosaicing
After crop image collection, multiple small-range images
must be spliced into a complete large image by
matching same-name pixels between the overlapping
frames. Geometric reference boards were installed in
R = IS - ID
IW - ID
Plant Science Today | 2026 | Volume 13(sp9) | Page 9
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
various navigation belts as global positioning system
(GPS) reference points to improve the accuracy of
image mosaics. This allowed for the creation of
successive mosaics based on precise geographic
coordinates. The GPS-tagged images were processed
through software like Agisoft PhotoScan, which is
capable of stitching images automatically and
Pix4Dmapper offers automated image stitching
capabilities and enables production of high-precision
orthographic projection images (35, 36).
In addition to classical correction methods,
deep-learning-based approaches have started to
complement existing pre-processing methods. The use
of hyperspectral imaging and deep learning can be used
to detect viral diseases at an early stage along with
performing signal extraction of relevant features without
using any additional pre-processing techniques. This is
indicative of the shift from classical statistical correction
techniques to learned, task-adaptive pre-processing
methods (37).
Feature extraction techniques
The feature extraction process is considered the most
crucial phase in hyperspectral-based classification,
particularly for detecting plant diseases. The goal of
feature extraction is the formation of new feature vectors
using the optimisation of the spectral, spatial and texture
information. Subsequently, the developed feature
vectors are then fed into various classifiers or ML
algorithms to increase the precision and resilience of
disease identification (14).
Feature extraction techniques employ specific
linear transformations to effectively diminish noise
present in feature datasets. These algorithms are
typically organised into two categories: supervised and
unsupervised methods. Unsupervised methods
encompass PCA, maximum noise fraction (MNF) and
independent component analysis (ICA), whereas linear
discriminant analysis (LDA) represents a supervised
approach. Despite their different architectures, both
methodologies serve a common purpose: they convert
hyperspectral imagery from high-dimensional space into
a more compact, lower-dimensional representation
while preserving the essential information content within
a reduced set of principal components (38).
High-dimensional data is reduced to a smaller
collection of uncorrelated variables known as principal
components using a dimensionality reduction method
called PCA. The Hughes phenomenon and the curse of
dimensionality, which together cause classification
accuracy to drastically decline when the number of
spectral features greatly exceeds the available training
samples, are two major obstacles that PCA helps to
overcome in hyperspectral classification. Principal
component analysis is utilised in hyperspectral imaging
to handle the enormous complexity of spectral data in
order to detect powdery mildew in wine grapes (39).
Independent components discriminant analysis
(ICDA) can be described as a non-parametric
discriminant analysis that employs the ICA technique as
the main approach for extracting independent features
out of hyperspectral high-dimensional data. As a result,
it becomes possible to overcome the curse of
dimensionality that negatively affects Bayesian classifier
performance. After applying the transformation that
maps data onto the independent feature space using
ICA, the class-conditional multivariate density estimation
problem that becomes difficult to tackle in high
dimensionality turns into a much simpler product of
univariate estimations for each of the extracted
independent components. The density function of each
extracted independent feature is estimated using a non-
parametric kernel density estimator and the Bayes
formula makes use of those estimations (40).
The MNF is a dimensionality reduction
technique utilised in hyperspectral images that
transforms data from high-dimensional spectra into
lower-dimensional space by reordering principal
components based on signal-to-noise ratio (SNR), while
PCA orders them according to image information. The
MNF algorithm is a very effective method in terms of
dealing with the problem of noise in images and works
by applying a linear transformation on the high-
dimensional data to produce low-dimensional data,
whereby noise covariance matrices are determined
through a multiple linear regression of the residuals
between image blocks. This approach not only lowers
the cost of computation significantly but also brings
down the execution time to roughly one fifth of the
baseline while improving classification accuracy and
robustness against noise (41).
The supervised dimensionality reduction
technique known as linear discriminant analysis (LDA) is
used to extract spectral information for classification by
reducing the number of spectral bands in a
hyperspectral image. The technique transforms the
high-dimensional spectral feature space into a low-
dimensional spectral feature space using the image's
known class labels. As a result, class separation is
improved to address overfitting problems by lowering
parameter estimate error and speeding up classification
computation. Nevertheless, the information needed for
classification is retained in a very small number of linear
discriminant components despite the decreased
spectral bands (42). The LDA was applied to the
spectral reflectance analysis of olive trees (Olea
europaea L.) to extract the most important features that
discriminate between normal trees and trees affected by
Fomitiporia mediterranea. Through LDA, the high-
dimensional spectral information was transformed into
linear combinations having maximum separation
between classes while minimising variance within each
class. The extracted linear discriminants served as the
key features enabling classification and also achieving
strong discrimination accuracy (43).
Partial least squares (PLS) were used as a
dimensionality reduction and feature extraction method
to address the problem of high-dimensional
hyperspectral data obtained from infested potato tubers
Plant Science Today | 2026 | Volume 13(sp9) | Page 10
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
with root-knot nematodes ( Meloidogyne spp.). Partial
least squares reduced the hyperspectral space to a
lower dimension by extracting latent variables having
maximum covariance between the spectral reflectance
data and the target variable, i.e., infestation. Thus,
redundant hyperspectral information is compressed into
useful components through PLS transformation, in
which these extracted PLS components can be used as
input features to the classification models for effective
discrimination of healthy and diseased potatoes (44).
Image segmentation is carried out to separate
the target items from the background prior to formal
spectral analysis or to generate the areas of interest
(ROIs) required for further data extraction and make a
mask (35). The different segmentation techniques
include the watershed algorithm, K-means, threshold-
based and edge detection (14).
Segmentation is the process of splitting an image
into several segments, each representing a homogeneous
area having similar spectral characteristics. It serves as the
initial step in image preprocessing after calibrating the
image in the hyperspectral image analysis process. It aids
in background elimination, makes the data simpler and
improves the efficiency of future classification activities, as
well as helping in reducing the size of data. Many different
methods are used for automatic image segmentation in
hyperspectral imaging, such as clustering, morphology
operations, edge detection, PCA and vegetative index
such as normalised difference vegetation index (NDVI),
along with deep neural network models developed using
RGB images producing binary masks of the hyperspectral
cube. The selection of segmentation depends upon the
purpose and the type of dataset. It is crucial for reflectance
studies in that it ensures purity of the spectral information in
the region of interest by getting rid of the signals outside
that region (12).
Vegetation indices-based segmentation
Vegetation indices are commonly used to extract
spectral features by combining two or more
wavelengths. Analysis of infected leaves indicates a
change in the amount of green colour of the infected
part compared to normal plants and VIs, which are
metrics of greenness in remote sensing image analysis,
are determined using the captured images of the
infected parts and are employed to remove the infected
parts from the images. The computed values of VIs are
used to segment the images and hence extract the
infected parts from the images of the leaves based on
the measurable drop in greenness caused by disease.
