论文信息

Ashish Chaudhary / Swas Kaushal / Het Samir Desai 等 10 人

Frontiers in Agronomy

论文解读

AI 辅助整理
方法
综述机器学习/深度学习、遥感、图像处理、杂草检测算法与靶向除草系统。
发现
AI精准技术可减少化学投入、分散选择压、提升农业生态系统韧性。
机会点
可研究物种特异性除草决策与生态友好型杂草群落保护的权衡模型。

论文摘要

杂草是全球作物生产的主要制约因素。对除草剂的过度依赖已引发环境担忧、抗除草剂杂草种群的筛选以及农业生态系统韧性下降。人工智能(AI)的最新进展,包括机器学习和深度学习,以及遥感、机器人和自动化领域的创新,正推动杂草管理从统一的田块尺度干预转向基于位置的植株水平决策。本综述综合了AI赋能精准杂草管理的当前进展,重点关注传感技术、图像处理流程、杂草检测算法和靶向杂草控制系统。我们还探讨了AI辅助的除草剂响应表型分析,包括作物药害、非靶标飘移和抗除草剂性评估。我们讨论了商业化和新兴的精准杂草管理技术,包括靶向除草剂喷洒、机器人机械除草和基于激光的杂草控制。除操作效率外,我们还提出了关于AI驱动的精准技术如何通过实现物种特异性干预——在抑制高竞争力杂草的同时保护生态促进性或中性植物群落——来支持向农业生态杂草管理范式转变的观点。总体而言,这些进展使AI赋能的精准杂草管理能够减少化学投入、多样化选择压力、增强农业生态系统韧性并支持可持续农业集约化。

全文

Frontiers in Agronomy

Towards sustainable weed
management: the role of arti ficial
intelligence, sensing, and robotics
Ashish Chaudhary1, Swas Kaushal1, Het Samir Desai2,
Navjot Singh3, Muskan Jakhar1, Devanshi Het Desai2,
Jaswinder Kaur4, Komal Mahajan1, Matthew Vredenburg1
and Sachin Dhanda1*
1Department of Agronomy, Horticulture and Plant Science, South Dakota State University, Brookings,
SD, United States, 2Department of Land Resources and Environmental Sciences, Montana State
University, Bozeman, MT, United States, 3Department of Agronomy and Plant Genetics, University of
Minnesota, St. Paul, MN, United States, 4Department of Crop and Soil Sciences, North Carolina State
University, Raleigh, NC, United States
Weeds are a major constraint to crop production worldwide. Overreliance on
herbicides has led to environmental concerns, selection of herbicide-resistant
weed populations, and reduced agroecosystem resilience. Recent advances in
artificial intelligence (AI), including machine learning and deep learning, along
with innovations in remote sensing, robotics, and automation, are shifting weed
management from uniform field-scale interventions to site-speci fic, plant-level
decision-making. This review synthesizes current developments in AI-enabled
precision weed management, with emphasis on sensing technologies, image-
processing work flows, weed detection algorithms, and targeted weed-control
systems. We also consider AI-assisted herbicide-response phenotyping, including
crop injury, off-target drift, and herbicide-resistance assessment. We discuss
commercial and emerging precision weed-management technologies, including
targeted herbicide spraying, robotic mechanical weeding, and laser-based weed
control. Beyond operational ef ficiency, we provide a perspective on how AI-
driven precision technologies can support a paradigm shift toward agroecological
weed management by enabling species-speci fic interventions that suppress
highly competitive weeds while conserving ecologically facilitative or neutral
plant communities. Collectively, these advances position AI-enabled precision
weed management for reducing chemical inputs, diversifying selection pressure,
enhancing agroecosystem resilience , and supporting sustainable agricul-
tural intensi fication.
KEYWORDS
agroecological weed management, deep learning, herbicide resistance, machine
learning, neural networks, precision weed management, site-speci fic weed control
1 Introduction
With the global population projected to reach 9.7 billion by 2050, ensuring sufficient and
sustainable food production has become one of the most critical challenges of the twenty-
first century ( Hamad and Tayel, 2026 ). This urgency is re flected by the current state of
global food security, with approximately 673 million people experiencing hunger and more
Frontiers in Agronomy frontiersin.org01
OPEN ACCESS
EDITED BY
Marija Popovic,
Delft University of Technology,
Netherlands
REVIEWED BY
Ke Xu,
Anhui Polytechnic University, China
Zhongzhong Niu,
Hewlett-Packard, United States
*CORRESPONDENCE
Sachin Dhanda
sachin.dhanda@sdstate.edu
RECEIVED 02 June 2026
REVISED 05 October 2026
ACCEPTED 26 October 2026
PUBLISHED 13 November 2026
CITATION
Chaudhary A, Kaushal S, Desai HS,
Singh N, Jakhar M, Desai DH, Kaur J,
Mahajan K, Vredenburg M and Dhanda S
(2026) Towards sustainable weed
management: the role of arti ficial
intelligence, sensing, and robotics.
Front. Agron. 8:1898506.
doi: 10.3389/fagro.2026.1898506
COPYRIGHT
© 2026 Chaudhary, Kaushal, Desai, Singh,
Jakhar, Desai, Kaur, Mahajan, Vredenburg
and Dhanda. This is an open-access
article distributed under the terms of the
Creative Commons Attribution License
(CC BY). The use, distribution or
reproduction in other forums is
permitted, provided the original
author(s) and the copyright owner(s) are
credited and that the original publication
in this journal is cited, in accordance
with accepted academic practice. No
use, distribution or reproduction is
permitted which does not comply with
these terms.
TYPE Review
PUBLISHED 13 November 2026
DOI 10.3389/fagro.2026.1898506

than 2.3 billion facing moderate or severe food insecurity in 2024
(FAO, 2025). Among the many constraints to agricultural produc-
tivity, weeds remain one of the most signi ficant biotic stressors.
Weeds compete with crops for light, water, nutrients, and space
while also serving as hosts for pests and diseases, ultimately
reducing crop yield and quality ( Soltani et al., 2017 ). Globally,
nearly 1,800 weed species cause an average 31.5% reduction in plant
production, resulting in an estimated annual economic loss of US
$32 billion (Kubiak et al., 2022). Uncontrolled weed growth can lead
to substantial yield losses or complete crop failure ( Gharde et al.,
2018). As global food production must increase by 50-60% by 2050
to meet growing demand, effective and sustainable weed manage-
ment will play a critical role in ensuring future food security ( Falcon
et al., 2022 ).
Herbicides have been widely used for weed management for
more than half a century because of their effectiveness, scalability,
and economic bene fits. However, conventional broadcast herbicide
applications treat entire fields regardless of weed distribution, often
leading to unnecessary herbicide use in weed-free areas. This
approach not only increases production costs but also contributes
to ecosystem disruption and accelerates the evolution of herbicide-
resistant weeds. To date, 548 unique cases of herbicide resistance
have been reported worldwide, involving 275 weed species resistant
to 21 of the 31 known herbicide sites of action across 102 crops in 76
countries ( Heap, 2026 ). Furthermore, excessive herbicide use can
adversely affect soil health, biodiversity, water quality, and human
health ( Bamal et al., 2024 ). These challenges have intensi fied
interest in precision weed management strategies that account for
the spatial variability of weed infestations and enable site-speci fic
interventions ( Huang et al., 2025 ; Zhang et al., 2022 ).
Targeted weed management seeks to apply control measures
only where weeds are present, thereby reducing inputs while
maintaining weed control ef ficacy. Advances in remote sensing,
computer vision, arti ficial intelligence (AI), robotics, and automa-
tion have accelerated the transition from uniform weed manage-
ment practices to site-speci fic and plant-level interventions
(Bagavathiannan and von Redwitz, 2025 ; Esposito et al., 2021 ).
Remote sensing platforms, including satellites, unmanned aerial
vehicles (UAVs), and ground-based sensors equipped with RGB,
multispectral, hyperspectral, and LiDAR technologies, now provide
high-resolution data capable of capturing the spatial heterogeneity
of weed populations within agricultural fields (Mogili and Deepak,
2018; Velusamy et al., 2023 ). Concurrently, machine learning (ML)
and deep learning (DL) algorithms have transformed agricultural
image analysis by enabling accurate weed identi fication, localiza-
tion, and classification under complex field conditions (Benos et al.,
2021; Rai et al., 2023 ).
The integration of AI with autonomous machinery has ex-
panded the scope of precision weed management beyond herbicide
application. AI-enabled technologies now support a broad range of
targeted tactics, including robotic mechanical weeders, laser-based
systems, and other site-speci fic interventions ( Steward et al., 2019 ;
Vijayakumar et al., 2025 ). These technologies can reduce depen-
dence on herbicides, delay the evolution of herbicide resistance,
lower environmental impacts, and improve the economic sustain-
ability of weed management. Re flecting their growing importance,
“Precision Weed Management and Robotics ” was recently identi-
fied as a top research priority by weed scientists worldwide
(Brainard et al., 2023 ). Beyond improving the ef ficiency of weed
control, emerging precision technologies also create new opportu-
nities to align weed management with agroecological principles
(Bagavathiannan and von Redwitz, 2025 ; Esposito et al., 2023 ).
Agroecological weed management recognizes that not all weed
species exert equal competitive effects and that functionally diverse
weed communities can provide ecosystem services ( Boinot et al.,
2024; Desai et al., 2026 ). By enabling species-speci fic detection and
plant-level interventions, AI-driven sensing, robotics, and precision
application technologies may support a shift from uniform weed
eradication to more selective, ecologically informed management
strategies that suppress highly competitive weeds while preserving
facilitative or neutral species.
Several recent reviews have addressed components of this field,
including deep learning ( Rai et al., 2023 ), deep learning-based
weed-crop recognition ( Qu and Su, 2024 ), drone and sensor
technologies for weed management ( Esposito et al., 2021 ; Roslim
et al., 2021 ), remote sensing ( Huang et al., 2025 ), robotic ap-
proaches ( Zhang et al., 2022 ), and intelligent mechanical weeding
(Das et al., 2025 ). More recently, the idea that precision agriculture
could selectively remove problematic weeds while retaining less
competitive or ecologically valuable species has also begun to
appear in the literature ( Esposito et al., 2023 ; Bagavathiannan and
von Redwitz, 2025 ). Building on these advances, the present review
is not intended as another catalog of sensors and AI models.
Instead, it connects these strands into a single, end-to-end frame-
work linking sensing, AI-based decision-making, selective actua-
tion, and agroecological management objectives, and extends the
discussion beyond weed detection to herbicide-re-
sponse phenotyping.
To this end, we synthesize recent advances in AI-, ML-, and
robotics-enabled weed management with four specific objectives: (i)
to summarize image processing work flow, (ii) to examine the
functionality and practical applications of integrating sensing tech-
nologies with data acquisition platforms and deep learning models
in weed detection and mapping, (iii) to highlight the commercial
precision-based chemical and non-chemical weed management
technologies, including precision spraying, mechanical weeding,
and laser weeding, and (iv) to provide a perspective on how these
technologies can support agroecological weed management.
2 Review methodology
This article is structured as a narrative review aimed at synthe-
sizing technological mechanisms and agroecological frameworks.
Our perspective-oriented synthesis draws on literature identi fied
through Web of Science, Scopus, and Google Scholar, supplemented
by publisher databases and the references of key articles. We
combined weed-management terms (e.g., weed detection, precision
and site-specific weed management, smart spraying, robotic and laser
weeding, herbicide resistance, and agroecological weed management)
with technology terms (e.g., machine learning, deep learning, CNN,
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org02

