Advanced Plant Phenotyping Technologies for Enhanced Detection and Mode of Action Analysis of Herbicide Damage Management

Abstract

Weed control is fundamental to modern agriculture, underpinning crop productivity, food security, and the economic sustainability of farming operations. Herbicides have long been the cornerstone of effective weed management, significantly enhancing agricultural yields over recent decades. However, the field now faces critical challenges, including stagnation in the discovery of new herbicide modes of action (MOAs) and the escalating prevalence of herbicide-resistant weed populations. High research and development costs, coupled with stringent regulatory hurdles, have impeded the introduction of novel herbicides, while the widespread reliance on glyphosate-based systems has accelerated resistance development. In response to these issues, advanced image-based plant phenotyping technologies have emerged as pivotal tools in addressing herbicide-related challenges in weed science. Utilizing sensor technologies such as hyperspectral, multispectral, RGB, fluorescence, and thermal imaging methods, plant phenotyping enables the precise monitoring of herbicide drift, analysis of resistance mechanisms, and development of new herbicides with innovative MOAs. The integration of machine learning algorithms with imaging data further enhances the ability to detect subtle phenotypic changes, predict herbicide resistance, and facilitate timely interventions. This review comprehensively examines the application of image phenotyping technologies in weed science, detailing various sensor types and deployment platforms, exploring modeling methods, and highlighting unique findings and innovative applications. Additionally, it addresses current limitations and proposes future research directions, emphasizing the significant contributions of phenotyping advancements to sustainable and effective weed management strategies. By leveraging these sophisticated technologies, the agricultural sector can overcome existing herbicide challenges, ensuring continued productivity and resilience in the face of evolving weed pressures.

1. Introduction

1.1. Challenges in Herbicide Development and Usage

Weed control is fundamental to modern agriculture, crucial for maximizing crop yields, ensuring food security, and maintaining the economic viability of farming operations [1]. Weeds aggressively compete with crops for vital resources such as sunlight, water, and nutrients, leading to significant reductions in productivity and quality [2]. Herbicides have long been the cornerstone of effective weed management due to their efficiency and ease of application over large agricultural areas, substantially contributing to increases in agricultural productivity over the past several decades [3]. However, the continued success of herbicides is now facing significant challenges.

One of the most pressing issues is the stagnation in the discovery of new herbicide modes of action (MOAs). Remarkably, no new MOA has been introduced in the past 30 years [4]. This stagnation poses a severe threat to sustainable weed management because novel MOAs are critical for combating herbicide-resistant weed populations [5]. Without new MOAs, the effectiveness of herbicides diminishes over time as weeds evolve resistance mechanisms, rendering existing herbicides less effective or even obsolete [6].

High research and development (R&D) costs significantly hinder the discovery and commercialization of new herbicides with novel modes of action (MOAs), often requiring over a decade and investments exceeding USD 250 million [7]. These substantial and risky expenditures, compounded by stringent regulatory requirements and uncertain market returns, lead many companies to modify existing compounds rather than innovate new ones. Simultaneously, the widespread adoption of glyphosate-resistant (GR) crops has resulted in an over-reliance on glyphosate-based herbicides [8]. This extensive use has accelerated the emergence of glyphosate-resistant weed species, which undermine the benefits of GR crops and pose significant challenges to global crop production. Consequently, there is an urgent need for alternative weed management strategies to address these growing concerns.

Given these challenges, there is a critical need for innovative solutions in weed management.

• Monitoring and detecting herbicide damage on crops is essential to minimize its detrimental effects. By utilizing advanced monitoring technologies such as sensor networks and remote sensing, farmers and agronomists can quickly identify instances of drift and implement timely mitigation strategies.
• Analyzing herbicide resistance in weeds is also paramount. Understanding the genetic and biochemical mechanisms behind resistance can inform the development of management strategies to mitigate its spread. This includes rotating herbicides with different MOAs and integrating non-chemical control methods.
• The development of new herbicides with novel MOAs remains a high priority. Discovering new targets for herbicide action can rejuvenate the herbicide pipeline and provide fresh tools to combat resistant weeds. Identifying herbicide MOAs and analyzing their interactions at the molecular level can lead to the design of more effective and selective compounds.

1.2. Role of Phenotyping and Advanced Technologies

Image-based plant phenotyping, which involves the identification and quantification of observable plant characteristics (phenotypes) through imaging technologies, has become a pivotal tool in addressing herbicide challenges [9,10,11,12,13]. This approach focuses on capturing detailed images using techniques such as hyperspectral, multispectral, fluorescence and thermal imaging methods to assess how plants interact with their environment and genetic makeup. In herbicide research, image phenotyping plays a critical role in understanding how different herbicides affect plant growth and development, as well as in identifying and characterizing herbicide resistance [14,15,16].

One of the paramount challenges is monitoring and detecting herbicide drift to minimize its detrimental effects. Herbicide drift can lead to the unintended exposure of non-target crops and ecosystems, causing damage that ranges from reduced crop yields to loss of biodiversity [17]. By employing advanced monitoring technologies like sensor networks and remote sensing within image phenotyping frameworks, farmers and agronomists can quickly identify instances of drift. Hyperspectral imaging, for example, can detect subtle changes in plant reflectance patterns indicative of herbicide exposure, allowing for timely mitigation strategies to be implemented [10,18].

Analyzing herbicide resistance in weeds is another critical need in modern agriculture. Understanding the genetic and biochemical mechanisms behind resistance informs the development of management strategies to mitigate its spread. Image phenotyping plays a vital role in this regard by identifying phenotypic markers associated with resistance mechanisms [18,19]. Techniques such as chlorophyll fluorescence imaging measure disruptions in photosynthetic processes caused by herbicides, which can indicate resistance [16]. By integrating machine learning algorithms with imaging data, complex patterns of resistance can be deciphered, facilitating the rotation of herbicides with different modes of action (MOAs) and the integration of non-chemical control methods.