This VI-based segmentation approach provides a robust
and automated means of isolating lesion areas on rice
leaves, which then feeds into the classification stage to
distinguish between brown spot and blast diseases (45).
Full spectrum feature extraction uses
reflectance values from all wavelengths. This preserves
complete spectral information as mean reflectance for
each wavelength for all pixels in a plant leaf was used
as features. This method allows ML models to identify
complex spectral patterns and provides better
classification accuracy than using selected indices alone
(29).
Probabilistic topic modelling is a data-driven
feature extraction method. It is an unsupervised ML
method used to reduce hyperspectral data dimensions
and extract useful features for disease classification
(46). The latent dirichlet allocation (LDA) model is
trained to identify hidden spectral patterns (topics) within
large datasets. Each hyperspectral pixel is treated as a
document and each wavelength-reflectance pair is
treated as a word. The algorithm analyses these
spectral “documents” and discovers hidden spectral
patterns called topics. Each topic represents a
characteristic spectral signature and each pixel is
assigned a probability distribution over multiple topics.
The LDA works through a generative process: first, a
topic is selected based on probability and then a
spectral word is chosen from that topic. By repeating
this process, LDA learns the relationship between
spectral features and their topic distributions (29).
In plant disease detection, LDA helps to
minimise the high dimensionality of hyperspectral data
by transforming complex spectral information into a
smaller set of meaningful topic proportions. These topic
proportions capture subtle spectral differences between
healthy and diseased tissues, making them highly useful
for classification and disease monitoring (47).
Enabling deep architectures to automatically
learn hierarchical and discriminative spectral-spatial
patterns rather than depending on manually created
indices, integrating deep learning with conventional ML
enhances feature representation for plant disease
diagnosis. This hybrid approach, which reflects a larger
trend in the field toward automated, learned feature
extraction over manually created spectral indices and
statistical techniques, combines the interpretability and
efficiency of traditional classifiers with the
representational strength of deep learning (48). In deep
learning-based feature extraction for hyperspectral
disease detection, interpretability remains a largely
unresolved challenge. One of the few attempts to
address the "black box" limitation applies explainable 3D
deep learning to hyperspectral images, using saliency-
based visualisation to reveal which spectral-spatial
regions drive the model's classification decisions (49,
50). The lack of comparable work highlights
explainability as an underdeveloped area relative to the
rapid growth of deep learning-based feature extraction
methods in this field.
Vegetation indices for disease detection
Hyperspectral reflectance measurements are usually
transformed into VIs to provide more relevant
information or adjust for atmospheric or soil-related
background interference. Vegetation indices are
arithmetical combinations of reflectance in various
spectral wavebands which correspond to particular plant
biophysical characteristics like the amount of
chlorophyll, biomass, water content, etc. Vegetation
indices were calculated by combining two or more
Plant Science Today | 2026 | Volume 13(sp9) | Page 11
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
spectral wavelengths to extract features from spectral
reflectance that highlight the biophysical and spectral
characteristics of green vegetation. It aids in
distinguishing them from other aspects of the image. A
few VIs employ a variety of wavelength measurements
to explain physiological characteristics of plants and
other vegetation generic characteristics or the growth
parameters of a particular plant (51).
The mechanism through which vegetative
indices operate is by assessing the contrast between
the red and near-infrared reflectance of the vegetation.
This works well since the red spectrum is absorbed by
the chlorophyll while the leaf mesophyll structures cause
the NIR reflectance. In the event where there is a huge
difference between the red and NIR reflectance, it
means that there are plenty of healthy plants. On the
other hand, when there is very little contrast between
the red and NIR reflectance, it can be concluded that
there is less vegetation within the area (51).
The use of a single VI for plant disease
detection often lacks sufficient discriminatory power to
distinguish diseased from healthy plants and also
different vegetation indices correspond to different plant
properties (29). A more effective approach involves
combining multiple vegetation indices to capture diverse
spectral characteristics that respond differently to
disease-induced changes in plant physiology. By
integrating several indices as classification features and
applying them in ML algorithms, higher accuracy and
early disease detection are achieved. This multi-index
approach takes into account that different diseases
change plant spectral signatures differently, so disease-
specific index combinations are more effective than any
single index alone. The usage of multiple VIs with
support vector machines to classify spectral data of
sugar beetroot plants got classification accuracies of
around 95 % for Cercospora beticola and Uromyces
betae and more than 92 % for Erysiphe betae while
comparing with single VI accuracy (52). Different VIs
were important for different diseases, highlighting the
importance of selection of disease-specific indices in
developing detection systems.
Standard vegetation indices, such as NDVI,
spot general plant stress but cannot tell specific
diseases apart. This leaves farmers clueless about
which pathogen is hitting their crops. The spectral
disease indices come in to help here by using custom
spectral formulas. These formulas target the unique light
reflectance from each individual disease as different
pathogens affect plants differently. The disease specific
spectral indices are customised mathematical formulas
tailored to detect particular plant pathogens (e.g.,
powdery mildew, rust, blight). Instead of using general
VIs, developing unique spectral signatures caused by
each disease’s physiological effects on plants and
create targeted indices that maximise sensitivity to those
specific changes. This approach enables differentiation
between different diseases and healthy plants using
remote sensing data, allowing precise, pathogen-level
diagnosis at field scale- critical for targeted disease
management without unnecessary fungicide application
(53). Table 1. summarises spectral vegetation indices
commonly applied in disease detection.
Machine learning approaches in data analysis
After the required preprocessing is completed and the
data has been separated into training and test sets, the
final stage in image analysis is to choose a model and
apply it to the data. The model is trained using the
training dataset. Training is the most important stage in
machine learning. The prepared data is fed into the ML
model during training, allowing it to recognise patterns
and generate predictions. The model completes the task
and learns from the data (49).
Table 1. Common spectral vegetation indices used in hyperspectral plant disease detection
Sl. No Vegetation indices Formula Significance Reference
1.
Normalised
difference vegetation
index (NDVI)
Changes in chlorophyll content,
examine green and healthy vegetation (54)
2.
Photochemical
reflectance index
(PRI)
Sensitive to carotenoid pigment
changes, changes in photosynthetic
effiency
(55, 56)
3.
Anthocynanin
reflectance index
(ARI)
Elevated anthocyanin concentration
indicates deteriorating vegetation,
allowing anthocyanin- based spectral
indices to serve as diagnostic markers
of vegetation stress
(57)
4.
Modified chlorophyll
absorption ratio
index (MCARI)
Leaf area Index and chlorophyll
estimation (57, 58)
5.