transformer, hyperspectral, multispectral, UAV, remote sensing, and
sensor fusion). We synthesized information from peer-reviewed
studies from 2000-2026, prioritizing those reporting quantitative
performance, field-scale evaluation, or conceptual advances.
3 Sensing technologies and data
processing work flows for weed
detection
3.1 Sensing technologies and data
acquisition platforms
Imaging and sensor technologies are mounted on data acqui-
sition platforms for field-level operations, providing the founda-
tional data for machine-learning and deep-learning models used in
weed detection, classi fication, mapping, and site-speci fic manage-
ment. While sensing technologies include RGB, multispectral,
hyperspectral, fluorescence, ultrasonic, and LiDAR systems, data
acquisition platforms range from ground-based vehicles and
smartphones to unmanned aerial vehicles (UAVs). These technol-
ogies differ in cost, spectral resolution, data complexity, and
suitability for real-time deployment. Here, we provide a detailed
workflow for the required processing, information on the function-
ality, feasibility, and limitations of these imaging platforms, and
highlight successful attempts to detect weeds and crop injury. Also,
we emphasize multimodal sensor fusion, in which spectral, spatial,
structural, textural, and temporal information are integrated to
improve weed detection accuracy under variable field conditions (El
Sakka et al., 2025 ). The characteristics of the major sensing
technologies compared in this review are summarized in Table 1.
3.2 Image processing pipeline
Following the capture of high-resolution field images using
different sensors, a series of image-processing steps are performed
to convert raw images with overlapping layers into analysis-ready
data suitable for model development (Figure 1). Because each image
captures only a small portion of the field, individual images cannot
be used directly for spatial analysis. Therefore, the overlapped
images are processed using photogrammetric software (Pix4D
Mapper or Pix4D Field), although open-source alternatives such
as OpenDroneMap are also available to obtain a single, highly
accurate map of the study area.
For UAV-based imagery, Ground Control Points (GCPs) are
clearly visible markers commonly established prior to flight to
support accurate georeferencing, although modern real-time kine-
matic (RTK) positioning can reduce or, in some work flows, elim-
inate the need for GCPs. In agricultural fields, GCPs are usually
placed at plot corners to ensure accurate spatial positioning and
correct image distortion, enabling precise analysis, mapping, and
interpretation of remotely sensed data. These markers are usually
square and feature a highly contrasting black-and-white chessboard
pattern for easy recognition on high-contrast backgrounds. The
coordinates of GCPs can be acquired using RTK Global Navigation
TABLE 1 Comparison of the major sensing technologies used for AI-assisted weed detection and herbicide-response phenotyping.
Sensor
type
Spatial
resolution
Spectral
information
Relative
cost
Calibration
needs
Data
volume/
complexity
Real-
time
suitability
Key strengths and limitations
RGB High (visible) Three broad
visible bands (R,
G, B)
Low Minimal Low High Low cost and simple to deploy; limited by
illumination, occlusion and soil
background; poor for spectrally similar
species.
Multispectral High-
moderate
Several discrete
bands, incl. NIR
and red-edge
Moderate Radiometric
calibration
required
Moderate Moderate Adds vegetation and stress information;
requires calibration and careful flight
planning.
Hyperspectral Moderate Hundreds of
contiguous
narrow bands
High Extensive
calibration
High (3-D
spectral cube)
Low
(currently)
Detects subtle biochemical and
physiological changes before visible
symptoms; costly, data-heavy and mostly
research-scale.
Fluorescence Moderate Chlorophyll
fluorescence
(e.g., Fv/Fm,
FPSII, NPQ)
Moderate-
high
Controlled light/
dark protocols
Moderate Low-
moderate
Direct, early readout of PSII/
photosynthetic stress and herbicide mode
of action; sensitive to ambient light and
often needs controlled conditions.
Thermal Moderate Canopy
temperature
(thermal
infrared)
Moderate Emissivity and
ambient
correction
Low-moderate Moderate Captures stomatal/water-status
differences; strongly affected by weather
and time of day.
LiDAR/depth High
(structural)
3-D structure
and plant height
(no spectral)
Moderate-
high
Sensor-specific
calibration
Moderate-high Moderate Adds height and canopy architecture for
crop-row and canopy separation; limited
biochemical information.
Ultrasonic Low-moderate Canopy height/
presence (no
spectral)
Low Minimal Low High Inexpensive height/presence sensing;
coarse and limited for species
discrimination.
Qualitative ratings (low/moderate/high) are indicative and vary with platform and application.
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org03

Satellite Systems receivers and L iDAR-based surveys, whereas
lower-accuracy sources such as Google Earth provide only approx-
imate positions and are not appropriate when survey-grade accu-
racy is required. These surveyed GCP coordinates are used to align
imagery during stitching, imp roving the spatial accuracy of
orthomosaic images in photogrammetric processing.
3.2.1 Image stitching using photogrammetric
software
Pix4D Mapper is widely used photogrammetric software for
stitching images and generating orthomosaics due to its survey-
grade accuracy, robust multicamera compatibility, and high-density
output. Prior to image processing, calibrated re flectance panel
images, along with all UAV-acquired images, are imported for
radiometric calibration, which compensates for changing light
conditions during flight. After importing the imagery, the standard
Ag Multispectral processing template is typically used. This tem-
plate is specifically designed for an agricultural multispectral dataset
to generate a re flectance map, an index map, and an application
map for precision agriculture app lications. This template can
automatically con figure the work flow, which consists of three
main processing stages: (1) initial processing, (2) point cloud and
mesh generation, and (3) Digital Surface Model (DSM)
orthomosaic and Re flectance Map generation.
3.2.2 Image classi fication
ArcGIS (used here as a representative example, with open-
source alternatives such as QGIS also available) is widely used for
spatial analysis in precision agriculture due to its comprehensive
geoprocessing capabilities, integrated 2D and 3D analysis tools, and
support for machine- and deep-learning work flows. A Pix4D-
generated orthomosaic (GeoTIFF) can be imported into ArcGIS
for spatial analysis, training data development, classi fication, and
vegetation index calculation, depending on the sensor. Pixel-based
classification and object-based image analysis (OBIA) are two major
approaches widely adopted for separating vegetation from non-
vegetation areas and reducing background noise from soil and
residue, using training samples and machine learning ( Figure 1 ).
In pixel-based classification, each pixel is treated as an individual
observation, and a speci fic class is assigned based on its re flectance
value derived from vegetation indices or spectral bands. In ArcGIS
Pro, this can be performed using supervised classi fication methods,
including Random Forest (RF), Support Vector Machine (SVM), and
Maximum Likelihood classification (MLC). Training samples such as
crop, weed, soil, and crop residue are classi fied manually, and then
the model is trained and applied to the Orthomosaic. Among these
different classification approaches, RF has been most widely adopted
because it can handle high-dimensional multispectral data, non-
linear relationships, and correlated predictor variables while main-
taining higher classification accuracy (Breiman, 2001).
Object-based image analysis (OBIA) classi fies an image by
grouping neighboring pixels based on their shape, texture, spatial,
and spectral information ( Baatz, 2000 ). This method typically in-
volves two primary phases: (1) image segmentation and (2) object
classification. In segmentation, neighboring pixels with similar char-
acteristics are grouped into homogeneous objects using a segmenta-
tion algorithm, such as the mean shift method, among others.
Classification includes image annotation by drawing polygons over
different land-cover classes (crop, weeds, soil) and supervised ma-
chine learning methods such as RF and SVM to classify the
segmented objects. Object-based image analysis can also incorporate
crop-row structure, canopy texture, and plant height derived from a
DSM orthomosaic, thereby reducing the salt-and-pepper effect com-
monly observed in pixel-based classification (Pu et al., 2011 ).
Stephen and Kumar (2023) compared the RF, SVM, and OBIA
for mapping weeds in sugar beetfields and reported overall accuracies
FIGURE 1
Image processing work flow from data acquisition to model development.
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org04