The development of new herbicides with novel MOAs remains a high priority to combat resistant weed populations. Discovering new targets for herbicide action can rejuvenate the herbicide pipeline and provide fresh tools against resistant weeds. Image phenotyping contributes to this effort by enabling the study of herbicide mechanisms of action through detailed phenotypic responses. For instance, thermal imaging can detect changes in plant canopy temperature resulting from herbicide-induced stress, offering insights into how plants metabolize and respond to different compounds at physiological levels [20]. Identifying herbicide MOAs and analyzing their interactions at the plant level through imaging techniques can lead to the design of more effective and selective compounds [21].

Advancements in image phenotyping, such as high-throughput phenotyping platforms and the integration of multiple sensor technologies, enhance the ability to address these challenges [22]. Automated systems utilizing robotics and conveyor belts allow for the rapid processing of large numbers of plant samples, essential for screening herbicide efficacy and resistance on a broad scale [23,24]. Machine learning algorithms further amplify the capabilities of image phenotyping by analyzing complex datasets to identify patterns and predict herbicide resistance or susceptibility with high accuracy [10,25,26].

1.3. Outline

This article is structured into three main sections, meticulously detailing the application of image phenotyping technologies to address herbicide challenges in weed science. Section 1 provides an overview of plant phenotyping technologies applied in weed science, discussing various sensor types, including hyperspectral, multispectral, RGB, fluorescence, and thermal imaging. Section 2 explores the applications of imaging techniques in herbicide challenges, focusing on monitoring and detecting herbicide drift, analyzing herbicide resistance in weeds, and aiding in the development of new herbicides and discusses unique findings in plant phenotyping related to herbicides, highlighting significant breakthroughs and innovative applications. Section 3 examines the limitations and future work for plant phenotyping, addressing current challenges and suggesting potential research directions.

Papers were searched using the Web of Science database, and research published after 2015 concerning the application of plant phenotyping in crop injury detection, weed resistance investigation, and herbicide mode of action identification was obtained. Although plant phenotyping is a trend in the field of agriculture, its application in the area of herbicides is still at the development stage. This review will summarize the achievements and proof of concept from the perspective of engineering.

2. Overview of Plant Phenotyping Technologies

The development of image-based phenotyping has significantly advanced the field of herbicide response assessment. These non-destructive methods have enabled high-throughput screening and detailed analysis of plant responses to herbicide treatment, crucial for evaluating the sometimes severe damage caused by herbicides. Such technologies necessitate non-destructive data collection techniques to assess these impacts accurately. Utilizing a range of sensors and platforms, imaging-based phenotyping captures intricate data on plant traits, providing unique advantages tailored to specific research needs.

2.1. RGB and Multispectral Sensing

RGB imaging, utilizing standard digital cameras capturing red, green, and blue, forms the foundation of many image-based phenotyping studies. Its accessibility, low cost, and ease of use has made it a valuable tool, particularly for initial screening and assessing gross phenotypic changes in response to herbicide application. RGB images can effectively capture visible traits such as plant growth, overall vigor, the presence and extent of chlorosis (loss of green color), and necrosis (tissue death). These visual symptoms are often indicative of herbicide stress and can be used to assess herbicide efficacy and resistance. For example, Ramirez-Rojas et al. [27] used RGB imaging to evaluate the effects of mesotrione on corn plants, observing chlorosis as a key indicator of herbicide impact. While RGB imaging provides valuable visual information, its limited spectral information restricts its capacity for the quantitative assessment of subtle physiological changes or biochemical alterations within plants.

Multispectral imaging extends beyond the visible spectrum, employing sensors that capture light across multiple discrete wavelengths, including near-infrared (NIR) and sometimes shortwave infrared (SWIR) regions. This broader spectral range provides significantly richer data compared to RGB images, allowing for the calculation of vegetation indices (VIs). VIs are mathematical combinations of reflectance values at different wavelengths and serve as quantitative proxies for various plant properties, including biomass, chlorophyll content, and overall plant health. The use of VIs significantly enhances the objectivity and quantitative nature of herbicide phenotyping. For instance, Duddu et al. [28] used multispectral imaging to calculate the optimized soil-adjusted vegetation index (OSAVI) in faba beans, demonstrating its superior precision compared to traditional visual ratings for assessing herbicide tolerance. The enhanced information content of multispectral imaging allows for a more accurate and consistent assessment of herbicide effects, facilitating more reliable comparisons across treatments and genotypes.

2.2. Hyperspectral Imaging

Hyperspectral imaging (HSI) is a non-destructive optical sensing technology that combines continuous wavelengths and imaging to capture and analyze a wide range of wavelengths across the electromagnetic spectrum [9,25,29,30,31]. Unlike conventional imaging systems that capture images in three broad bands corresponding to red, green, and blue, HSI systems acquire images across hundreds of narrow, contiguous spectral bands, producing a three-dimensional dataset known as a hypercube [25,30,31]. This hypercube contains both spatial and spectral information for each pixel, allowing for the detailed analysis of the chemical and physical properties of plant tissues [32]. For example, glyphosate applications can alter the chemical composition of leaves, affecting light absorption patterns in glyphosate-resistant and glyphosate-susceptible plants, which can be detected using HSI [8].

The effectiveness of hyperspectral imaging (HSI) in herbicide research is critically dependent on advanced hardware components and platforms. Hyperspectral cameras, essential for capturing detailed spectral data, operate across a broad spectral range, including the visible and near-infrared (VNIR) regions, with high spectral resolution. These cameras are often mounted on unmanned aerial vehicles (UAVs) [12] or handheld devices like LeafSpec [22] for field applications, or integrated into conveyor-based systems for high-throughput screening in controlled environments. Spectrometers such as the analytical spectral device (ASD) FieldSpec provide precise measurements essential for plant analysis [31]. Stable illumination sources, like quartz-tungsten-halogen lamps, ensure consistent lighting conditions necessary for accurate spectral data capture. Calibration with reference panels and the control of environmental variables in laboratory settings are crucial for maintaining data integrity, making the hardware configuration and software integration foundational to leveraging HSI in modern agricultural practices.