Carotenoid
reflectance index
(CRI)
Vegetation under stress accumulates
higher carotenoid levels, making
carotenoid indices useful indicators of
plant stress
(59)
NDVI = (NIR - RED)(NIR + RED)
PRI = (ρ531 - ρ570)(ρ531 + ρ570)
-ARI = 1ρ550 1ρ700
MCARI =
(ρ701 - ρ671) - 0.2(ρ701 - ρ549)
(ρ701 - ρ761)
-CRI = 1ρ510 1ρ550
Plant Science Today | 2026 | Volume 13(sp9) | Page 12
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
6. Structure insensitive pigment index (SIPI) Increases the index’s sensitivity to the bulk carotenoids to chlorophyll ratio (56)
7.
Normalized
difference nitrogen
index (NDNI)
Measures relative nitrogen content in
plant canopies (48)
8. Plant pigment ratio (PPR)
Detects foliar pigments by the ratio of
green to blue reflectance. A high value
is obtained for strongly pigmented
vegetation and a low value for weakly
pigmented foliage.
(60, 61)
9.
Plant senescence
reflectance index
(PSRI)
Detect the carotenoid to chlorophyll
ratio to quantitatively measure leaf
senescence and fruit ripening
(61)
10.
Green normalised
difference vegetation
index (GNDVI)
Similar to NDVI but measures the
green spectrum, Sensitive to
chlorophyll concentration
(51, 61)
11. Red green index Indicator of plant chlorophyll content and physiological condition (61)
12.
Triangular
vegetation index
(TVI)
TVI =
0.5 * (120(ρ750 - ρ550) - 200(ρ670 - ρ550))
The area of a hypothetical triangle
with vertices at red, green and NIR
wavelengths is the basis for TVI. It is
affected by both the leaf area index
and the amount of chlorophyll. When
the absorption of chlorophyll is high,
the triangle's total area increases.
(51)
13. Modified simple ratio(MSR) Enhances the sensitivity of vegetation biophysical measurements (57, 62)
14.
Transformed
chlorophyll
absorption index
(TCARI)
Measures chlorophyll content and leaf
area index with minimal soil effects
(57, 58,
61)
15. Nitrogen reflectance index (NRI) Assess plant nitrogen status (61, 63)
16.
Red edge
normalised
difference vegetation
index (RENDVI)
Modifying the NDVI by using red edge
instead of absorption and reflectance
peaks to improve sensitivity of small
changes in canopy foliage content, gap
fraction and senescence
(61)
17. Simple ratio (SR)
Estimate vegetation biophysical
parameters like leaf area index and
photosynthetically active radiation
absorption
(29, 51,
62)
18.
Normalised
difference lignin
index (NDLI)
Estimate lignin content in vegetation
based on reflectance at specific
infrared wavelengths
(29, 48,
51)
GNDVI = (ρ750 - ρ540 + ρ570)(ρ750 + ρ540 - ρ570)
RGI =
ρ690
ρ550
TCARI =
3 (ρ700 - 670 0.2 * (ρ700 - 550) *
ρ700
ρ670
SIPI =(ρ800 - ρ445)(ρ800 + ρ680)
NDNI =
1
ρ1510log - log 1ρ1680
1
ρ1510log + log 1ρ1680
PPR = (ρ550 - ρ450)(ρ550 + ρ450)
PSRI = (ρ660 - ρ510)(ρ760
MSR =
- 1NIRRed
- 1NIRRed
NRI = (ρ570 - ρ670)(ρ570 + ρ670)
RENDVI = (ρ750 - ρ705)(ρ750 + ρ705)
SR =
ρ900
ρ680
NDLI =
1log ρ1754 - log 1ρ1680
1log ρ1754 + log 1ρ1680
Plant Science Today | 2026 | Volume 13(sp9) | Page 13
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
19.
Red- edge stress
vegetation index
(RVSI)
Detect plant stress by analysing
reflectance in the red edge wavelength
region
(57, 64)
20. Normalised pigment
chlorophyll ratio
index (NPCI)
Measure carotenoid-to-chlorophyll ratio
for assess nitrogen status and
senescence
(56, 57,
63)
21. Greenness index (GI) Assess vegetation greenness and detecting changes in leaf pigmentation (51, 65)
22. Water band index (WBI) Evaluate water content and stress in vegetation (56, 65)
23. Chlorophyll index (CI) Estimate chlorophyll concentration (66)
24.
Ratio analysis of
reflectance of
spectral chlorophyll
a (RARSa)
Estimate chlorophyll a concentration
(61, 66,
67)
25.
Ratio analysis of
reflectance of
spectral chlorophyll
b (RARSb)
Estimate chlorophyll b concentration
(61, 66,
67)
26.
Ratio analysis of
reflectance of
spectral chlorophyll
c (RARSc)
Estimate carotenoids concentration
(61, 66,
67)
27. Pigment specific
simple ratio
(PSSRa)
Estimate chlorophyll a concentration,
strong linear relationships with pigment
concentration
(29, 68)
28. Pigment specific
simple ratio
(PSSRb)
Estimate chlorophyll b concentration,
strong linear relationships with pigment
concentration
(29, 68)
29. Canopy chlorophyll index (CCI)
2D remote sensing index calculated
from NDVI and NDRE to analyse the
chlorophyll content in vegetation
isolating leaf signal from soil
background
(51)
30.
Modified triangular
vegetation index
(MTVI)
MTVI =
1.2 * (1.2(ρ800 - ρ550) - 2.5 * (ρ670 - ρ550))
Measure leaf area index, less sensitive
to chlorophyll content variations
(58, 61,
66)
31. Health index (HI) Detect and classify whether plant
leaves are healthy or diseased based
on their reflectance charateristics
(56, 64)
Disease specific indices
32. Yellow rust index (YRI) Detect yellow rust (56, 64)
33. Powdery mildew- index (PMI) Identify powdery mildew infected leaves (56, 64)
ρ – reflectance in nm; NIR: near infra-red band; red: red band; red edge: red edge band.
Machine learning methods are able to handle
hundreds of spectral bands and have become well
established tools for detecting, quantifying and
characterising plant diseases in field, greenhouse and
laboratory conditions. These methods infer or learn
model properties from the data itself rather than
employing a predetermined model. Supervised methods
require labelled training datasets. While unsupervised
methods operate without the training datasets (16).