of 95%, 96%, and 95%, respectively. However, the Kappa coef ficient
was higher for SVM (0.93) and RF (0.92) than for OBIA (0.81),
suggesting that the pixel-based classification approach provides more
reliable results for crop-weed discrimination than the object-based
classification method. A similar study on weed density mapping in
cotton ( Gossypium hirsutum L.) fields using high-resolution RGB
images combined with Hough Transformation, OBIA, and machine
learning classification reported overall accuracy of 89.16%, 85.83%,
and 83.33%, and Kappa values of 0.84, 0.79, and 0.75, in high,
medium, and low-density weed infestation, respectively ( Sapkota
et al., 2020). Gao et al. (2018) demonstrated the integration approach
of a pixel-based classi fier with OBIA to map weeds in maize that
achieved 94.5% overall accuracy with a 0.91 Kappa value using
Hough transformation followed by RF classi fication.
3.2.3 Annotation and model development
Recently, advancements in arti ficial intelligence have enabled
the integration of UAV imagery with deep learning-based semantic
segmentation for crop-weed discrimination, combining pixel- and
object-based approaches ( Sa et al., 2018 ). Following image process-
ing and classi fication in GIS software, Geo TIFF orthomosaic
images are typically divided into smaller tiles and annotated using
platforms such as CVAT, Robo flow, or Label Studio ( Figure 1 ).
These annotation tools enable users to manually annotate crops,
weeds, soil, and crop residue, or even individual crop-weed species,
using polygon masks, and generating a training dataset for seg-
mentation. These annotated datasets can be exported in COCO,
Pascal VOC, YOLO, or segmentation mask and imported into
Python-based environments and deep learning frameworks (e.g.,
Anaconda, PyTorch, or TensorFlow) for model development. Prior
to training, model robustness is commonly improved using data
augmentation techniques such as image rotation, flipping, scaling,
cropping, and brightness or contrast adjustment to increase dataset
diversity (Figure 1). Semantic segmentation models such as U-Net,
U-Net++, DeepLabV3+, SegNet assign a class label to every pixel,
whereas Mask R-CNN performs instance segmentation by deline-
ating individual plant instances; Transformer-based architectures
are increasingly applied to both tasks. Model performance is
commonly assessed using parameters such as overall accuracy,
precision, recall, F1-score, and Intersection Over Union (IoU).
Once validated, the trained model can be used to generate pixel-
level crop-weed-segmentation maps, a weed density map, and site-
specific herbicide application maps f or precision weed management.
Furthermore, thesedeep learning-based models can also be deployed
for real-time crop-weed detectionon UAVs, sprayers, and autonomous
robots. In the sugarcane system, six deep learning models, including
AlexNet, GoogLeNet, Inceptio nV3, ResNet50, Xception, and
DarkNet53, were compared to map weeds using field-acquired RGB
imagery and found that the DarkNe t53 model achieved the highest
accuracy (>99%) and F1 score (>99 %) with the lowest error (<1%)
(Modi et al., 2023 ). Furthermore, Sunil et al. (2022) evaluated the
CNNs VGG16 and ResNet50 for weed detection on imagery with
various backgrounds and found thatboth achieved F1 scores ranging
from 92 to 99%, indicating robust weed classi fication across diverse
backgrounds. Representative AI models and datasets discussed in this
review are summarized in Table 2.
3.3 Integration of RGB and data acquisition
platforms
Red-green-blue imaging is a widely used sensing technology for
digital weed detection and herbicide-based crop injury due to its
low cost, ease of deployment, and compatibility with data acquisi-
tion platforms ( Qu and Su, 2024 ). The RGB camera captures
information across the visible red, green, and blue wavelengths,
enabling the extraction of color, texture, canopy structure, and
spatial context of the vegetation ( Roslim et al., 2021 ). The RGB-
based models are especially attractive for real-time applications
because they require relatively inexpensive sensors and generate
smaller datasets than multispectral or hyperspectral systems.
However, RGB image sensors are only effective at distinguishing
crops from weeds based on visual characteristics.
Peña et al. (2015) demonstrated that the detection of weed
plants using RGB-mounted UAVs required 1 –4 cm pixel -1 spatial
resolution and a flight altitude of 40–100 m. Similarly, weed patches
were also found to be distinguishable from crop fields using an RGB
camera mounted on a multirotor UAV, with a required spatial
resolution of <5 cm pixel -1 and a flight altitude of <50 m ( Huang
et al., 2018a ). Abrantes et al. (2021) successfully detected dicamba-
mediated drifts on soybean ( Glycine max (L.) Merr.) by assessing
spatial differences in vegetative indices obtained from an RGB
camera integrated with a UAV. İeri (2025) showed the reasonable
accuracy of smartphone-based RGB imagery (Canopeo mobile
application) in quantifying vegetation ground cover percentage.
Despite these successful attempts, the performance of RGB imagery
is highly contingent on variable illumination across phases of the
day, occlusion, soil background i nterference, and overlapping
canopies ( Roslim et al., 2021 ; Rai et al., 2023 ; Qu and Su, 2024 ).
For example, Papadopoulos et al. (2018) concluded that images
captured at a certain time of day using RGB-mounted ground-based
platforms compromised the accuracy of percent weed cover esti-
mates compared with ground-truth data. The development of deep
learning has substantially enhanced the utility of RGB imagery.
Convolutional neural networks (CNNs) and object detection
models such as YOLO and Faster R-CNN can identify weed targets
from RGB images, while segmentation models such as U-Net and
DeepLab can delineate weed pixels or plant boundaries ( Qu and Su,
2024; Rai et al., 2023 ).
3.4 Integration of multispectral imagery
and data acquisition platforms
Multispectral sensing extends RGB imaging by adding spectral
bands, commonly near-infrared and red-edge wavelengths. These
bands are sensitive to vegetation structure, chlorophyll content,
canopy vigor, and plant stress, enabling better discrimination
between crops and weeds when visual differences are limited
(Singh et al., 2026 ). Multispectral imagery is widely used for weed
patch mapping, crop-weed separation, and variable-rate herbicide
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org05

applications. UAV-mounted multispectral sensors are particularly
valuable because they combine high spatial resolution with biolog-
ically informative spectral data ( Roslim et al., 2021 ).
Two decades ago, aerial multispectral imagery was used to
detect green foxtail [ Setaria viridis (L.) Beauv] and velvetleaf
(Abutilon theophrasti Medik.) in a soybean field with >60% accu-
racy (Gibson et al., 2004 ). Peña et al. (2015) noted that lower flight
altitudes should be used for multispectral-mounted UAVs com-
pared with RGB to improve weed detection accuracy. Betitame et al.
(2024) evaluated the dose-response relationship between dicamba
drift and soybean injury with moderate yet meaningful accuracy
(46%) using a multispectral camera on a small UAV. Pérez-Ortiz
et al. (2015) integrated a multispectral camera on a quadcopter to
classify crops, weeds, and soil in a sun flower (Helianthus annuus L.)
field. Various other studies have indicated that airborne multispec-
tral imagery is widely used for vegetation mapping, including weeds
in crop fields (de Castro et al., 2012; Goel et al., 2002; Sa et al., 2018).
Despite these promising research-level attempts, commercial adop-
tion may be constrained by a lack of expertise in operating high-
quality multispectral cameras that require radiometric calibration,
careful flight planning, and larger data-processing work flows.
3.5 Integration of hyperspectral imaging
and data acquisition platforms
Hyperspectral imagery captures hundreds of narrow, contigu-
ous spectral bands, generating a detailed spectral signature for each
pixel. This enables characterization of plant biochemical and
physiological properties, including chlorophyll, water, and nutrient
content ( Lu et al., 2020 ). Unlike multispectral imagery,
hyperspectral imagery produces continuous spectral data for each
pixel, forming a three-dimensional hyperspectral cube. This allows
much finer discrimination between crops and weeds, even when
they appear similar in RGB images. Hyperspectral imaging has been
successfully applied t o crop-weed classi fication, invasive weed
detection, species-level discrimination, and early stress or herbicide
response assessment. When comb ined with machine learning
algorithms, hyperspectral data can achieve high classi fication accu-
racy by exploiting subtle spectral differences among plant species
(Sulaiman et al., 2022 ). A key advantage of hyperspectral imaging is
its sensitivity to subtle physiological and biochemical changes that
precede visible symptoms. Herbicide-induced stress can alter leaf
spectral signatures through changes in chlorophyll and carotenoid
content, water status, photosynthetic activity, cellular structure, and
canopy morphology, producing measurable shifts particularly in the
visible, red-edge, and near-infrared regions often within hours to
days of exposure and before injury is visible to the naked eye ( Lu
et al., 2020 ). This early, non-destructive sensitivity underpins the
use of hyperspectral imaging not only for crop-weed discrimination
but also for herbicide-resistance screening and herbicide-re-
sponse phenotyping.
Hyperspectral imagery from handheld ground-based systems
was used to discriminate among multiple broadleaf weed species in
a soybean field, achieving an average accuracy of 65% ( Henry et al.,
TABLE 2 Representative arti ficial-intelligence models and datasets for weed detection, classi fication and herbicide-response phenotyping discussed in
this review.
Study Crop/weed Sensor and platform Model/
method
Task Reported
performance
Setting
Stephen and
Kumar (2023)
Sugar beet/weeds UAV RGB RF, SVM, OBIA Weed mapping 95-96% overall accuracy;
Kappa 0.81-0.93
Field
Sapkota et al.
(2020)
Cotton/weeds UAV RGB Hough transform +
OBIA + ML
Weed density mapping 83-89% overall accuracy;
Kappa 0.75-0.84
Field
Gao et al. (2018) Maize/weeds UAV imagery Pixel + OBIA
(Hough + RF)
Weed mapping 94.5% overall accuracy;
Kappa 0.91
Field
Modi et al.
(2023)
Sugarcane/weeds Field RGB Six CNNs (AlexNet-
DarkNet53)
Weed identi fication DarkNet53 >99% accuracy
and F1
Field
Sunil et al.
(2022)
Crop/weed
species
RGB (varied backgrounds) VGG16, ResNet50 Weed classi fication F1 92-99% Controlled/
varied
Reddy et al.
(2014)
Palmer amaranth Ground hyperspectral Spectral +
classification
Resistant vs susceptible
discrimination
94% (greenhouse), 96%
(field)
Greenhouse
& field
Singh et al.
(2026)
Wild oat Hyperspectral (204 bands) Machine learning Herbicide-resistance
classification
~88% Controlled/
field
Zhang et al.
(2021)
Sod/monocot &
dicot weeds
UAV hyperspectral CNN Weed classi fication 57-90% Field
Xia et al. (2023) Barnyardgrass,
velvetleaf
UAV hyperspectral + RGB
(Phantom 4 RTK)
Deep fusion network Resistance
discrimination
78% Field
Eide et al. (2021) Redroot pigweed UAV thermal + multispectral Random tree Glyphosate-resistance
discrimination
87% Field
Upadhyay et al.
(2024)
Row-crop weeds Machine vision (edge) YOLOv4 Detection for site-
specific spraying
100% (indoor), 95%
(outdoor) precision
Indoor &
field
Deng et al.
(2025)
Vegetable weeds Machine vision YOLOv11 Real-time detection 99% mAP (indoor), 69%
mAP ( field)
Indoor &
field
Reported performance is as stated in the cited studies.
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org06