Despite its advantages, hyperspectral imaging (HSI) has inherent limitations that must be considered. The high dimensionality of hyperspectral data introduces redundancy and multicollinearity among spectral bands, complicating data analysis and interpretation [33]. To effectively mine HSI data, there is a growing application of both supervised and unsupervised machine learning methods. Statistical machine learning techniques are widely used to extract features from the spectral dimensions of HSI. Additionally, neural networks, recognized as robust and powerful supervised tools, are employed to utilize both the spatial and spectral dimensions of HSI data. The unique and large data characteristics of HSI also make it challenging to apply existing pre-trained RGB deep learning models directly. Training neural networks for HSI requires substantially more computational resources compared to RGB images. Additionally, high-resolution HSI systems can be costly, potentially limiting their accessibility for widespread use in both research and practical applications.

2.3. Fluorescence Imaging

Fluorescence imaging is a non-destructive optical sensing technique that measures the re-emission of light by chlorophyll molecules during photosynthesis. This method provides detailed insights into the photosynthetic efficiency and physiological status of plants by assessing parameters related to photosystem II (PSII) activity [29,34]. In herbicide research, fluorescence imaging has been extensively applied to detect herbicide-induced stress, analyze herbicide resistance in weeds, and aid the development of new herbicides with novel modes of action. The following

Table 1. Fluorescence parameters used in herbicide plant phenotyping.

Parameter	Definition	Formula	Key Applications	References
$Fv/Fm$	Maximum quantum yield of PSII photochemistry	${\displaystyle \frac{F_m-F_0}{Fm}}$	Used to screen for metabolic perturbations, detect early stress responses, and identify herbicide-resistant weeds	[19,29,33,35]
$\Phi PSII$	Operating efficiency of PSII	${\displaystyle \frac{F_s-F_0}{F_s}}$	Serves as a bioindicator of photosynthetic machinery damage and stress evaluation	[29,33,34,36,37]
NPQ	Reflects heat dissipation of excess energy in PSII antenna complexes	${\displaystyle \frac{Fm- {Fm}^{\prime }}{{Fm}^{\prime }}}$	Evaluates photoprotection mechanisms and the degree of thermal energy dissipation under stress conditions	[29,35,36,38]
$Fd/Fm$	Fraction of maximum fluorescence dissipated under steady-state conditions	${\displaystyle \frac{F_m-F_s}{F_m}}$	Contributes to understanding the balance between photochemical utilization and energy dissipation	[33]
qP	Coefficient of photochemical quenching	${\displaystyle \frac{{F_m}^{\prime }-F_s}{{F_m}^{\prime }-F_0}}$	Reflects the proportion of open PSII reaction centers and is used to assess the efficiency of the photochemical phase	[29]

shows some common fluorescence parameters used in fluorescence research.

However, this technique has several limitations. The measurement protocol requires exposing plants to continuous actinic red light (617 nm; 600 μmol) and a saturating pulse (cool white; 2000 μmol) following 25 min of dark adaptation [35], which makes the imaging process more suitable for indoor or laboratory proximal sensing applications. The throughput is also lower compared with other phenotyping technology like hyperspectral methods. The equipment required for high-resolution fluorescence imaging can be costly, potentially limiting accessibility for some researchers [35]. Despite these challenges, fluorescence imaging remains a powerful tool for researchers in breeding programs and herbicide development groups, where controlled conditions facilitate precise measurements and analyses.

2.4. Thermal Imaging

Thermal imaging technology has been used because of its capacity to detect differences in temperature between herbicide-resistant and -susceptible weed populations. This method hinges on the observation that susceptible plants, particularly when treated with herbicides like glyphosate [20,39], exhibit a reduction in photosynthesis and stomatal conductance, leading to decreased transpiration and a subsequent increase in leaf temperature. Conversely, resistant plants do not undergo similar physiological changes, thus maintaining lower temperatures.

Thermal imaging employs long-wave infrared (LWIR) cameras to capture the radiation emitted from plant canopies, converting it into temperature readings. These cameras can be deployed on unmanned aerial vehicles (UAVs) for expansive field studies or utilized in more controlled environments such as greenhouses. The subsequent steps involve extracting temperature data from the images, focusing primarily on the plant canopy, and using these data to discern temperature differences between resistant and susceptible plant biotypes.

However, the application of thermal imaging in field environments presents challenges. Although greenhouse studies have shown promising results, such as classifying glyphosate-resistant horseweed with up to 89% accuracy, field studies have yielded inconsistent outcomes [40]. Factors such as uneven solar radiation, spatial variability within weed populations, and varying degrees of resistance impact the reliability of thermal imaging as a predictive tool for herbicide resistance. Additionally, the presence of crops and other vegetation can skew thermal readings, further complicating data interpretation.

Given these limitations, research suggests that alternative sensing technologies might provide more dependable results in field conditions [20,40]. For instance, multispectral imaging techniques that leverage indices like the Normalized Difference Vegetation Index (NDVI) or specific wavelength reflectance have demonstrated superior efficacy in some studies.