Supervised machine learning
Supervised ML models employ labelled datasets with
sensor data, images, or additional pertinent information
to identify patterns associated with different crops or
RSVI = (ρ712 + ρ752 )2 - ρ732
NCPI = (ρ680 - ρ430)(ρ680 + ρ430)
GI =
ρ554
ρ677
WBI =
ρ900
ρ970
CI = - 1ρ840
ρ570
RARSa =
ρ675
ρ700
RARSb = ρ675(ρ700 * ρ650)
RARSc =
ρ760
ρ500
PSSRa =
ρ800
ρ680
PSSRb =
ρ800
ρ635
CCI =
(NIR - RED EDGE)
(NIR + RED EDGE)
(NIR - RED)
(NIR + RED)
HI = + 0.5 * ρ403(ρ739 - ρ402)(ρ739 + ρ402)
YRI = + 0.5 * ρ403(ρ730 - ρ419)(ρ730 + ρ419)
PMI = + 0.5 * ρ738(ρ515 - ρ698)(ρ515 + ρ698)
Plant Science Today | 2026 | Volume 13(sp9) | Page 14
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
diseases (6). The models can be either regression
models (for predicting and analysing the association
between the target variables and the spectral response)
or classification models (for detecting and classifying
diseases), depending on the study's goals (14). Table 2.
summarises classification and detection accuracies
reported across the reviewed studies, along with the
scale, sensor platform, feature set and model where
available, to enable direct comparison across crop–
pathogen systems.
Classification models
Classification models classify each data sample based
on their classes by training on labelled data. These
models can operate at two levels: 1) the image level,
where the overall content of an image (or an entire plant
when only one is present) is interpreted and assigned a
category; 2) pixel level, in which each pixel in the image
is given a meaningful description, such as "healthy
tissue" or "disease symptom," resulting in a fully
classified image (15).
The commonly used method for classifying
pixels is the spectral angle mapper (SAM) that uses a
cosine distance measure in assessing similarity among
spectra. It involves assigning the class for all pixels that
lie within a certain angle range in the feature region
compared to the reference data point. The major
advantage of using cosine distance is that it takes care
of the scaling effects arising from differences in lighting
conditions among other issues, hence making
comparisons among the spectra easy. The SAM was
applied in detecting and locating different disease
symptoms on sugar beets caused by three diseases
(21), detecting citrus canker in citrus plants (50) and
fusarium head blight in wheat ears (69).
Table 2. Cross-study comparison of hyperspectral disease detection methods and their reported performance metrics
Crop Pathogen Country
Leaf/
canopy/
field
Sensor and
platform
Spectral
range Feature set Model N (train/test) Reported metric value
Sugar
beet
Cercospora
leaf spot
(Cercospora
beticola)
Germany Leaf
ImSpector
V10E line-
scan
hyperspectra
l imaging
spectrograph
- mirror
scanner,
controlled
chamber,
ASD-Pro-
Lamps
illumination
400–1000 nm
(2.8 nm
spectral
resolution)
. Pixel-wise
reflectance
spectra;
. 3 endmember
classes:
healthy /
margin /
necrotic centre
Spectral
angle
mapper
(SAM)
. 20 fully
developed
symptoms/
disease for
spectral
signature
extraction;
. 10
symptomat
ic areas
analyzed
daily; pixel-
level
classificati
on
8 dai: 98.90 %
accuracy, κ =
0.99, 0 %
diseased area
classified
11 dai: 89.01 %
accuracy, κ =
0.53, 6.56 %
diseased
14 dai: 96.58 %
accuracy, κ =
0.92, 7.61 %
diseased
17 dai: 98.73 %
accuracy, κ =
0.98, 20.92 %
diseased
Citrus
Huanglongbing
(HLB) / citrus
greening
USA
Canopy
(pixel/tree-
canopy
level)
. AISA
EAGLE
VNIR
hyperspectra
l imager,
airborne (640
m altitude);
.
WorldView-2
satellite
multispectral
(for
comparison)
. 400–1000
nm, 128
bands, 5 nm
resolution
(HS);
. WV-2: 8
bands, 400–
1040 nm (MS)
. Savitzky–
Golay-
smoothed
reflectance
spectra;
. SVM-derived
canopy mask;
. VCA pure
endmembers;
spectral angle
(SAM);
. red-edge
position (REP,
linear
extrapolation)
. ESAM
(SVM mask
→ VCA
endmembe
r extraction
→ SAM
classificatio
n → REP
false-
positive
filtering);
. compared
against
Mahalanobi
s distance
and K-
means
. Block 8a
(mature):
26 healthy - 23 HLB /
25 healthy - 22 HLB;
. Block 2b
(intermedia
te): 14/14
HLB; Block
5c (young):
33/33 HLB
Detection
accuracy -
. Block 8a: 82.6
% (calibration),
86.3 %
(validation);
. Block 2b: 64.2
%/71.4 %;
. Block 5c: 72.7
%/63.6 %. MS
image (SAM):
78.2 %/63.6 %
Wheat Fusarium head blight Germany
Ear
(organ-
level, both
lab and
field)
. Lab
hyperspectra
l pushbroom
imager
(ImSpector
V10E +
Pixelfly qe
CCD),
· 400–1000
nm, 2.5 nm
resolution
(HS);
· fluorescence
at 685–700
nm emission
band (CFA)
. Raw
reflectance
spectra (512
bands) for SAM;
pixelwise Fv/Fm
distributions and
cumulative
Fv/Fm curve for
. Spectral
Angle
Mapper
(SAM, ROI-
based
endmembe
rs) for HS;
cumulative
. HS: 12
infected +
12 control
ears (time
series);
CFA lab: 6
infected +
6 control
. HS-SAM vs.
visual rating:
R²=0.964;
best detection
window 14 ± 2
dai, up to 50 %
severity;
CFA lab:
Plant Science Today | 2026 | Volume 13(sp9) | Page 15
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
bench-
mounted;
FluorCAM
700MF
chlorophyll
fluorescen
ce imager,
bench and
field
(shielded
box)
CFA
Fv/Fm
threshold
(0.3)
classificatio
n for CFA
(time
series), 15
ears at
BBCH75;
· CFA field:
30 ears (of
50 imaged)
differentiates in
10 % severity
steps from 6 dai;
· CFA field: R² =
0.658 (RMSE=17
%), rising to R² =
0.80 after outlier
removal
Tomato
Bacterial leaf
spot
(Xanthomonas
spp.)