2004). The same study achieved 97% accuracy in classifying
soybeans. Reddy et al. (2014) evaluated a ground-based
hyperspectral camera to identify glyphosate-resistant and
glyphosate-susceptible Palmer amaranth ( Amaranthus palmeri S.
Watson) plants with 94% and 96% accuracy in the greenhouse and
field conditions, respectively. Huang et al. (2018b) and Huang et al.
(2022) further tested a similar approach to identify glyphosate-
resistant and glyphosate-sus ceptible Italian ryegrass ( Lolium
perenne L. ssp. multiflorum (Lam.) Husnot) and Johnsongrass
[Sorghum halepense (L.) Pers.] populations. A ground-based
system with a hyperspectral camera was able to detect glyphosate-
and dicamba-resistant kochia [ Bassia scoparia (L.) A. J. Scott]
populations associated with various crops with a >65% success
rate (Nugent et al., 2018). Singh et al. (2026) used 204 hyperspectral
bands ranging from 400 to 1000 nm wavelength to differentiate
fenoxaprop- and imazamox-resistant wild oat ( Avena fatua L.)
populations with ~88% accuracy.
While the aforementioned studies demonstrated the feasibility
of hyperspectral imagery for weed detection, its practical feasibility
for routine field-scale deployment remains challenging due to the
considerable expense, sensitivity to illumination variability, and the
generation of large, high-dimensional datasets that require exten-
sive calibration and computational processing ( Lu et al., 2020 ;
Mensah et al., 2025 ). Dimensionality reduction, feature selection,
and advanced deep learning methods are often necessary to reduce
computational burden and prevent over fitting ( Sulaiman et al.,
2022). Consequently, hyperspectral imaging is currently most
valuable for research applications, model development, and con-
trolled experiments, whereas RGB and multispectral systems are
more feasible for operational weed management.
3.6 Fluorescence imaging
Chlorophyll fluorescence imaging provides a direct, non-de-
structive window into photosynthetic performance and is particu-
larly valuable for detecting herbicides that impair photosynthetic
electron transport. When a leaf absorbs light, a small fraction of the
energy is re-emitted as chlorophyll fluorescence, and changes in this
signal re flect the ef ficiency of photosystem II (PSII). Widely used
parameters include the maximum quantum yield of PSII (Fv/Fm),
the effective quantum yield of PSII ( FPSII), and non-photochem-
ical quenching (NPQ). Because photosynthesis-inhibiting herbi-
cides (e.g., PSII inhibitors such as triazines and substituted ureas)
disrupt electron transport almost immediately, fluorescence param-
eters can reveal herbicide stress within hours of application, well
before visible injury develops, and the response patterns can help
distinguish herbicide modes of action ( Hassannejad et al., 2020 ).
Imaging-based fluorescence systems extend these point measure-
ments to the whole-leaf or canopy scale, enabling spatially resolved
detection of early stress, differentiation of resistant and susceptible
biotypes, and high-throughput screening. When combined with
machine learning, fluorescence and hyperspectral data can be fused
to improve the sensitivity and speci ficity of early herbicide-
stress detection.
3.7 AI-assisted herbicide-response
phenotyping
Beyond distinguishing crops from weeds, AI-assisted sensing is
increasingly used for herbicide-response phenotyping, defined as the
quantitative characterization of how plants respond to herbicides.
This emerging area includes crop-injury assessment, off-target drift
diagnosis, herbicide-resistance characterization, and discrimination
of herbicide sites or modes of action. Hyperspectral imaging com-
bined with machine learning has been used to quantify dicamba
injury on non-dicamba-tolerant soybean and to distinguish recover-
able from unrecoverable damage with overall accuracies above 90%
(Zhang et al., 2019 ). Shallow convolutional neural networks applied
to hyperspectral data predicted early herbicide stress in wheat
cultivars before visible symptoms appeared ( Chu et al., 2022 ), while
high-throughput hyperspectral imaging discriminated among herbi-
cide sites of action as early as one day after treatment, offering a novel
route to accelerate herbicide discovery (Niu et al., 2023). Chlorophyll
fluorescence measurements have similarly been used to identify
herbicide modes of action from characteristic changes in photosyn-
thetic parameters ( Hassannejad et al., 2020 ). Together with the
herbicide-resistance screening studies discussed above (e.g., Reddy
et al., 2014; Nugent et al., 2018; Singh et al., 2026), these applications
show that AI-enabled sensing extends weed science from detection
and mapping to herbicide-response phenotyping, resistance manage-
ment, and mechanism-of-action research.
3.8 Integration of sensor fusion with
machine/deep learning
The development of comprehensive sensing and analytical
frameworks that integrate multiple sensors with machine-learning
and deep-learning algorithms is emerging as an important strategy
for improving weed detection under variable field conditions.
Instead of relying on a single data source, these multimodal
approaches combine complementary data from different sensors,
such as RGB, multispectral, hyperspectral, thermal, LiDAR, or
depth sensors. This allows weed detection systems to capture
visual, spectral, structural, and physiological differences between
crops and weeds more effectively than single-sensor approaches.
RGB imagery provides information related to color, shape, texture,
and canopy architecture, while multispectral and hyperspectral
sensors capture vegetation re flectance associated with chlorophyll
content, water status, and biochemical composition. Combining
these data sources improves weed identi fication, particularly in
environments where illumination, soil background, crop growth
stage, and weed morphology vary substantially ( Xu et al., 2025 ).
Sensor fusion can occur at multiple stages of the analytical pipeline.
Data-level fusion combines raw sensor inputs before training;
feature-level fusion integrates extracted features into a shared
representation, and decision-level fusion combines outputs from
separate models to produce a final classi fication. Feature-level and
decision-level fusion are commonly preferred because they balance
computational efficiency, flexibility, and predictive performance.
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org07

Castellano et al. (2023) proposed integrating semantic segmen-
tation, multispectral imagery, machine learning, and a lightweight
vision transformer for weed mapping in in-crop conditions. Eide
et al. (2021) investigated the use of thermal infrared and multi-
spectral imaging for discriminating glyphosate-resistant redroot
pigweed ( Amaranthus retro flexus L.) populations through a
random tree classi fier and achieved 87% accuracy. Zhang et al.
(2021) integrated UAV-based hyperspectral imagery with a convo-
lutional neural network to classify several monocot and dicot weed
species in a sod farm with 57% to 90% accuracy. Xia et al. (2023)
sourced imagery from two sen sors: a hyperspectral camera
mounted on a six-rotor UAV and an RGB camera with a DJI
Phantom 4 RTK and used a deep fusion network model to achieve
an enhanced accuracy of 78% in identifying atrazine-resistant and
atrazine-susceptible barnyard grass [ Echinochloa crus-galli (L.)
Beauv] and A. theophrasti populations. A novel hybrid architecture
integrating vision transformers, convolutional neural networks, and
multimodal imaging (RGB + near-infrared + hyperspectral) re-
cently achieved 95% accuracy in real-time weed detection ( Pittala
et al., 2026 ).
Recent developments in transformer-based architectures and
attention mechanisms further enhance multimodal learning by
capturing both local plant features and broader field context
(Jácome Galarza et al., 2025 ). These models are particularly useful
in complex agricultural fields where crops and weeds overlap or
exhibit similar visual characteristics. However, multimodal systems
also increase sensor costs, calibration requirements, processing
complexity, and data management demands. Challenges such as
sensor alignment, temporal synchronization, radiometric correc-
tion, and spatial resolution differences must be addressed before
reliable fusion can be achieved. Future research should therefore
focus not only on improving detection accuracy but also on
developing cost-effective and computationally ef ficient fusion
frameworks suitable for real-world deployment.
4 Precision weed management
4.1 Targeted herbicide application and
smart spraying
Deep learning-enabled weed de tection has accelerated the
development of targeted herbicide application systems aimed at
reducing the environmental impacts of broadcast spraying.
Segmentation models such as U2-Net have demonstrated the ability
to separate weeds from complex field backgrounds, enabling highly
selective spraying systems that may reduce herbicide use by up to
90% ( Rizvi et al., 2024 ). Real-time object detection frameworks,
particularly the YOLO ( Redmon and Farhadi, 2018 ) and Faster R-
CNN families, are widely integ rated into smart sprayers and
autonomous robotic platforms in field evaluation systems have
achieved meaningful reductions in herbicide use, with primary
smart-sprayer studies reporting postemergence savings that com-
monly range from 30% to more than 70% depending on weed
pressure, crop and platform ( Avent et al., 2024; Houser et al., 2024).
However, field deployment remains challenging because models
that perform well under controlled conditions often experience
substantial declines in accuracy in real-world settings. Upadhyay
et al. (2024) designed and developed a machine vision-based system
for weed identi fication and precise spray applications using a
YOLOv4 model and reported 100% and 95% precision in indoor
and outdoor experiments, respectively. Deng et al. (2025) used a
YOLOv11 model for real-time weed detection, enabling precise
nozzle activation for targeted spraying, achieving 99% mean average
precision(mAP) in indoor conditions but a reduction to 69% mAP
during field trials.
Greeneye Technology and John Deere have developed precision
sprayers equipped with cameras and processing units to detect
weeds and apply herbicides in real time ( Cross Implement, 2024 ;
NuWay-K&H Cooperative, 2023 ). Avent et al. (2024) reported
average POST herbicide savings of 28% to 62% using See & Spray
technology in dicamba- , glufosinate-, and glyphosate-resistant
soybean. Targeted applications improved pro fitability by reducing
herbicide use and increasing application ef ficiency, providing aver-
age savings of US$43.22 ha -1 to $129.19 ha -1 relative to a broadcast
postemergence cost of $227.22 ha -1 (Avent et al., 2025 ). A study
conducted in Iowa reported an average herbicide saving of 76% and
an economic savings of $39 ha -1 in soybean using John Deere See
and Spray™ Ultimate sprayer ( Houser et al., 2024 ).
Precision herbicide spray systems can support multiple best
weed resistance management practices by enabling real-time weed
scouting, species identi fication, density assessment, and
georeferenced mapping of infestations ( Norsworthy et al., 2012 ).
These data can be integrated with historical weed emergence
patterns and soil characteristics to generate prescription maps for
variable-rate residual herbicide applications, optimizing herbicide
use and reducing selection pressure for resistance to postemergence
herbicides ( Zhu et al., 2026 ). Early-season weed control through
residual herbicides can further reduce weed populations at post-
emergence application timings, maximizing the value of targeted
spray systems. Additionally, dual-tank spray systems enable the
simultaneous application of herbicides with different modes of
action, potentially lowering herbicide application costs, minimizing
antagonistic interactions, and allowing spray parameters to be
optimized for each herbicide ( Leise et al., 2025 ). Adoption of
precision sprayer technologies is primarily driven by farm size,
profitability, and grower characteristics, with larger farms benefiting
from lower per-acre costs and greater returns on investment. Wider
adoption can be facilitated through retro fit solutions, technical
support, and regulatory incentives that improve affordability, flex-
ibility, and environmental outcomes ( Ghatreshsamani et al., 2020 ;
Tey and Brindal, 2012 ).
4.2 Targeted tillage for weed control
Tillage, including emerging site-speci fic approaches, has been
practiced for millennia as a primary form of weed control. Hand
hoeing and manual weed removal represent the earliest methods
and were widely used long before Jethro Tull popularized the horse-
drawn hoe in the early eighteenth century ( Lal, 2009 ). Recent
advances have integrated automation, robotics, and computer
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org08