2.5. Integration of Machine Learning with Image Data for Herbicide Resistance Detection

A variety of machine learning algorithms—ranging from traditional approaches to advanced deep learning models—have been successfully employed to enhance the prediction accuracy of herbicide resistance using image-based plant phenotyping. Traditional machine learning methods, such as Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbor (KNN), Partial Least Squares Discriminant Analysis (PLS-DA), Fisher Linear Discriminant Analysis (FLDA), and Naive Bayes, have been widely applied in this context. For example, Nugent et al. [15] utilized an SVM with a radial basis function kernel on hyperspectral data to differentiate herbicide-resistant kochia strains, while Zhang et al. [26] demonstrated that SVMs, along with RF and KNN, could effectively assess soybean and corn damage, respectively. Random Forest, known for its robustness in handling high-dimensional and multicollinear data, has achieved high accuracy in classifying soybean responses to dicamba [41]. PLS-DA and FLDA have also been applied successfully to reduce data dimensionality and enhance class separation, providing effective means to handle the complex spectral information from hyperspectral imagery. The details of advantages and limitations of statistical machine learning methods was well discussed by Tanzeel et al. [42].

Deep learning approaches, particularly Convolutional Neural Networks (CNNs) [43] and Artificial Neural Networks (ANNs), have revolutionized image processing due to their ability to automatically extract hierarchical features directly from raw data. Shallow CNNs, such as the SCNN-ATT model proposed by Chu et al. (2022) [44], have been tailored for herbicide type classification using hyperspectral imaging, offering advantages over deeper networks when data are limited. Models like AlexNet [45], VGG-16 [46], ResNet50 [47], and DenseNet121 have also been employed to process both hyperspectral and UAV-based RGB images for the early detection of herbicide stress, with notable improvements in classification accuracy. In addition, specialized models such as HerbiNet [48] and ATT-RES have been developed for predicting maize resistance to herbicides like nicosulfuron; these models incorporate residual connections and attention mechanisms to capture long-range dependencies in spectral data. Deep learning models not only excel in handling high-dimensional data, but also provide improved generalizability through techniques such as transfer learning, thereby enhancing the robustness of herbicide resistance prediction across diverse environmental conditions and crop varieties.

3. Applications of Imaging Techniques in Herbicide Challenges

3.1. Detection of Herbicide Damage on Crops

The primary advantages of integrating deep learning with image-based phenotyping include automatic feature extraction, which reduces the need for manual feature engineering, and the ability to learn complex, non-linear relationships between spectral features and plant resistance phenotypes. These capabilities enable the detection of subtle physiological changes that might otherwise be missed by traditional methods. Moreover, deep learning models are particularly effective in processing large volumes of high-dimensional data, making them well suited for hyperspectral images where each pixel contains information across hundreds of wavelengths.

Traditionally, herbicide injury on crops is assessed visually in field trials. Researchers and agronomists inspect treated plots and compare them to untreated controls, looking for characteristic damage symptoms. Key visible indicators include leaf discoloration (chlorosis or other abnormal coloration), the wilting or drooping of foliage, the development of necrotic (dead) tissue spots or margins, and general growth inhibition or stunting [17,49]. These symptoms often appear within days or weeks after herbicide application, depending on the chemical’s mode of action and the crop’s sensitivity. By noting the type and pattern of injury (e.g., interveinal yellowing, leaf tip burn, malformed new growth), experts can often infer which class of herbicide caused the damage.

The following

Table 2. Application of image phenotyping on detection of herbicide damage on crops.

Crop Type	Sensor Type	Herbicide Name	Herbicide Group Number	Herbicide Group Name	Reference
Wheat	Hyperspectral	Mesosulfuron-methyl	2	ALS inhibitors	[44]
Corn	Hyperspectral	Nicosulfuron	2		[50]
Maize	Hyperspectral	Nicosulfuron	2		[48]
Soybean	Hyperspectral	2,4-D	4	Synthetic auxins	[18]
Cotton	Hyperspectral	2,4-D	4		[51]
Soybean	Hyperspectral, RGB	Dicamba	4		[52]
Soybean	RGB	Dicamba	4		[41]
Soybean	Hyperspectral	Dicamba	4		[18]
Soybean	Hyperspectral	Dicamba	4		[36]
Wheat	Hyperspectral	MCPA-Na	4		[44]
Sugar Beet	Chlorophyll Fluorescence Imaging	Desmedipham	5	Photosystem II inhibitors	[53]
Sugar Beet	Chlorophyll Fluorescence Imaging	Phenmedipham	5		[53]
Wheat	Hyperspectral	Isoproturon	7	Photosynthesis inhibitors	[44]
Sugar Beet	Chlorophyll Fluorescence Imaging	Lenacil	7		[53]
Soybean	Multispectral	Glyphosate	9	EPSP synthase inhibitors	[54]
Soybean	Hyperspectral	Glyphosate	9		[55]
Black nightshade	Hyperspectral	Glyphosate	9		[30]
Corn	Hyperspectral	Glyphosate	9		[56]
Maize	Hyperspectral	Glyphosate	9		[56]
Black nightshade	Hyperspectral	Glufosinate	10	Glutamine synthetase inhibitor	[30]
Soybean	Multispectral	Glufosinate	10		[57]
Sugar Beet	Chlorophyll Fluorescence Imaging	Ethofumesate	16	HRAC Group F3	[53]

summarizes key studies that have employed different phenotyping technologies, ranging from hyperspectral and multispectral imaging to chlorophyll fluorescence imaging. Each entry details the type of crop studied, the specific sensors used, the herbicides applied, and their respective action groups.

According to the table, dicamba, 2,4-D, and glyphosate were detected frequently as a popular weed management method for crops like corn and soybeans. Injury can be divided into two types: direct injury and drifting damage. The mean difference is the dosage of the herbicide applied, where drifting damage is caused by herbicides applied at a much lower rate than the rate of field application. This leads to differences in the level of stress.

These approaches can achieve over 80% accuracy in the detection of specific herbicide damage on crops within 72 h, which is much earlier than can be achieved by human eyes. This proves the advantage of plant phenotyping, and high-throughput sequencing and sensors such as hyperspectral cameras can provide more information than visual observation.