USA Leaf
Resonon
benchtop
HSI (Pika L +
Pika
NIR320), lab
imaging
400–1700 nm
(VISNIR 388–
1025 nm +
SWIR 896–
1715 nm)
Full spectra
(leaf- and
pixel-level)
and 14
vegetation
indices (VIs)
LDA,
SVM,
KNN, RF,
GBM,
XGBoost,
MLP
(best:
LDA for
full
spectra;
RF for
pixel/VI
data)
Leaf-level:
36–92
leaves/sta
ge across
4 stages;
pixel-level:
648
pixels/rang
e across 9
classes
Leaf-level LDA
accuracy 0.55–
0.74
(VISNIR/SWIR,
pre- to post-
symptomatic);
pixel-level RF
accuracy 0.63 ±
0.13 (VISNIR),
0.57 ± 0.16
(SWIR);
VI-based models
outperformed raw
data by 26–37 %
Sugar
beet
Cercospora
beticola,
Uromyces
betae, Erysiphe
betae
Germany Leaf
ASD
FieldSpec
Pro FR,
handheld
leaf-clip
400–1050 nm 8 VIs + SPAD
SVM
(RBF);
compared
vs. DT,
ANN
3,570 leaf
spectra
(pooled,
10-fold)
· Acc. 97 %
(healthy vs.
diseased); 86 %
(4-class);
· 65–90 %
presymptomatic
Brinjal,
Cluster
beans,
Long
Beans
India Leaf
ASD
FieldSpecPr
o
spectroradio
meter,
ground-
based/in-lab
(fibre optic,
8° FOV,
halogen
lamp
illumination)
350–2500 nm
(2151 bands),
analyzed in 5
sub-regions:
Visible (350–
670 nm),
Red Edge
(671–780 nm),
NIR (781–
1000 nm),
SWI (1301–
1500 nm),
SWII (1701–
1900 nm)
350–2500 nm
(2151 bands),
analyzed in 5
sub-regions:
Visible (350–
670 nm),
Red Edge (671–
780 nm),
NIR (781–1000
nm),
SWI (1301–
1500 nm),
SWII (1701–
1900 nm)
Random
Forest
classifier
(with PCA
or LDA as
pre-filter)
25 leaf
samples
each of
Brinjal,
Cluster
Beans,
Long
Beans
(healthy +
infected),
10 spectral
iterations/s
ample
Best region =
Visible/Red Edge;
mean accuracy
(PCA): Brinjal
0.864, Cluster
Beans 0.92, Long
Beans 0.90
(Kappa
0.66/0.83/0.78);
Long beans red
edge PCA hit 1.0
accuracy/Kappa
Chardon
nay wine
grape
(Vitis
vinifera)
Powdery
mildew
(Erysiphe
necator)
Australia
Berry
(detached)
and bunch
(intact),
lab-
controlled
HySpex
VNIR 1600 +
HySpex
SWIR-320m-
e line-scan
cameras,
benchtop/tra
nslation
stage
VNIR 400–
1000 nm (3.7
nm resolution,
160 bands);
SWIR 970–
2500 nm (6
nm resolution,
256 bands)
LDA/PCA-
reduced
spectral bands
- spatial-
spectral texture
features (mean,
SD,
homogeneity via
integral images
at multiple block
sizes)
Modified
Random
Forest
classifier
(pixel-wise
spectral vs.
spatial-
spectral
texture-
based)
Detached
berries:
128
healthy +
136
diseased
(training/C
V);
intact
bunches:
30 (10
healthy, 10
infected,
10
severely
diseased)
× 2 scans
= 60
images
(validation)
Detached berries:
accuracy up to
0.998 ± 0.003
(LDA+spatial-
spectral, VNIR);
intact bunches:
overall accuracy
improved from
0.76 (spectral-
only) to 0.86
(spatial-spectral);
severely diseased
bunches detected
with high
accuracy,
healthy/infected
discrimination
harder
Plant Science Today | 2026 | Volume 13(sp9) | Page 16
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
Avacado
Laurel wilt
(Raffaelea
lauricola)
USA Leaf
SVC
HR-1024,
handheld;
greenhouse
- field
350–1000 nm
(66 bands, 10
nm)
PCA scores
(99.9 %
variance)
LDA, QDA,
NB, BDT
Random
75:25 split
· 249/83 (healthy
vs
asymptomatic);
209/69 (healthy
vs symptomatic)
Cucumb
er
Angular leaf
spot
Pseudomonas
syringae pv.
lachrymans
China Leaf
Line-scan
HSI camera,
benchtop
380–1030 nm
(476 used for
modeling)
Regressionco
efficient-
selected
important
wavelengths
(IWs):
9 bands for
Chl (408, 484,
533, 582,
686, 708,
750, 911, 976
nm);
5 bands for
Car (484,
558, 675,
705, 765 nm)
PLSR (RC-
PLSR)
130
calibration
/ 64
prediction
(194 total
leaves)
· Chl-HSI:
Rc=0.901,
Rcv=0.889,
Rp=0.871,
RMSEP=0.25
mg/g;
· Car-HSI:
Rc=0.803,
Rcv=0.793,
Rp=0.876,
RMSEP=0.05
mg/g
Wheat
Septoria tritici
blotch
(Zymoseptoria
tritici) +
Stagonospora
nodorum blotch
(Parastagonos
pora nodorum)
Russia
Leaf and
ear (in situ,
field, non-
imaging
point
measurem
ents)
Ocean
Insight
"Flame" VIS-
NIR non-
imaging
hyperspectro
meter,
handheld/fiel
d (0.3 nm
resolution)
338–1018 nm
(downsampled
to 350–950
nm, smoothed
to 6 nm step
for NN)
Full
hyperspectral
reflectance
signatures (100
values/curve
after
smoothing);
red-edge slope
"angle" metric
derived from
curve shape
Perceptron
neural
network
(PNN), 2
layers (5+2
neurons)
2000
spectral
curves/leaf
dataset,
2000–
4000/ear
dataset;
8 leaf + 8
ear
datasets
total
(healthy,
light,
medium,
severe,
dead)
· PNN (10,000
epochs): 0.99
accuracy, 0.94
precision, 0.89
recall, 0.91 F-
score;
· healthy/severe
leaf classification
88.1–97.7 %;
light/medium
severity harder
(38–66 %);
· Cohen's kappa
0.90–0.98
(healthy/severe)
vs. 0.13–0.64
(light/medium),
overall κ=0.69
Apple Marssonina blotch USA
Leaf-level
(indoor
detached-
leaf
imaging),
with a
supplemen
tary test on
an outdoor
canopy/tre
e image
SPECIM PS-
V10E
pushbroom
camera, lab
setup
356–1000 nm
(437 bands
used; 10
selected)
OSP-selected
bands (mostly
NIR 680–957
nm
Decision
tree,
weighted
KNN,
ensemble
bagged
10,000 px
(2,000 × 5
classes);
5-fold CV,
no explicit
train/test
split;
outdoor
image as
external
test
Accuracy:
84.3 %
(ensemble),
79.8 % (DT), 71.3
% (KNN)
Barley
Powdery
mildew
(Blumeria
graminis)
Germany
Leaf/
cellular
(microscop
ic)
Hyperspectra
l line-scan
camera
(Specim PFD
V10E) on
stereo
microscope
foreoptic,
7.3x
magnification
400–1000 nm
(analyzed 420
–830 nm),
2.73 nm
resolution
Full-spectrum
reflectance
signatures;
SiVM-derived
extreme
components/sim
plex
representation
Simplex
Volume
Maximizatio
n (data-
driven
matrix
factorizatio
n) +
Dirichlet-
aggregation
regression
(leaf
traces);
LLR-based
binary
disease
maps
3
inoculated - 4 healthy
leaves per
genotype
(n=3
genotypes:
WT,
Mla12,
mlo3)
Automated vs.