vision into mechanical weeding, giving rise to intelligent mechan-
ical weeders. These systems employ sensors and arti ficial intelli-
gence algorithms for real-time weed detection and site-speci fic
control, minimizing crop disturbance and operator dependency.
Walsh et al. (2020) evaluated a pre-commercial prototype, the
“Weed Chipper, ” which integrated commercially available weed-
detection cameras for field-scale operation. The system achieved
90% weed control ef ficacy with only 1.8% soil disturbance under
low weed density conditions (<1.0 plant per 10 m 2) in a commercial
fallow field. Quan et al. (2022) developed an intra-row robotic
mechanical weeding system based on deep learning for crop and
weed detection, reporting an 86% weed removal rate with 1% corn
(Zea mays L.) injury. Das et al. (2025) recently summarized the
intelligent mechanical weed management systems. Given the recent
advances in machine vision, arti ficial intelligence, and robotics,
targeted mechanical weeders are likely to play an increasingly
important role in integrated weed management programs.
4.3 Targeted laser weeding
Laser technology is an emerging tool for precision weed
management because coherent laser beams, in which light waves
are spatially and temporally sync hronized, can deliver highly
concentrated energy to targeted plants ( Coleman et al., 2021 ).
This capability enables precise, nonchemical weed control with
minimal disturbance to the surrounding crop and soil. In early
2022, Carbon Robotics introduced a commercial laser weeder
designed for large-scale vegetable production and are currently
operating primarily in the western United States ( Sosnoskie et al.,
2025). Field experiments conducted by Sosnoskie et al. (2025) in
beet ( Beta vulgaris L.), spinach ( Spinacia oleracea L.), and pea
(Pisum sativum L.) demonstrated that laser weeding reduced weed
cover by ≥45% and weed density by ≥66%, resulting in ≥97% lower
weed biomass at the end of the growing season compared with
untreated controls. Mwitta et al. (2024) developed an autonomous
robotic platform that integrated deep learning and machine vision
algorithms for weed detection and laser diodes for weed elimination
in cotton. The system achieved 72% weed control, highlighting the
potential of autonomous laser-based weed management using small
robotic platforms. Overall, while laser weeding shows strong po-
tential as a precise, nonchemical weed management strategy, its
widespread adoption remains constrained by high capital costs,
with individual units estimated to cost at least US$500,000
(Sosnoskie et al., 2025 ).
5 Model generalization, validation, and
real-time field deployment
Although deep learning models frequently achieve high accu-
racy under controlled conditions, their performance often declines
substantially when deployed in operational fields. This gap arises
because field environments introduce variability that is poorly
represented in many training datasets, including changing
illumination and weather, soil background and crop residue,
overlapping canopies, differences in crop and weed growth stage,
species diversity, field history, and sensor or calibration differences
across platforms. Models trained under limited conditions often
overfit to dataset-speci fic cues and perform poorly when applied to
new fields, seasons, or sensors. A related and frequently underap-
preciated issue is validation design. When images from the same
field, date, or plot are shared between training and testing sets,
spatial and temporal autocorrelation can in flate reported accuracy
and give an overly optimistic impression of real-world performance.
Robust evaluation therefore requires independent partitioning,
ideally cross- field and cross-year (or cross-site and cross-season)
validation, so that reported metrics re flect genuine generalization
rather than memorization of local conditions. Strategies such as
diverse and well-documented data collection, data augmentation,
domain adaptation, transfer learning, semi-supervised and self-
supervised learning, and uncertainty-aware models can improve
robustness. However, transparent reporting of dataset composition
and validation protocols remains essential for interpreting pub-
lished accuracies Kaushal et al. (2026) .
Translating an accurate detection model into an operational
weed-control system introduces engineering constraints that extend
well beyond classi fication accuracy. For smart sprayers and robotic
platforms, effective performance depends on the entire sensing-to-
actuation chain, including on-board (edge) computing capacity,
inference latency, camera field of view and camera-to-nozzle (or
camera-to-tool) offset, platform forward speed, nozzle or actuator
response time, actuator precision, geolocation accuracy, and precise
temporal synchronization betw een detection and actuation. A
model that is accurate but too slow, or a system with mismatched
timing between detection and nozzle firing, can miss target weeds or
spray crop plants even when the underlying classi fier performs well.
Consequently, real-world evaluation should report not only detec-
tion metrics but also system-level outcomes such as targeting
accuracy, hit rate, spray or actuation precision, throughput, and
crop-injury risk at realistic operating speeds. This is consistent with
the marked accuracy reductions observed when models validated
indoors are deployed in the field (e.g., the decline from 99% to 69%
mAP reported by Deng et al., 2025 ), underscoring that detection
and operational performances must be assessed together. Table 3
compares the major precision we ed-control technologies with
respect to efficacy, input savings, crop-injury risk, cost, and deploy-
ment barriers.
6 Emerging AI directions and
deployment considerations
Positioning these tools as advanced also involves engaging with
the leading current trends in agricultural AI. Beyond conventional
CNNs, YOLO detectors, and U-Net-type segmentation, recent work
increasingly explores domain adaptation and domain generalization
to bridge the controlled-to- field gap, while self-supervised, semi-
supervised, and active learning reduce dependence on large
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org09

annotated datasets ( Rai and Sun, 2024 ; Shen et al., 2025 ). Recent
work also explores foundation vision-language models, multimodal
learning, synthetic and simulated training data, and uncertainty-
aware prediction to improve transferability and decision reliability
so that models can be adapted to new crops and weeds with limited
labels (Awais et al., 2025 ; Nasir et al., 2026 ). For field deployment,
model compression and quantization are important to reduce
latency, computational, and energy requirements on edge devices
(Wang et al., 2025 ). Therefore, AI systems should not be evaluated
by accuracy alone. Robustness to distribution shift, inference speed,
resource requirements, and the agronomic costs of false positives
and false negatives are equally important for practical
weed management.
7 From herbicide reduction to holistic
sustainability
Reduced herbicide use is an important bene fit, but it is not
sufficient on its own to establish overall sustainability. A rigorous
assessment should distinguish four related but distinct dimensions:
technological feasibility (whether a system reliably detects and acts on
weeds under field conditions), agronomic efficacy (whether it delivers
acceptable weed control and protects yield), economic feasibility
(capital cost, energy consumption, machinery throughput, mainte-
nance, sensor replacement, computing requirements, and accessibil-
ity to small and medium-size d farms), and environmental
sustainability (including labor implications and, ideally, life-cycle
environmental impacts). The high capital cost of laser weeding
illustrates the importance of the economic dimension, but compara-
ble considerations apply across smart spraying, robotic mechanical
weeding, and other sensor-intensive platforms. Table 3 summarizes
these technologies with respect to efficacy, input savings, crop-injury
risk, cost, and deployment barriers; we caution against treating
technological feasibility, agronomic ef ficacy, economic feasibility,
and environmental sustainability as interchangeable.
8 Precision weed management from
an agroecological perspective
Weeds of agroecosystems, selected by anthropogenic distur-
bance, are often seen and perhaps marketed as ‘unwanted’ plants, a
perspective increasingly criticized as ecologically and evolutionarily
unsustainable ( Boinot et al., 2024 ). This reductionist construct has
led to weed management programs centered on complete eradica-
tion, thereby compromising the resilience of agroecosystems.
Current weed management practices (i.e., herbicides, tillage) act
as strong filters that have selected for fewer, highly adaptive and
competitive ‘super-weed’ species, significantly reducing weed diver-
sity and negatively impacting the multifunctionality of
agroecosystems and the entire food chain ( Boinot et al., 2024 ).
These challenges warrant a paradigm shift from complete eradica-
tion (i.e., zero tolerance) to the preservation of ecologically neutral
or even facilitative weed communities through agroecological weed
management (AWM). The central tenet of AWM is to integrate
fundamental ecological and biological principles to manipulate
landscape-level abiotic and biotic filters in ways that suppress
highly competitive weed species while preserving the ecosystem
services associated with functionally diverse weed communities.
Recent advances in precision-based weed detection, particularly
the integration of deep learning-based neural networks (e.g., CNNs,
RNNs, ViTs) with high-resolution sensors, could enable a targeted
approach to managing only problematic weed species, thereby
establishing neutral weed communities and potentially preserving
ecosystem services (Bagavathiannan and von Redwitz, 2025). In this
instance, fundamental ecological and evolutionary concepts, such as
trait- and phylogenetic-mediated crop-weed interactions, may be
integrated with advanced remote sensing and analytical techniques.
While some functional traits can be partially inferred from remote
sensing data, phylogenetic information cannot be directly sensed.
Instead, it must be obtained through species identi fication along
with external ecological and phylogenetic databases. It is therefore
crucial to differentiate between direct measurement and inference
when employing these metrics t o prioritize management of
TABLE 3 Comparison of precision weed-control technologies based on studies and commercial systems discussed in this review.
Technology Representative
system/study
Crop
system
Reported ef ficacy Crop-
injury
risk
Relative
cost
Key deployment
barriers
Targeted
spraying/smart
sprayers
See & Spray ( Avent et al.,
2024, 2025; Houser et al.,
2024)
Soybean and
row crops