In herbicide research, HSI has been instrumental in monitoring and detecting herbicide drift. It identifies spectral changes in plants resulting from unintended herbicide exposure, enabling the detection of herbicide drift [18,58]. Herbicide drift can cause physiological and biochemical alterations in non-target plants, affecting their spectral reflectance properties. By detecting these subtle spectral variations, HSI allows for the early and non-destructive detection of herbicide drift, facilitating timely interventions to minimize crop damage. For instance, researchers used a field spectrometer and a hyperspectral imaging camera to measure grass sods treated with glyphosate, drought-stressed plants, and control plants [29]. They found that certain spectral indices, such as the normalized difference lignin index (NDLI) and indices related to photosynthetic pigments like the carotenoid reflectance index (CRI-1) and the photochemical reflectance index (PRI), were sensitive to glyphosate treatment as early as two days after application. Another study utilized visible/near-infrared hyperspectral imaging (Vis/NIR HSI) to distinguish glyphosate types and stress levels in wheat seedlings at different stress durations, observing spectral reflectance differences at specific wavelengths [25].

Although hyperspectral imaging combined with advanced machine learning techniques consistently achieves high accuracy (often exceeding 80%) in detecting herbicide-induced crop stress, a proportion of undetectable cases persists. These undetectable instances frequently result from the complex polygenic adaptations that confer crop resistance to herbicides. Polygenic resistance typically involves multiple subtle physiological changes, including modified enzyme activity, altered herbicide uptake, reduced translocation, or enhanced metabolic detoxification pathways, resulting in minimal visual or biochemical alterations following herbicide exposure [51]. Such subtle or delayed plant responses, especially at early herbicide exposure stages or low dosage levels, present significant challenges for detection, as typical imaging methods rely heavily on distinct physiological or biochemical perturbations induced by herbicides. For instance, resistant plants might exhibit spectral and physiological characteristics nearly identical to untreated control plants, leading to high misclassification rates when relying solely on average spectral indices without integrating spatial or temporal variability information.

Previous studies explicitly highlight this issue by noting that slight physiological differences between resistant and susceptible genotypes may go undetected when employing standard vegetation indices or generalized spectral analysis [21]. The inability of standard vegetation indices or spectral analysis to distinguish subtle polygenic adaptations reflects the complexity of resistance mechanisms, including altered enzyme activity, reduced herbicide translocation, and enhanced metabolic detoxification pathways, which do not always result in immediate or easily measurable physiological shifts. Consequently, undetectable proportions may vary significantly depending on crop species, the genetic complexity of resistance traits, and the specific herbicide modes of action. Addressing this limitation necessitates developing imaging techniques with higher spectral sensitivity, improved spatial resolution, and more sophisticated analytical frameworks capable of capturing minute spectral shifts associated with polygenic traits.

Future efforts to enhance detection capabilities may benefit from integrating complementary analytical strategies. Utilizing hyperspectral imaging with deep learning models that integrate temporal and spatial data, as well as biochemical markers identified through targeted metabolomics, could substantially increase sensitivity to subtle phenotypic changes resulting from polygenic adaptations. Further empirical research under diverse field conditions and involving various crop genotypes and herbicide combinations is critical to accurately quantify the proportion of undetectable herbicide responses. Such advancements will ultimately contribute to more robust methodologies for identifying complex resistance mechanisms and ensuring timely and effective resistance management strategies in agriculture.

The analysis, while comprehensive, reveals the necessity for the broader environmental validation of sensor technologies to ensure reliability under varied field conditions. Future research should aim to establish standardized protocols for deploying these technologies in different climates and across a wider range of soil types.

Moreover, expanding the database to include underrepresented herbicides and crops could provide a more complete picture of the global state of herbicide impact. Integrating sensor feedback with automated application systems could also revolutionize real-time herbicide management, enhancing both efficacy and sustainability.

3.2. Weed Herbicide Resistance Analysis

Classical resistance testing relies heavily on whole-plant assays in greenhouse or field conditions. Seeds or individuals from a suspected herbicide-resistant weed population are grown alongside a known susceptible population, and both are treated with the herbicide in question. By comparing survival or growth, researchers can confirm resistance if the suspect population withstands doses that kill the susceptible plants. Dose–response experiments are a gold-standard approach: plants are treated with a range of herbicide doses to generate a response curve (e.g., percent biomass reduction or mortality). From these, one can determine metrics like ED₅₀ or LD₅₀ for each population. Performing a full dose–response (instead of just a single label rate) evaluation reveals the magnitude of resistance and is especially useful for quantifying how much more of a herbicide the resistant biotype tolerates [59].

This process can take weeks or months: seeds must be matured and harvested, often after the growing season, and then broken from dormancy and grown to a certain stage for spraying. These assays demand significant greenhouse space and tending (watering, transplanting, spraying), which means high labor and facility costs. Because each test occupies space and time, only a limited number of weed populations can be screened concurrently, hindering scalability [60].

DNA-based methods (e.g., PCR tests for known resistance mutations) offer a faster alternative to growing plants. They can be scaled up in the lab and even made high-throughput (for instance, using real-time PCR to skip gel electrophoresis and handle many samples at once) [59]. These assays can turn around results in days and process large numbers of plants or even pooled samples. However, they require prior knowledge of the specific gene mutation conferring resistance. They will not detect unknown resistance mechanisms (such as metabolic resistance or new target-site mutations) and can give false negatives if the resistant population does not carry the tested marker. Moreover, setting up molecular assays for each weed species and each herbicide mode of action demands significant upfront labor in method development and still involves per-sample costs for DNA extraction and reagents.

Given the above constraints, classical resistance screening cannot easily keep up with the scale of the problem. Weed scientists face a growing number of cases across large regions, but high-throughput phenotyping remains a technical bottleneck in resistance management. The integration of sensor data with sophisticated analytical methods is pivotal for assessing the efficacy of herbicides and the adaptive resistance traits exhibited by these plants. Such studies are crucial for refining management strategies and mitigating the impacts of resistance across agricultural systems.

Table 3. Herbicide resistance research utilizing plant phenotyping.