manual spectral
difference: WT
−0.04 to 0.04;
Mla12 −0.02 to
0.03;
RSD 0.6–14.7 %
across
genotypes/time
Plant Science Today | 2026 | Volume 13(sp9) | Page 17
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
Ash
Ash dieback
(Hymenoscyph
us fraxineus)
UK
Canopy/
landscape
(forest site,
airborne)
Hyspex
VNIR 1800
hyperspectra
l camera,
human-
crewed
aircraf
· 410–1001
nm, 186
bands, 3.26
nm resolution;
· 0.32 m native
/ 1.28 m
downsampled
spatial
resolution
Raw pixel
spectra; pixel
purity η(x) via
HySime + VCA
unmixing;
density-purity
harmonic mean
ζ(x); diffusion
distances
PLSDA
(species
mapping)
→ Random
Forest
(supervised
disease
classes:
healthy/
infected/
severely
infected)
vs. D-VIS
(unsupervis
ed diffusion
- VCA
clustering,
K=2:
healthy/die
back)
PLSDA:
166
Madingley
crowns +
256
crowns
from 3
other sites;
RF test
set: 16
trees/class
;
D-VIS:
n=72,775
ash pixels
(401×279
scene
PLSDA accuracy
85.3 %;
RF accuracy 77.1
%;
D-VIS vs. RF
overall accuracy
71.0 %, average
accuracy 71.3 %
Tobacco
Tobacco
mosaic virus
(TMV)
China Leaf
(ImSpector
V10E +
Hamamatsu
C8484-05
CCD),
benchtop
380–1023 nm
(450–1000
nm/434 bands
used)
SPA-selected 8
EWs;
GLCM texture
(contrast,
correlation,
entropy,
homogeneity);
EW+texture
fusion
· PLS-DA,
RF, SVM,
LS-SVM,
ELM,
BPNN
120
samples
for
calibration
(training)
and 60 for
prediction
(testing)
Best: SPA-ELM
overall accuracy
98.33 % (EWs);
BPNN 93.33 %
(texture);
BPNN 95.00 %
(fusion
Wheat
Puccinia
triticina (leaf
rust)
Iran Leaf
ASD
spectroradio
meter
(contact
probe) +
RGB camera
(Canon IXUS
85 IS)
450–1000 nm
22 SVIs (NDVI,
NBNDVI, GI,
PRI, ARI,
TCARI, MSR,
PSRI, PhRI,
TVI, RVSI, NRI,
SIPI, NPCI,
CARI, MCARI +
derivative-
based
variables)
Threshold/
statistical
classificatio
n (healthy
vs. rust)
across 10
disease-
severity
classes
300
spectral
samples
Best:
NBNDVI 83 %,
NDVI 81 %,
GI 77 %,
ARI 75 %
classification
accuracy
Oilseed
rape Alternaria spp Poland Leaf
VNIR
camera
(ImSpector
V10E) +
SWIR
camera
(N25E 2/3",
SPECIM) on
conveyor
belt;
FLIR SC620
thermal
camera (8–
13 µm)
· 400–1000
nm (VNIR) +
1000–2500
nm (SWIR);
· classification
used 430–
2376 nm at
~32 nm
increments (61
channels)
Second-
derivative
transformed
reflectance
spectra
(Savitzky-Golay,
4th-order,
11 points);
raw/averaged
reflectance also
examined
· 8
classifiers
compared
(NB, SL,
SVM,
LibLINEAR,
BNN, FT,
RF, J48);
· best =
Back-
Propagatio
n Neural
Network
(BNN)
251
instances
(training/te
sting) + 28
instances
(validation)
Days-after-
inoculation
classification:
BNN 90.5 %
(train/test), 93.3
% (validation);
Alternaria species
classification:
BNN 80.5 %
(train/test), 82.3
% (validation)
Onion
Burkholderia
cepacia (sour
skin)
USA
Bulb/organ
level (neck
& shoulder
region)
SWIR
pushbroom
hyperspectra
l imager:
LCTF +
InGaAs
camera, lab
chamber with
halogen
lamps
950–1650 nm
(2 optimal
bands
selected: 1070
& 1400 nm)
Log-ratio image
(1070/1400 nm)
→ 12
statistical/textur
al features
(GLCM-based),
reduced via
stepwise
discriminant
analysis to 3
LDA on
segmented
pixel count
(threshold=
0.45);
SVM (RBF
kernel,
γ=1.5) on 3
image
features
140
images
total (75
onions × 2
timepoints,
70
inoculated - 5
controls)
Accuracy:
LDA 80 %
(healthy 97.14 %,
diseased 62.86
%);
SVM 87.14 %
(healthy 81.43 %,
diseased 92.86
%)
Plant Science Today | 2026 | Volume 13(sp9) | Page 18
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
(max, contrast,
homogeneity);
also pixel-count
from
thresholded log-
ratio images
Soybean
Charcoal rot
(Macrophomina
phaseolina)
USA
Stem
(excised,
benchtop
imaging)
Pika XC
hyperspectra
l line imaging
scanner
(Resonon),
240
wavebands,
2.5 nm
resolution
400–1000 nm
Raw
hyperspectral
data cubes (3D
spatial-spectral
patches,
64×64×240
pixels);
no hand-crafted
indices
3D deep
convolution
al neural
network
(3D-CNN, 2
conv+pooli
ng layers, 2
fully-
connected
layers,
softmax
output) with
class-
balanced
weighted
binary
cross-
entropy
loss
Training:
1090
patches
(940
healthy/15
0 infected);
Validation:
194;
Test: 539
(from 111
original
hyperspect
ral stem
images: 64
healthy/47
infected)
Test accuracy
95.73 %; recall
0.92; precision
0.82; F1 (infected
class) 0.87;
5-fold CV mean
accuracy 95.49 %
(SD 2.01 %)
Sweet/
bell
pepper
Tomato spotted
wilt virus
(TSWV)
USA
Pixel-level
within
whole-
plant (top-
canopy)
images
Hyperspectra
l reflectance
imaging
system,
Specim
ImSpector
V10E
camera
mounted on
a Motorman
SL robotic
manipulator;
· lab setup
with two
halogen
lamps at 45°
395–1005 nm
(83 bands
used out of full
range;
8 bands
sufficient after
MVPCA band
selection)
Pixel-wise raw
reflectance
spectra (83
bands), no
dimensionality
reduction
beyond band
selection;
also evaluated
with
MVPCA/FDPC/
SUBS-selected
bands + OR-
AC-GAN-
augmented
synthetic
spectra
OR-AC-
GAN
(outlier-
removal
auxiliary
classifier
GAN);
benchmark
ed against
1D-CNN
and
standard
AC-GAN;
final
classificatio
n via built-
in
discriminat
or (or KNN
when band-
selected)
Training
pixels:
103,769
backgroun
d, 105,561
healthy,
2,071
TSWV
(from 6
images);
Test: 54
plants/ima
ges
Pixel accuracy
98.03 %;
TSWV false-
positive rate in
healthy plants
1.47 % (avg,
worst 8.21 %);
plant-level
sensitivity 92.59
%;
overall plant-level
classification
accuracy 96.25 %
ASD: analytical Spectral Devices; SVM: support vector machine; VNIR: visible and near-infrared; VCA: vertex component analysis;
REP: red edge position; ESAM: extended spectral angle mapping; HLB: huanglongbing; MS image: Multispectral image; ROI: region
of interest; CFA Lab: colour feature analysis laboratory; KNN: K-nearest neighbours; LDA: linear discriminant analysis; SWIR: short-
wave infrared; PCA: principal component analysis; QDA: quadratic discriminant analysis; NB: naïve bayes; BDT: binary decision
tree; PLSR: partial least squares regression; ELM: extreme learning machine; BPNN: backpropagation neural network; PLS-DA:
partial least squares discriminant analysis; EWS: early warning system; LS-SVM: least-squares support vector machine; NDVI:
normalised difference vegetation index; GI: greenness index; ARI: anthocyanin reflectance index; RBF: radial basis function; CNN:
convolutional neural network.