90% weed control Low-
moderate
High capital
(scales with
farm size)
In-field detection
accuracy, timing/
synchronization, cost.
Robotic
mechanical
weeding
Weed Chipper ( Walsh et al.,
2020); intra-row robot ( Quan
et al., 2022 )
Fallow; row
crops (corn)
90% control (1.8% soil
disturbance); 86% removal (1%
crop injury)
Moderate
(mechanical)
Moderate-
high
Speed/throughput, intra-
row precision, soil and
residue variability.
Laser weeding Carbon Robotics ( Sosnoskie
et al., 2025 ); diode-laser robot
(Mwitta et al., 2024 )
Vegetables
(beet, spinach,
pea); cotton
≥45% cover and ≥66% density
reduction; ≥97% lower biomass;
72% control (robotic)
Low
(selective)
Very high
(≥US
$500,000 per
unit)
Capital cost, throughput,
energy, safety.
Performance values (e.g., ef ficacy, input savings, and injury risk) re flect findings from specific cited trials under distinct crops, weed pressures, and field conditions, and are therefore not directly
comparable across technologies.
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org10

problematic species. Although some research efforts have demon-
strated the potential of remote sensing-based measurement of plant
functional traits across various ecosystems ( Abelleira Mart ı́nez
et al., 2016 ; Cherif et al., 2023 ; Hayes et al., 2025 ; Ustin and
Gamon, 2010 ), substantial work remains to characterize crop-
weed functional traits. This approach can also be scaled up to
develop a systematic weed trait database, which is currently lacking,
particularly for North American agroecosystems.
Achieving this vision involves tra nsforming ecological concepts
into a clear, actionable decision f ramework. This framework would
connect (i) detection and species identi fication, (ii) estimating weed
density and distribution, (iii) crop growth stages and critical crop-weed
competition periods, (iv) weed traits and phylogenetic links, (v) the risk
posed by weeds and their ecosystem bene fits, and (vi) management
actions, from no action for low-risk or helpful species to targeted
control for problematic weeds. Advances in real-time detection,
imagery-based trait retrieval, and decision-support models are
making each component more feasible, though integrating them into
validated,field-ready systems remains a significant research challenge.
Tolerance-based or selective weed management also carries im-
portant agronomic and regulatory risks that must be weighed explic-
itly. Allowing some weeds to survive and reproduce can replenish the
soil seedbank. Tolerating regulated noxious or invasive species may
violate legal requirements, and retained weeds can contaminate
harvested produce, host insect pests or pathogens, interfere with
harvest operations, and increase management complexity. Grower
risk tolerance, economic thresholds, and market or certi fication
requirements will therefore strongly in fluence adoption. A credible
agroecological approach must incorporate these constraints, prioritiz-
ing the suppression of high-risk s pecies and safeguarding against
seedbank replenishment, so that conserving facilitative or neutral
weed communities does not compromise regulatory compliance.
9 Conclusion and future opportunities
Artificial intelligence and machine learning are rapidly trans-
forming digital weed management by enabling automated weed
detection, mapping, and precision management. RGB, multispec-
tral, and hyperspectral sensing platforms provide complementary
capabilities, while deep learning architectures increasingly deliver
highly accurate classi fication and segment ation performance.
However, signi ficant barriers remain, including limited annotated
datasets, poor model generalization across environments, compu-
tational constraints, and challenges in real-time deployment. Future
research should focus on developing robust, high-accuracy detec-
tion methods for various weed species and functional traits (phe-
nological, physiological, and morphological) of crops and weeds,
which are scalable across the range of agroecosystems. Open-access
global weed datasets, multimodal sensor fusion, domain adaptation,
self- and semi-supervised learning, foundation and vision-language
models, synthetic training data, model compression for edge
computing, uncertainty-aware AI, and AI-integrated decision sup-
port systems are likely to play central roles in advancing sustainable
site-speci fic weed management. Lastly, we emphasize that the
holistic use of precision-bas ed weed management technology,
grounded in fundamental ecological and evolutionary concepts,
has strong potential to reduce reliance on standalone management
approaches (e.g., herbicides, tillage), thereby enhancing the sustain-
ability of agroecosystems.
Author contributions
AC: Conceptualization, Writing– original draft, Writing– review
& editing. SK: Conceptualization, Writing – original draft, Writing –
review & editing. HD: Conceptualization, Writing – original draft,
Writing – review & editing. NS: Writing – original draft, Writing –
review & editing. MJ: Writing – original draft, Writing – review &
editing. DD: Writing – original draft, Writing – review & editing. JK:
Writing – original draft, Writing – review & editing. KM: Writing –
original draft, Writing – review & editing. MV: Writing – review &
editing, Writing – original draft. SD: Conceptualization, Writing –
original draft, Writing – review & editing.
Funding
The author(s) declared that financial support was not received
for this work and/or its publication.
Conflict of interest
The author(s) declared that this work was conducted in the
absence of any commercial or financial relationships that could be
construed as a potential con flict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the
creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this
article has been generated by Frontiers with the support of arti ficial
intelligence and reasonable e fforts have been made to ensure
accuracy, including review by the authors wherever possible. If
you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their affiliated organizations,
or those of the publisher, the editors and the reviewers. Any product
that may be evaluated in this article, or claim that may be made by its
manufacturer, is not guaranteed or endorsed by the publisher.
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org11