Weed Scientific Name	Sensor Type	Herbicide Group Number	Herbicide Group Name	Reference
Alopecurus myosuroides	Chlorophyll Fluorescence Imaging	1	ACCase inhibitors	[61]
Alopecurus myosuroides	Chlorophyll Fluorescence Imaging	2	ALS inhibitors	[16]
Papaver rhoeas	Chlorophyll Fluorescence Imaging	2	ALS inhibitors	[62]
Stellaria media	Chlorophyll Fluorescence Imaging	2	ALS inhibitors	[62]
Kochia scoparia, marestail, Conyza canadensis, Chenopodium album	Hyperspectral	4	Synthetic auxins	[63]
Kochia scoparia	Hyperspectral	4	Synthetic auxins	[15]
Echinochloa crus-galli	Multispectral, RGB	5	Photosystem II inhibitors	[64]
Abutilon theophrasti	Multispectral, RGB	5	Photosystem II inhibitors	[64]
Amaranthus palmeri	Hyperspectral	9	Glyphosate (EPSP synthase inhibitors)	[8]
Kochia, Conyza canadensis, Chenopodium album	Hyperspectral	9	Glyphosate (EPSP synthase inhibitors)	[63]
Kochia scoparia	Hyperspectral	9	Glyphosate (EPSP synthase inhibitors)	[15]
Amaranthus rudis	Thermal	9	Glyphosate (EPSP synthase inhibitors)	[40]
Conyza canadensis	Thermal	9	Glyphosate (EPSP synthase inhibitors)	[40]
Amaranthus rudis, Kochia scoparia, Ambrosia artemisiifolia	Multispectral	9	Glyphosate (EPSP synthase inhibitors)	[65]
Kochia scoparia	Thermal, Multispectral	9	Glyphosate (EPSP synthase inhibitors)	[20]
Amaranthus retroflexus	Multispectral	10	Glutamine synthetase inhibitors	[57]
Amaranthus retroflexus	Multispectral	14	PPO inhibitors	[66]

presents a compilation of research efforts that employ different sensor technologies to investigate herbicide resistance in diverse weed species.

Herbicide resistance, an escalating issue in agricultural management, profoundly affects crop yield and economics due to the resilience that weeds develop against various herbicide actions. This resistance stems from several complex mechanisms, such as target-site mutations that alter the protein targeted by the herbicide, non-target-site resistance mechanisms including enhanced metabolism or reduced translocation, and gene amplification. Notably, resistance is not just widespread, but also varied—glyphosate resistance is notably pervasive, and ALS inhibitors are frequently rendered ineffective across numerous farming regions. The implication of such widespread resistance extends beyond increased management costs; it fundamentally alters crop management strategies and necessitates the adoption of integrated weed management systems.

Thermal imaging provides a powerful tool for differentiating between herbicide-resistant and -susceptible plants. Herbicide-resistant plants may exhibit distinct temperature patterns compared to susceptible plants in response to herbicide treatment. This difference in thermal response can be attributed to variations in physiological processes affected by herbicide. For example, a resistant plant might maintain a lower leaf temperature due to sustained transpiration, while a susceptible plant might exhibit a higher leaf temperature due to reduced transpiration resulting from herbicide-induced stress. This method offers a rapid and non-destructive way to screen for herbicide resistance, facilitating the efficient high-throughput screening of large plant populations [20]. This high-throughput screening capability is especially valuable in breeding programs where large numbers of plants need to be evaluated for herbicide tolerance.

Fluorescence imaging has been instrumental in analyzing herbicide resistance in weeds [62,67]. Resistant plants often exhibit less pronounced changes in fluorescence parameters following herbicide application, reflecting their ability to maintain photosynthetic function despite the presence of herbicides [62]. By comparing fluorescence responses between resistant and susceptible weed biotypes, the mechanisms of resistance can be better understood [67].

3.3. Discovery of Herbicide Mode of Actions

Unraveling a herbicide’s mode of action (MOA) traditionally involves extensive experimentation. Researchers must observe what symptoms the herbicide causes, run physiological assays (e.g., measure photosynthesis rate, pigment levels, growth inhibition), and perform biochemical tests targeting suspected sites (such as enzyme activity assays for a certain pathway). This investigative process is iterative and time-intensive, often proceeding by elimination. Each hypothesis about the MOA (for instance, “perhaps it inhibits ALS enzyme”) requires a separate experiment to confirm or deny. Such molecular and biochemical analyses are slow and costly, sometimes taking months per compound [68]. If the first guesses are wrong, scientists have to try other approaches, further extending the timeline. In practice, discovering a new herbicide’s target site can span years of work, especially without any high-throughput clues. In contrast, high-throughput phenotyping (HTP) technologies offer a transformative solution to these limitations.

The characteristics of image phenotyping and its effectiveness are closely influenced by the herbicide’s mechanism of action, as different herbicide groups induce distinctive physiological and biochemical responses in plants.

Table 4. Summary of herbicide groups, their respective modes of action, and the suitability of sensor types for detecting physiological responses.