Support vector machine (SVM) is a usual
technique for hyperspectral data analysis. It operates on
the empirical risk minimisation principle which is
effective at establishing class boundaries in high-
dimensional feature spaces even when a limited amount
of training data is available. Support vector machine is
more effective regarding classification precision,
processing speed and exhibit low sensitivity to the
Huges phenomenon resulting in reducing the phase that
takes a long time and is necessary for any other feature
reduction technique (70). Application of SVM in foliar
disease identification from hyperspectral data, identifies
the initial symptoms even before they become apparent
to the human eye (71).
Using both visible–near-infrared (VISNIR) and
SWIR spectral ranges, linear discriminant analysis
(LDA) is used to separate inoculated from uninoculated
tomato leaves in hyperspectral imaging data as early as
two hours after inoculation. Linear discriminant analysis
consistently exhibited the best classification
performance across all four bacterial leaf spot disease
progression stages at the leaf level, while also providing
interpretable visualisations through LDA plots that
revealed clear spectral separation between healthy,
Plant Science Today | 2026 | Volume 13(sp9) | Page 19
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
bacterially infected and abiotically stressed leaf tissues.
The LDA proves its effectiveness as both a classification
and visualisation tool in pre-symptomatic plant disease
detection (2).
A versatile ML technique called K-nearest
neighbours (K-NN) is applied to both classification and
regression issues. This algorithm is referred to as an
instance-based algorithm, because instead of learning a
model from training samples, all the training samples
are stored in memory and predictions are made based
on similarity assessment between a test sample and
training samples (72).
extreme gradient boosting (XG Boost) is
employed specifically for precise categorisation of plant
diseases in pepper, tomato and potato crops along with
convolutional neural networks (CNN) which extracts
features from images (73).
Random forest (RF) is used in classifying
healthy and infected vegetable leaves (brinjal, cluster
beans, long beans) from hyperspectral reflectance data,
where it creates multiple decision trees by randomly
selecting samples and generates a separate tree for
each sample. Each tree gives its own prediction (vote)
and the final product is chosen from the class that gets
the most votes, making it resistant to overfitting since no
single tree dominates the final result (30). Modified
random forest classifiers combined with LDA-reduced
hyperspectral features were developed to detect and
classify powdery mildew infection severity levels in
Chardonnay grape bunches (39).
Decision tree model- bagged decision tree
(BDT) is a non-linear classifier used for classifying into
four categories: healthy, asymptomatic-infected,
symptomatic-infected and freeze-damaged avocado
leaves from VISNIR spectral data, where multiple
classification trees are developed using bootstrap
replicates (randomly resampled subsets of training data)
to strengthen and stabilise the final classification. Its
main advantage over a single decision tree is that the
bagging (bootstrap aggregating) process lowers
variance and overfitting, making it more robust and
efficient for detecting laurel wilt even in asymptomatic
leaves where spectral differences are minimal (74).
Regression models
Regression models generate continuous numerical
outputs rather than fixed categories, making them well
suited for tasks like measuring pigment concentrations
or quantifying disease severity (15).
The frequent method employed to interpret
hyperspectral images is partial- least square regression
(PLSR) because it can handle a large number of highly
correlated bands. The PLSR is used to manage a lot of
bands that are highly connected. The degradation of
carotenoids and chlorophyll caused by an angular leaf spot
infection is measured using PLSR (75). Using
hyperspectral reflectance data, the extent of damage
caused by sclerotinia rot in celery was modelled using
PLSR, which effectively managed the high-dimensional
space of spectral data and multicollinearity of spectral
bands. This model helped find significant wavelengths that
were highly responsive to disease-induced physiological
alterations in plant tissues for the prediction of disease
extent based on hyperspectral imaging analysis (76).
Gaussian process regression (GPR), PLSR
and v support vector regression (v-SVR) methods were
employed in wheat leaf rust disease detection. The
GPR outperformed both PLSR and SVR in detecting
wheat leaf rust disease severity from hyperspectral data.
All three ML regression techniques surpassed traditional
spectral vegetation indices, indicating that ML methods
are more effective for disease severity detection,
especially when disease symptoms and varying training
sample sizes are considered (76).
Unsupervised learning
Unsupervised methods find hidden patterns and
connections within data and therefore fall under the
category of pattern recognition (16).
Unsupervised ML (spectral unmixing, clustering
and structure detection) was applied to classify
hyperspectral images with low spatial resolution, proving
that unsupervised approaches can effectively detect and
classify surfaces without requiring labelled training data,
offering a practical solution when ground truth information
is unavailable in remote sensing applications (77).
K-means clustering was used as an unsupervised
method to segment the leaf from the background in the
hyperspectral images, by exploiting the unique spectral
signature difference between vegetation and surrounding
materials (soil, plastics, metals etc.). It classifies image
spectra into two clusters (leaf and background), generating
a 2D leaf mask to ensure only leaf spectral measurements
were used for further disease classification (29).
Simplex volume maximisation (SiVM) was used
as a data-driven, unsupervised matrix factorisation
technique to automatically extract characteristic spectral
signatures from hyperspectral microscopy data of barley
leaves affected with powdery mildew (Blumeria graminis).