References
Abelleira Mart ı́nez, O. J., Fremier, A. K., Günter, S., Ramos Bendaña, Z., Vierling, L.,
Galbraith, S. M., et al. (2016). Scaling up functional traits for ecosystem services with
remote sensing: concepts and methods. Ecol. Evol. 6, 4359 –4371. doi: 10.1002/
ece3.2201
Abrantes, T. C., Queiroz, A. R. S., Lucio, F. R., Mendes Junior, C. W., Kuplich, T. M.,
Bredemeier, C., et al. (2021). Asses sing the effects of dicamba and 2, 4
dichlorophenoxyacetic acid (2, 4D) on soybean through vegetation indices derived
from unmanned aerial vehicle (UAV) based RGB imagery. Int. J. Remote Sens. 42,
2740–2758. doi: 10.1080/01431161.2020.1832283
Avent, T. H., Norsworthy, J. K., Patzoldt, W. L., Schwartz-Lazaro, L. M., Houston, M.
M., Butts, T. R., et al. (2024). Comparing herbicide application methods with See &
Spray™ technology in soybean. Weed Technol. 38, e74. doi: 10.1017/wet.2024.70
Avent, T. H., Norsworthy, J. K., Zimmer, M., Young, B. G., Contreras, D. J., Everman,
W. J., et al. (2025). Commercial sprayer ef ficiency for application success on targeted
weeds. Weed Technol. 39, e65. doi: 10.1017/wet.2025.36
Awais, M., Alharthi, A. H. S. A., Kumar, A., Cholakkal, H., and Anwer, R. M. (2025).
“Agrogpt: Ef ficient agricultural vision-language model with expert tuning ”, in: 2025
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (IEEE), pp.
5687–5696.
Baatz, M. (2000). “Multiresolution segmentation: an optimization approach for high
quality multi-scale image segmentation ”,i n : Angewandte Geographische
Informationsverarbeitung XII. Karlsruhe, Germany , 12, 12 –23. Available online at:
https://cir.nii.ac.jp/crid/1572261550679971840.
Bagavathiannan, M. V., and von Redwitz, C. (2025). Precision agriculture can foster
cropland weed diversity and protect endangered species. Weed Sci. 73, e48.
doi: 10.1017/wsc.2025.10024
Bamal, D., Duhan, A., Pal, A., Beniwal, R. K., Kumawat, P., Dhanda, S., et al. (2024).
Herbicide risks to non-target species and the environment: A review. Environ. Chem.
Lett. 22, 2977 –3032. doi: 10.1007/s10311-024-01773-9
Benos, L., Tagarakis, A. C., Dolias, G., Berruto, R., Kateris, D., and Bochtis, D. (2021).
Machine learning in agriculture: A comprehensive updated review. Sensors 21, 3758.
doi: 10.3390/s21113758
Betitame, K., Koparan, C., Zhang, Y., Howatt, K., Ostlie, M., Bajwa, S. G., et al. (2024).
Evaluation of dicamba drift injury and yield loss on soybean using small unmanned
aircraft systems (sUAS) and multispectral imaging technologies. J. Natural Resour.
Agric. Ecosyst. 2, 63–76. doi: 10.13031/jnrae.15686
Boinot, S., Alignier, A., and Storkey, J . (2024). Landscape perspectives for
agroecological weed management. A review. Agron. Sustain. Dev. 44, 7. doi: 10.1007/
s13593-023-00941-5
Brainard, D. C., Haramoto, E. R., Leon, R. G., Kells, J. J., Van Wychen, L. R., Devkota,
P., et al. (2023). A survey of weed research priorities: key findings and future directions.
Weed Sci. 71, 330 –343. doi: 10.1017/wsc.2023.24
Breiman, L. (2001). Random forests. Mach. Learn. 45, 5 –32. doi: 10.1023/
A:1010933404324
Castellano, G., De Marinis, P., and Vessio, G. (2023). Weed mapping in multispectral
drone imagery using lightweight vision transformers. Neurocomputing 562, 126914.
doi: 10.1016/j.neucom.2023.126914
Cherif, E., Feilhauer, H., Berger, K., Dao, P. D., Ewald, M., Hank, T. B., et al. (2023). From
spectra to plant functional traits: Transferable multi-trait models from heterogeneous and
sparse data. Remote Sens. Environ. 292, 113580. doi: 10.1016/j.rse.2023.113580
Chu, H., Zhang, C., Wang, M., Gouda, M., Wei, X., He, Y., et al. (2022). Hyperspectral
imaging with shallow convolutional neural networks (SCNN) predicts the early
herbicide stress in wheat cultivars. J. Hazard. Mater. 421, 126706. doi: 10.1016/
j.jhazmat.2021.126706
Coleman, G., Betters, C., Squires, C., Leon-Saval, S., and Walsh, M. (2021). Low energy
laser treatments control annual ryegrass (Lolium rigidum). Front. Agron. 2, 601542.
doi: 10.3389/fagro.2020.601542
Cross Implement (2024). John Deere See & Spray Sprayer Demos Spring 24 . Available
online at: https://crossimplement.com/news/article/2024/02/johndeere-see-and-spray-
sprayer-demos-spring-24 (Accessed May 10, 2026).
Das, S., Upadhyay, A., and Sun, X. (2025). Technology-driven approaches to intelligent
mechanical weed control: a systematic review for sustainable weed management. Front.
Plant Sci. 16, 1734507. doi: 10.3389/fpls.2025.1734507
de Castro, A. I., Jurado-Expo ́sito, M., Peña-Barraga ́n, J. M., and Lo ́pez-Granados, F.
(2012). Airborne multi-spectral imagery for mapping cruciferous weeds in cereal and
legume crops. Precis. Agric. 13, 302 –321. doi: 10.1007/s11119-011-9247-0
Deng, B., Lu, Y., and Brainard, D. (2025). Semi-supervised weed detection in vegetable
fields: in-domain and cross-domain experiments. IFAC-PapersOnLine 59, 96 –101.
doi: 10.1016/j.ifacol.2025.11.770
Desai, H. S., Menalled, F. D., and Shergill, L. S. (2026). Functional trait-mediated
limiting similarity shapes weed biomass acquisition, fecundity, and reproductive
phenology. Sci. Rep. 16, 24514. doi: 10.1038/s41598-026-63511-1
Eide, A., Koparan, C., Zhang, Y., Ostlie, M., Howatt, K., and Sun, X. (2021). UAV-
assisted thermal infrared and multispectral imaging of weed canopies for glyphosate
resistance detection. Remote Sens. 13, 4606. doi: 10.3390/rs13224606
El Sakka, M., De Pourtales, C., Chaari, L., and Mothe, J. (2025). “AgriPotential: A novel
multi-spectral and multi-temporal remote sensing dataset for agricultural potentials ”,
in: 2025 International Conference on Content-Based Multimedia Indexing (CBMI)
(IEEE), 1–6. doi: 10.1109/CBMI66578.2025.11339350
Esposito, M., Crimaldi, M., Cirillo, V., Sarghini, F., and Maggio, A. (2021). Drone and
sensor technology for sustainable weed management: A review. Chem. Biol. Technol.
Agric. 8, 18. doi: 10.1186/S40538-021-00217-8
Esposito, M., Westbrook, A. S., Maggio, A., Cirillo, V., and DiTommaso, A. (2023).
Neutral weed communities: the intersection between crop productivity, biodiversity,
and weed ecosystem services. Weed Sci. 71, 301 –311. doi: 10.1017/wsc.2023.27
Falcon, W. P., Naylor, R. L., and Shankar, N. D. (2022). Rethinking global food demand
for 2050. Population Dev. Rev. 48, 921 –957. doi: 10.1111/padr.12508
FAO (2025). The State of Food Security and Nutrition in the World 2025: Addressing
High Food Price In flation for Food Security and Nutrition (Rome, Italy: Food and
Agriculture Organization of the United Nations).
Gao, J., Liao, W., Nuyttens, D., Lootens, P., Vangeyte, J., Piz ̌urica, A., et al. (2018).
Fusion of pixel and object-based features for weed mapping using unmanned aerial
vehicle imagery. Int. J. Appl. Earth Obs. Geoinf. 67, 43 –53. doi: 10.1016/
j.jag.2017.12.012
Gharde, Y., Singh, P. K., Dubey, R. P., and Gupta, P. K. (2018). Assessment of yield and
economic losses in agriculture due to weeds in India. Crop Prot. 107, 12 –18.
doi: 10.1016/j.cropro.2018.01.007
Ghatrehsamani, S., Wade, T., and Ampatzidis, Y. (2020). The adoption of precision
agriculture technologies by Florida growers: a comparison of 2005 and 2018 survey
data. Acta Hortic. 1279, 311 –316. doi: 10.17660/ActaHortic.2020.1279.44
Gibson, K. D., Dirks, R., Medlin, C. R., and Johnston, L. (2004). Detection of weed
species in soybean using multispectral digital images. Weed Technol. 18, 742 –749.
doi: 10.1614/WT-03-170R1
Goel, P. K., Prasher, S. O., Patel, R. M., Smith, D. L., and DiTommaso, A. (2002). Use of
airborne multi–spectral imagery for weed detection in field crops. Trans. ASAE 45, 443.
doi: 10.13031/2013.8513
Hamad, A., and Tayel, A. (2026). Food 2050 concept: the trends that shape the future of
our food. J. Future Foods 6, 1053 –1066. doi: 10.1016/j.jfutfo.2025.03.003
Hassannejad, S., Lotfi, R., Ghafarbi, S. P., Oukarroum, A., Abbasi, A., Kalaji, H. M., et al.
(2020). Early identi fication of herbicide modes of action by the use of chlorophyll
fluorescence measurements. Plants 9, 529. doi: 10.3390/plants9040529
Hayes, S., Cawkwell, F., Bacon, K. L., and Wingler, A. (2025). Remote sensing of
grassland plant biodiversity and functional traits. Ecol. Evol. 15, e71829. doi: 10.1002/
ece3.71829
Heap, I. (2026). The International Herbicide-Resistant Weed Database. Available online
at: http://www.weedscience.org (Accessed May 31, 2026).
Henry, W. B., Shaw, D. R., Reddy, K. R., Bruce, L. M., and Tamhankar, H. D. (2004).
Remote sensing to distinguish soybean from weeds after herbicide application. Weed
Technol. 18, 594 –604. doi: 10.1614/WT-03-097R
Houser, D., Darr, M., and Huffman, R. (2024). Precision Spraying Technology. Iowa
State University Extension and Outreach Integrated Crop Management. Available online
at: https://crops.extension.iastate.edu/cropnews/2024/08/precision-spraying-technol-
ogy (Accessed May 31, 2026).
Huang, Y., Lee, M. A., Nandula, V. K., and Reddy, K. N. (2018b). Hyperspectral
imaging for differentiating glyphosate-resistant and glyphosate-susceptible Italian
ryegrass. Am. J. Plant Sci. 9, 1467 –1477. doi: 10.4236/ajps.2018.97107
Huang, Y., Reddy, K. N., Fletcher, R. S., and Pennington, D. (2018a). UAV low-altitude
remote sensing for precision weed management. Weed Technol. 32, 2–6. doi: 10.1017/
wet.2017.89
Huang, Y., Sun, X., Rai, N., Yao, H., Reddy, K. N., and Jenkins, J. (2025). Remote
sensing for precision weed management. Pest. Manage. Sci. 81, 4190 –4199.
doi: 10.1002/ps.8833
Huang, Y., Zhao, X., Pan, Z., Reddy, K. N., and Zhang, J. (2022). Hyperspectral plant
sensing for differentiating glyphosate ‐resistant and glyphosate ‐susceptible johnson-
grass through machine learning algorithms. Pest. Manage. Sci. 78, 2370 –2377.
doi: 10.1002/ps.6864
İeri, O. (2025). Usability of smartphone-based RGB vegetation indices for steppe
rangeland inventory and monitoring. Precis. Agric. 26, 2. doi: 10.1007/s11119-024-
10195-0
Jácome Galarza, L., Realpe, M., Viña ́n-Ludeña, M. S., Calderón, M. F., and Jaramillo, S.
(2025). Agritransformer: A transformer-based model with attention mechanisms for
enhanced multimodal crop yield prediction. Electronics 14, 2466. doi: 10.3390/elec-
tronics14122466
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org12