Herbicide Name	Herbicide Group Number	Herbicide Group Name	Sensor Type	Reference
Pinoxaden	1	ACCase inhibitors	Chlorophyll Fluorescence Imaging	[67]
U-46 Combi Fluid	2	ALS inhibitors	Chlorophyll Fluorescence Imaging	[13]
Penoxsulam			RGB, Thermal, Chlorophyll Fluorescence Imaging	[33]
Chlorimuron			Hyperspectral	[21]
Amidosulfuron			Raman spectroscopy, chlorophyll fluorescence imaging	[69]
Cruz	4	Synthetic auxins	Chlorophyll Fluorescence Imaging	[13]
2,4-D			Hyperspectral	[70]
Atrazine	5	Photosystem II inhibitors	Hyperspectral	[21]
Bentazon	6		Chlorophyll Fluorescence Imaging	[19]
Basagran			Chlorophyll Fluorescence Imaging	[13]
Bromicide			Chlorophyll Fluorescence Imaging	[13]
Dinoseb			Hyperspectral	[21]
Glyphosate	9	EPSP synthase inhibitors	RGB, Thermal, Chlorophyll Fluorescence Imaging	[33]
Glyphosate			Hyperspectral	[30]
Glyphosate			Hyperspectral	[21]
Glufosinate	10	Glutamine synthetase inhibitors	RGB, Thermal, Chlorophyll Fluorescence Imaging	[33]
Glufosinate			Hyperspectral	[30]
Glufosinate			Hyperspectral	[21]
Diflufenican	12	Carotenoid biosynthesis inhibitors	Raman spectroscopy, chlorophyll fluorescence imaging	[69]
Clomazone	13	Long-chain fatty acid inhibitors	Raman spectroscopy, chlorophyll fluorescence imaging	[69]
Tiafenacil	14	PPO inhibitors	RGB, Thermal, Chlorophyll Fluorescence Imaging	[33]
Flumioxazin			Hyperspectral	[21]
Carfentrazone-ethyl			Raman spectroscopy, chlorophyll fluorescence Imaging	[69]
Gramoxone	22	Photosystem I electron diverter	Chlorophyll Fluorescence Imaging	[13]
Paraquat	22		RGB, Thermal, Chlorophyll Fluorescence Imaging	[33]
Paraquat	22		Hyperspectral	[21]
Isoxaflutole	27	HPPD inhibitors	RGB, Thermal, Chlorophyll Fluorescence Imaging	[33]
Mesotrione	27	HPPD inhibitors	Raman spectroscopy, chlorophyll fluorescence Imaging	[69]
Indaziflam	29	Cellulose biosynthesis inhibitors (CBIs)	Hyperspectral	[21]
Pyroxsulam + Florasulam	2	ALS inhibitors	Chlorophyll Fluorescence Imaging	[67]
Lumax (S-metolachlor + Mesotrione + Terbuthylazine)	5	Photosystem II inhibitors	Chlorophyll Fluorescence Imaging	[13]
	15	Seedling Growth Inhibitors	Chlorophyll Fluorescence Imaging	[13]
	27	HPPD Inhibitors	Chlorophyll Fluorescence Imaging	[13]

summarizes the current herbicide mode action research conducted with certain types of imaging sensors.

Glyphosate, an EPSPS inhibitor, disrupts the shikimate pathway, causing the depletion of aromatic amino acids, alterations in chlorophyll and carotenoid pigments, and the accumulation of lignin. These physiological disruptions result in measurable spectral changes, particularly in the visible and near-infrared (VNIR) regions, detectable by hyperspectral imaging through indices like the NDLI, CRI-1, and PRI. Similarly, chlorophyll fluorescence imaging can detect glyphosate-induced stress via altered fluorescence parameters associated with impaired photosynthesis. Additionally, glyphosate application can cause an inhibition of stomatal conductance [40]. When stomatal conductance is reduced, the plant’s ability to perform transpiration is lowered. As a result, less heat is dissipated from the leaves, leading to an increase in the canopy temperature. Research using UAVs equipped with thermal cameras to detect glyphosate resistance in field conditions found it to be unreliable due to the susceptibility of thermal reflectance to environmental conditions.

Auxinic herbicides such as dicamba and 2,4-D induce pronounced morphological abnormalities, including leaf cupping and epinasty, making high spatial resolution hyperspectral imaging crucial for accurately detecting these herbicide-induced morphological changes. Spectral–spatial analysis significantly enhances classification accuracy over traditional spectral methods, making sensors capable of capturing fine spatial details, such as RGB cameras, multispectral cameras, and high-resolution hyperspectral cameras.

Photosynthesis inhibitors (e.g., atrazine, diuron) cause direct disruptions in the photosynthetic electron transport chain, leading to immediate alterations in chlorophyll fluorescence parameters and spectral shifts in the red-edge wavelengths due to declining chlorophyll content. Phenotyping technologies capable of measuring sun-induced fluorescence (SIF) or specific red-edge spectral shifts thus offer direct and sensitive detection of these herbicide impacts. Glutamine synthetase inhibitors, like glufosinate, cause disruptions in nitrogen metabolism and ammonia accumulation, which manifest as subtle spectral reflectance shifts detectable through hyperspectral imaging. Early detection is achievable by identifying key spectral bands sensitive to nitrogen stress and chlorophyll degradation.

Herbicides belonging to ALS inhibitors, PPO inhibitors, and cellulose synthesis inhibitors induce distinct physiological changes, including pigment degradation, leaf discoloration, and morphological alterations, each potentially linked to unique spectral signatures detectable through advanced imaging techniques. Typically, these herbicides are classified as contact herbicides, causing rapid and severe visible symptoms, such as leaf necrosis or chlorosis, shortly after field application. Consequently, hyperspectral imaging technologies, which provide high spectral resolution across visible and near-infrared (VNIR) wavelengths, effectively capture these quick physiological responses. Chlorophyll fluorescence and hyperspectral imaging are particularly effective, as they can detect subtle biochemical shifts within hours post treatment by capturing spectral regions sensitive to pigment changes and photosynthetic disruption. However, the detection accuracy can decline significantly if the imaging window is missed, especially for cell membrane disruptor herbicides (e.g., PPO inhibitors—Group 14 and Photosystem I electron diverters—Group 22). As these herbicides cause rapid cellular damage followed by potential plant tissue recovery, imaging too late may fail to capture the initial stress indicators, resulting in confusion in classification models. Thus, the precise timing of data acquisition relative to herbicide application and symptom manifestation is critical to maintaining accuracy in detecting and classifying herbicide-induced damage using phenotyping technologies.