The automated SiVM methodology produced results
comparable to the labour-intensive and time-consuming
manually assessed hyperspectral signatures, but it was
assessed in a shorter period of time than traditional
approaches. Additionally, it was employed to produce
intuitive summaries of several barley genotypes, including
fingerprints and traces of hyperspectral dynamics (78).
The unsupervised diffusion and volume-
maximising interface segmentation (D-VIS ) clustering
algorithm efficiently divides hyperspectral image data into
significant spectral classes by combining a volume-
maximising geometric clustering technique with diffusion
maps or nonlinear dimensionality reduction. This technique
works completely without labelled training data, which
makes it suitable for identifying minute spectral changes in
complicated hyperspectral imaging that are linked to plant
diseases like ash dieback (Hymenoscyphus fraxineus)
(79).
Plant Science Today | 2026 | Volume 13(sp9) | Page 20
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
Application of six ML classifiers back propagation
neural network (BPNN), extreme learning machine (ELM),
least squares support vector machine (LS-SVM), SVM,
random forest and partial least squares discriminant
analysis (PLS-DA) on hyperspectral imaging data to pre-
symptomatically detect tobacco mosaic virus (TMV)
infection in tobacco leaves, where BPNN, ELM and LS-
SVM outperformed others achieving up to 95 % accuracy
via data fusion, enabling early detection as fast as 48 hr
post-infection, far before visible symptoms appear (5–11
days), proving hyperspectral imaging with ML as a
powerful non-invasive tool for precision agriculture (80).
Deep learning approaches in data analysis
Conventional ML approaches frequently rely on
predetermined features, while deep learning models
spontaneously uncover significant patterns and
representations in data throughout the learning process
to solve a given task. This makes deep learning suitable
for analysing hyperspectral images with large datasets
containing hundreds of spectral bands for detecting
specific diseases. Deep learning approaches handle the
challenges of hyperspectral data: higher dimensionality
and limited labelled training data and automatically learn
features without human guidance (16).
Non-imaging hyperspectral signatures of
healthy and diseased (S. tritici and S. nodorum) wheat
tissues were distinguished from each other by studying
the red-edge reflectance patterns (705–745 nm) of the
wheat. The high spectral variability was linked with the
severity of the disease. A perceptron neural network
was capable of correctly classifying 88–99 % of cases
and predicted yield losses (35–45 % for severe STB, 20
–40 % for severe SNB). The results offer a fast,
objective method for later-stage disease assessment
and crop loss estimation in wheat in the field (20).
Application of back propagation neural
networks (BNN) on second-derivative-transformed
hyperspectral (VNIR/SWIR) data of oilseed rape leaves
to classify Alternaria fungal infection, achieving 90.5 %
accuracy for identifying days after inoculation and 80.5
% for distinguishing fungal species (81).
Artificial neural network (ANN)-based
hyperspectral-based plant disease detection shows
exceptional performance by providing near-perfect
classification accuracy for presymptomatic oil palm
disease by incorporating both spectral and weather data
as model inputs. However, accuracy for the paired
healthy class was as low as 66.7 % and the dataset had
fewer than 50 spectral measurements per class, so this
should be read as a preliminary, class-imbalanced result
rather than a generalisable finding. Among the various
ANN types evaluated, BPNN and extreme learning
machine (ELM) were able to classify healthy versus
diseased tobacco leaves with up to 95 % accuracy,
while multilayer perceptron using multi-layer perceptron
(MLP)-autoreggressive distributed lag (ARD) or
automatic relevance determination and XY-fused
networks (spectral (X) and spatial (F) information) were
able to detect tomato chlorosis virus with over 85 %
accuracy. These results collectively demonstrated ANN
as a trusted and adaptable technique for early plant
disease, through claims of perfect accuracy warrant
scrutiny of sample size and validation design (9).
Automatic extraction of hierarchical spatial
features from both hyperspectral and RGB images,
CNNs have been used to detect yellow rust in wheat.
This method offers superior classification accuracy by
learning intricate spectral-spatial patterns across
different stages of disease progression. The two hidden
layer MLP utilising bayesian regularisation was
implemented, incorporating spectral band data and
lesion index as input features to effectively categorise
wheat canopy into healthy or diseased classes (82).
Three dimensional (3D) deep convolutional
neural network (3D-DCNN) extracts spatial and spectral
features from hyperspectral image cubes to classify
soybean stems as healthy or charcoal rot infected. A
visualisation technique based on saliency maps is
employed to determine the most sensitive pixel
locations and hyperspectral wavelengths (particularly in
the NIR region) that make significant contribution to
classification accuracy (83, 84).
The outlier removal auxiliary classifier
generative adversarial nets (OR-AC-GAN) model, which
is based on generative adversarial nets (GAN), is an all-
in-one framework that integrates spectrum classification,
image classification and plant segmentation to detect
tomato spotted wilt virus (TSWV) from hyperspectral
images even before symptoms become apparent. The
GAN model consists of two components: 1) the
generator increases the quantity of the training dataset
and enhances the precision of current disease detection
techniques by producing artificial spectrum data that
resembles real plant data. 2) the discriminator, which
serves as a filter to reject inaccurate or noisy data,
making sure that healthy plants are rarely mistakenly
identified as unhealthy (85). Fig. 2 shows the overall
process for hyperspectral-based plant disease
detection.
Conventional ML classifiers (SVM, RF) are
appropriate for specific, well-defined disease
classifications because they often perform well on
limited, carefully selected datasets and provide
interpretability and minimal computational cost. On
complicated, multi-class, or early-stage detection tasks,
DL techniques (CNN, 3D-DCNN, GAN) perform better
than ML; nevertheless, their field deployment in data-
scarce crop systems is limited because they require
larger labelled datasets and more processing power.
Applications of hyperspectral technology in plant
disease diagnosis
Plant disease diagnosis mainly involves detection,
differentiation, identification, evaluation of disease
severity and evaluation of genotype resistance. These
are carried out using corresponding measurement
scales at various levels of biological organisation. In
laboratories, hyperspectral microscopes are used to
Plant Science Today | 2026 | Volume 13(sp9) | Page 21
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.17031
observe fungal spores and identify metabolic alterations
brought on by plant-pathogen interactions at the cellular
and tissue levels.
These studies are mainly conducted for
fundamental research, pathogen identification and
genotype resistance evaluation. At the individual organ
level (such as leaves, stems, roots, fruits and ears) and
the whole-plant level, hyperspectral measurements are
conducted in laboratories, greenhouses and field
settings to enable early disease detection and
differentiation. Hyperspectral technologies are widely
applied for large-scale disease mapping and field
conditions to enable early disease detection and
differentiation at the canopy level. Research has shown
that hyperspectral technologies are useful for
diagnosing plant diseases in t
论文原文与出版信息
Plant Science Today