Kaushal, S., Billah, M. M., Singh, M., Koupal, D., Sitaula, P. R., Halder, J., et al. (2026).
UAV‐based deep transfer learning to improve grain yield prediction in winter wheat
across temporal and spatial variability. Plant Phenome J. 9, e70085. doi: 10.1002/
ppj2.70085
Kubiak, A., Wolna-Maruwka, A., Niewiadomska, A., and Pilarska, A. A. (2022). The
problem of weed infestation of agricultural plantations vs. the assumptions of the
European biodiversity strategy. Agronomy 12, 1808. doi: 10.3390/agronomy12081808
Lal, R. (2009). The plow and agricultural sustainability. J. Sustain. Agric. 33, 66 –84.
doi: 10.1080/10440040802548555
Leise, A., Singh, M., Cafaro La Menza, N., Knezevic, S. Z., and Jhala, A. J. (2025). A dual
tank precision sprayer to evaluate the interaction of 2,4-D choline and quizalofop-p-
ethyl for control of vol unteer corn in corn. Agric. Environ. Lett. 10, e70015.
doi: 10.1002/ael2.70015
Lu, B., Dao, P. D., Liu, J., He, Y., and Shang, J. (2020). Recent advances of hyperspectral
imaging technology and applications in agriculture. Remote Sens. 12, 2659.
doi: 10.3390/rs12162659
Mensah, B., Betitame, K., Zhang, Y., Peters, T., Ram, B. G., Jalal, M., et al. (2025).
Hyperspectral imaging and machine learning for herbicide-resistant kochia identi fica-
tion in sugarbeet. J. Agric. Food Res. 23, 102284. doi: 10.1016/j.jafr.2025.102284
Modi, R. U., Kancheti, M., Subeesh, A., Raj, C., Singh, A. K., Chandel, N. S., et al.
(2023). An automated weed identi fication framework for sugarcane crop: a deep
learning approach. Crop Prot. 173, 106360. doi: 10.1016/j.cropro.2023.106360
Mogili, U. R., and Deepak, B. B. V. L. (2018). Review on application of drone systems in
precision agriculture. Proc. Comput. Sci. 133, 502–509. doi: 10.1016/j.procs.2018.07.063
Mwitta, C., Rains, G. C., and Prostko, E. P. (2024). Autonomous diode laser weeding
mobile robot in cotton field using deep learning, visual servoing and finite state
machine. Front. Agron. 6, 1388452. doi: 10.3389/fagro.2024.1388452
Nasir, M. F., Rehman, M. U., and Hussain, I. (2026). Vision-language models for zero-
shot weed detection and visual reasoning in UAV-based precision agriculture. Front.
Plant Sci. 16, 1735096. doi: 10.3389/fpls.2025.1735096
Niu, Z., Rehman, T., Young, J., Johnson, W. G., Yokoo, T., Young, B., et al. (2023).
Hyperspectral analysis for discriminating herbicide site of action: a novel approach for
accelerating herbicide research. Sensors 23, 9300. doi: 10.3390/s23239300
Norsworthy, J. K., Ward, S. M., Shaw, D. R., Llewellyn, R. S., Nichols, R. L., Webster, T.
M., et al. (2012). Reducing the risks of herbicide resistance: best management practices
and recommendations. Weed Sci. 60, 31–62. doi: 10.1614/WS-D-11-00155.1
Nugent, P. W., Shaw, J. A., Jha, P., Scherrer, B., Donelick, A., and Kumar, V. (2018).
Discrimination of herbicide-resistant kochia with hyperspectral imaging. J. Appl.
Remote Sens. 12, 016037 –016037. doi: 10.1117/1.JRS.12.016037
NuWay-K&H Cooperative (2023).Greeneye Faqs. Available online at: https://nuway-kandh.
com/wp-content/uploads/2023/02/Greeneye-FAQs-Web.pdf (Accessed May 31, 2026).
Papadopoulos, A. V., Kati, V., Chachalis, D., Kotoulas, V., and Stamatiadis, S. (2018).
Weed mapping in cotton using ground-based sensors and GIS. Environ. Monit. Assess.
190, 622. doi: 10.1007/s10661-018-6991-x
Peña, J. M., Torres-Sa ́nchez, J., Serrano-Pe ́rez, A., De Castro, A. I., and Lo ́pez-
Granados, F. (2015). Quantifying ef ficacy and limits of unmanned aerial vehicle
(UAV) technology for weed seedling detection as affected by sensor resolution.
Sensors 15, 5609 –5626. doi: 10.3390/s150305609
Pérez-Ortiz, M., Peña, J. M., Gutie ́rrez, P. A., Torres-Sa ́nchez, J., Herva ́s-Martı́nez, C.,
and Lo ́pez-Granados, F. (2015). A semi-supervised system for weed mapping in
sunflower crops using unmanned aerial vehicles and a crop row detection method.
Appl. Soft Comput. 37, 533 –544. doi: 10.1016/j.asoc.2015.08.027
Pittala, R. B., Ramakrishna, B., Ravi, C. V., Kiran, M. A., Sharma, N., Thaile, M., et al.
(2026). Integrating vision transformers and CNNs for high-accuracy weed detection in
crop fields. Eng. Lett. 34 (1), 50.
Pu, R., Landry, S., and Yu, Q. (2011). Object-based urban detailed land cover
classification with high spatial resolution IKONOS imagery. Int. J. Remote Sens. 32,
3285–3308. doi: 10.1080/01431161003745657
Qu, H. R., and Su, W. H. (2024). Deep learning-based weed –crop recognition for smart
agricultural equipment: A review. Agronomy 14, 363. doi: 10.3390/agronomy14020363
Quan, L., Jiang, W., Li, H., Li, H., Wang, Q., and Chen, L. (2022). Intelligent intra-row
robotic weeding system combining deep learning technology with a targeted weeding
mode. Biosyst. Eng. 216, 13 –31. doi: 10.1016/j.biosystemseng.2022.01.019
Rai, N., and Sun, X. (2024). WeedVision: A single-stage deep learning architecture to
perform weed detection and segmentation using drone-acquired images. Comput.
Electron. Agric. 219, 108792. doi: 10.1016/j.compag.2024.108792
Rai, N., Zhang, Y., Ram, B. G., Schumacher, L., Yellavajjala, R. K., Bajwa, S., et al.
(2023). Applications of deep learning in precision weed management: A review.
Comput. Electron. Agric. 206, 107698. doi: 10.1016/j.compag.2023.107698
Reddy, K. N., Huang, Y., Lee, M. A., Nandula, V. K., Fletcher, R. S., Thomson, S. J., et al.
(2014). Glyphosate ‐resistant and glyphosate ‐susceptible Palmer amaranth
(Amaranthus palmeri S. Wats.): hyperspectral re flectance properties of plants and
potential for classi fication. Pest. Manage. Sci. 70, 1910 –1917. doi: 10.1002/ps.3755
Redmon, J., and Farhadi, A. (2018). Yolov3: An incremental improvement. arXiv.
doi: 10.48550/arXiv.1804.02767
Rizvi, S. M. H., Naseer, A., Rehman, S. U., Akram, S., and Gruhn, V. (2024).
Revolutionizing agriculture: Machine and deep learning solutions for enhanced crop
quality and weed control. IEEE Access 12, 11865 –11878. doi: 10.1109/
ACCESS.2024.3355017
Roslim, M. H. M., Juraimi, A. S., Che ’Ya, N. N., Sulaiman, N., Manaf, M. N. H. A.,
Ramli, Z., et al. (2021). Using remote sensing and an unmanned aerial system for weed
management in agricultural crops: A review. Agronomy 11, 1809. doi: 10.3390/
agronomy11091809
Sa, I., Popović, M., Khanna, R., Chen, Z., Lottes, P., Liebisch, F., et al. (2018). WeedMap:
A large-scale semantic weed mapping framework using aerial multispectral imaging
and deep neural network for precision farming. Remote Sens. 10, 1423. doi: 10.3390/
rs10091423
Sapkota, B., Singh, V., Cope, D., Valasek, J., and Bagavathiannan, M. (2020). Mapping
and estimating weeds in cotton using unmanned aerial systems-borne imagery.
AgriEngineering 2, 24. doi: 10.3390/agriengineering2020024
Shen, Y., Ayanlade, T. T., Boddepalli, V. N., Saadati, M., Rairdin, A., Deng, Z. K., et al.
(2025). WeedNet: A foundation model-based global-to-local AI approach for real-time
weed species identi fication and classi fication. arXiv. doi: 10.48550/arXiv.2505.18930
Singh, K. D., Natarajan, M., Ravichandran, P., Tidemann, B. D., and Geddes, C. M.
(2026). Hyperspectral imaging and machine learning to classify herbicide-resistant
and-susceptible weeds in field crops. Can. J. Plant Sci. 106, 1 –18. doi: 10.1139/cjps-
2025-0244
Soltani, N., Dille, J. A., Burke, I. C., Everman, W. J., VanGessel, M. J., Davis, V. M., et al.
(2017). Perspectives on potential soybean yield losses from weeds in North America.
Weed Technol. 31, 148 –154. doi: 10.1017/wet.2016.2
Sosnoskie, L. M., Bill, K., Butler ‐Jones, A., Bouchelle, W. M., and Besanc ̧on, T. E.
(2025). Deep learning ‐based laser weed control compared to conventional herbicide
application across three vegetable production systems. Pest. Manage. Sci. 81, 4922 –
4936. doi: 10.1002/ps.8912
Stephen, S., and Kumar, V. (2023). Detection and analysis of weed impact on sugar beet
crop using drone imagery. J. Indian Soc Remote Sens. 51, 2577 –2597. doi: 10.1007/
s12524-023-01782-1
Steward, B., Gai, J., and Tang, L. (2019). “The use of agricultural robots in weed
management and control, ” in Robotics and Automation for Improving Agriculture
(Burleigh Dodds Science Publishing), 161 –186. doi: 10.19103/AS.2019.0056.13
Sulaiman, N., Che ’Ya, N. N., Mohd Roslim, M. H., Juraimi, A. S., Mohd Noor, N., and
Fazlil Ilahi, W. F. (2022). The application of hyperspectral remote sensing imagery
(HRSI) for weed detection analysis in rice fields: A review. Appl. Sci. 12, 2570.
doi: 10.3390/app12052570
Sunil, G. C., Koparan, C., Ahmed, M. R., Zhang, Y., Howatt, K., and Sun, X. (2022). A
study on deep learning algorithm performance on weed and crop species identi fication
under different image background. Artif. Intell. Agric. 6, 242 –256. doi: 10.1016/
j.aiia.2022.11.001
Tey, Y. S., and Brindal, M. (2012). Factors in fluencing the adoption of precision
agricultural technologies: a review for policy implications. Precis. Agric. 13, 713 –730.
doi: 10.1007/s11119-012-9273-6
Upadhyay, A., Sunil, G. C., Zhang, Y., Koparan, C., and Sun, X. (2024). Development
and evaluation of a machine vision and deep learning-based smart sprayer system for
site-specific weed management in row crops: An edge computing approach. J. Agric.
Food Res. 18, 101331. doi: 10.1016/j.jafr.2024.101331
Ustin, S. L., and Gamon, J. A. (2010). Remote sensing of plant functional types. New
Phytol. 186, 795 –816. doi: 10.1111/j.1469-8137.2010.03284.x
Velusamy, P., Rajendran, S., and John William, A. D. (2023). “Machine vision in UAV
data analytics for precision agriculture, ” in Drone Data Analytics in Aerial Computing
(Singapore: Springer Nature Singapore), 145 –162. doi: 10.1007/978-981-99-5056-0_8
Vijayakumar, V., Ampatzidis, Y., Lacerda, C., Burks, T., Lee, W. S., and Schueller, J.
(2025). AI-driven real-time weed detection and robotic smart spraying for optimised
performance and operational speed in vegetable production. Biosyst. Eng. 259, 104288.
Walsh, M. J., Squires, C. C., Coleman, G. R., Widderick, M. J., McKiernan, A. B.,
Chauhan, B. S., et al. (2020). Tillage-based, site-speci fic weed control for conservation
cropping systems. Weed Technol. 34, 704 –710. doi: 10.1017/wet.2020.44
Wang, X., Tang, Z., Guo, J., Meng, T., Wang, C., Wang, T., et al. (2025). Empowering
edge intelligence: A comprehensive survey on on-device AI models. ACM Comput.
Surv. 57, 1 –39. doi: 10.1145/3724420
Xia, F., Lou, Z., Sun, D., Li, H., and Quan, L. (2023). Weed resistance assessment
through airborne multimodal data fusion and deep learning: A novel approach towards
sustainable agriculture. Int. J. Appl. Earth Obs. Geoinf. 120, 103352. doi: 10.1016/
j.jag.2023.103352
Xu, K., Xie, Q., Song, M., Cap, W., Zhu, Y., and Ni, J. (2025). Development of weed
detection and management system using multi-modal information fusion for wheat
fields. T r a n s .C h i n .S o c .A g r i c .E n g .41, 175 –182. doi: 10.11975/j.issn.1002-
6819.202412210
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org13

Zhang, J., Huang, Y., Reddy, K. N., and Wang, B. (2019). Assessing crop damage from
dicamba on non-dicamba-tolerant soybean by hyperspectral imaging through machine
learning. Pest. Manage. Sci. 75, 3260 –3272. doi: 10.1002/ps.5448
Zhang, J., Maleski, J., Jespersen, D., Waltz, F. C., Rains, G., and Schwartz, B. (2021).
Unmanned aerial system-based weed mapping in sod production using a convolutional
neural network. Front. Plant Sci. 12, 702626. doi: 10.3389/fpls.2021.702626
Zhang, W., Miao, Z., Li, N., He, C., and Sun, T. (2022). Review of current robotic
approaches for precision weed management. Curr. Rob. Rep. 3, 139–151. doi: 10.1007/
s43154-022-00086-5
Zhu, H., Barlow, A., Hovinga, S., McConnell, L., Ramalingam, R., Ramanarayanan, T., et al. (2026).
Precision application of pesticides: challenges, opportunities, and impacts on crop production
sustainability.ACS Agric. Sci. Technol.6 (6), 783–796. doi: 10.1021/acsagscitech.5c01163
Chaudhary et al. 10.3389/fagro.2026.1898506
Frontiers in Agronomy frontiersin.org14

论文原文与出版信息

Frontiers in Agronomy
查看论文原文

继续探索