HSI also contributes to the development of new herbicides with novel modes of action by providing insights into the biochemical responses of plants to experimental compounds. By analyzing spectral changes over time, researchers can understand a herbicide’s mode of action, target sites, and impact on plant physiological processes. This information is crucial for screening and refining new herbicide candidates to enhance efficacy and minimize off-target effects. A study employing HSI to classify eight herbicides with different sites of action achieved an overall accuracy of 81.5% one day after treatment using support vector machine (SVM) models [21]. The research identified distinct spectral feature bands associated with specific herbicide modes of action, demonstrating the feasibility of using HSI for high-throughput herbicide screening and aiding in the discovery of new herbicide targets.

In the development of new herbicides with novel modes of action, fluorescence imaging contributes by revealing how experimental compounds affect plant physiology at the molecular level [35,37]. Studies have demonstrated that PSII-inhibiting herbicides cause measurable increases in chlorophyll fluorescence due to inhibited electron flow [19,34]. The early detection of herbicide stress has been achieved by monitoring decreases in Fv/Fm and ΦPSII parameters, enabling interventions before irreversible damage occurs [38,53].

Hyperspectral imaging technologies, particularly when combined with fluorescence measurements, provide a robust approach for elucidating herbicide mechanisms of action at a physiological and biochemical level. For instance, Chu et al. [71] demonstrated the potential of chlorophyll fluorescence (OJIP transient) and visible/near-infrared (Vis/NIR) hyperspectral imaging to effectively characterize the mode of action (MOA) of the herbicide isoproturon (ISO) and the safener gibberellin acid (GA3) in wheat seedlings. Specifically, hyperspectral imaging revealed distinctive spectral regions sensitive to phytochemical changes induced by herbicide stress, such as elevated malondialdehyde (MDA) levels indicative of oxidative damage, and fluctuating glutathione (GSH) concentrations associated with plant detoxification responses. Further, an analysis of the OJIP transient indicated significant disruptions in the photosystem II (PSII) electron transport chain upon ISO application, particularly the blockage between the primary quinone acceptor (Q_A) and the secondary quinone acceptor (Q_B). Such insights were made possible due to the non-invasive real-time detection capability of hyperspectral imaging, allowing immediate detection (within 12 h post treatment) of subtle physiological changes, a timeframe significantly shorter than conventional destructive assays. Additionally, by using partial least squares regression (PLSR) combined with competitive adaptive reweighted sampling (CARS), Chu et al. [71] identified characteristic wavelengths specifically correlated with biochemical responses such as glutathione (GSH), malondialdehyde (MDA), chlorophyll (Chl), and carotenoids (Car). This methodology significantly enhanced the precision of biochemical stress indicator predictions (R_P values reaching 0.89 for carotenoid content). Thus, hyperspectral imaging, complemented by robust statistical and machine learning models, not only elucidates herbicide mechanisms at the biochemical and physiological levels, but also offers a rapid, accurate, and scalable approach to evaluating herbicide safeners’ effectiveness under practical agricultural conditions.

The variability in plant responses to herbicides, influenced by species-specific physiology, developmental stages, and environmental conditions, introduces inconsistencies in research outcomes and complicates the generalization of findings. Addressing this requires rigorous experimental designs that encompass diverse developmental stages and environmental settings. Additionally, the limited spatial resolution of certain phenotyping techniques, like those using fiber-optic probes, restricts data comprehensiveness, potentially missing significant intra- and inter-plant variations. Enhanced imaging techniques with higher spatial resolution could improve data granularity and utility. The field also suffers from a lack of standardized methods for sensor-based measurements, hindering result comparability across studies. Establishing universally accepted protocols and calibration standards is essential for advancing plant phenotyping. Moreover, the prohibitive cost of advanced imaging systems, especially hyperspectral cameras, limits their accessibility, particularly in resource-constrained environments. Investments in more cost-effective technologies or the development of shared resources could mitigate these financial barriers. Lastly, while hyperspectral imaging effectively detects plant stress and damage, accurately linking these spectral signatures to specific stressors or damage mechanisms remains a significant challenge, necessitating further detailed physiological and biochemical studies.

4. Conclusions, Future Perspectives, and Challenges

In addition to sensor type and herbicide mechanisms, environmental conditions significantly influence the quality and reliability of sensor data in herbicide phenotyping research. These environmental factors—including temperature fluctuations, humidity variations, and dynamic lighting conditions—can alter sensor performance and plant physiological responses, thus affecting the accurate detection of crop stress or herbicide resistance. Consequently, careful consideration and mitigation strategies for these environmental influences are essential to enhance the robustness and practical applicability of phenotyping methodologies under field conditions.

Field experiments inherently encounter variable environmental conditions that can affect sensor data quality. Huang et al. [12] acknowledge uncertainties in sun intensity, sun angle, and wind interference, which are exacerbated in on-the-go line-scan hyperspectral imaging systems. Zhang et al. [56] further illustrate that the specific geographical location, soil composition, and buffer zones to minimize spray drift play a pivotal role in shaping plant responses. These studies highlight that while randomized block designs and controlled experimental setups in the field attempt to mitigate environmental variability, factors such as diurnal light variations and ambient weather conditions remain influential.

Temporal factors are essential in determining the plant’s spectral response to herbicide stress. Chu et al. [44] and several other studies emphasize that the duration of herbicide stress and the timing of data acquisition directly impact the spectral measurements and subsequent analyses. Experiments conducted at different intervals—ranging from hours after treatment to several weeks—demonstrate that the plant’s response evolves over time, thereby necessitating a careful consideration of temporal dynamics to enhance model accuracy and generalizability.

Geographical location and seasonal conditions significantly influence both plant physiology and sensor performance. Studies by Tao et al. [72] show that experimental outcomes are affected by the specific climate, soil characteristics, and seasonal variations of the study area. By integrating datasets collected across different years, planting seasons, and diverse locations, researchers can better assess the robustness and generalizability of their models. Such integration is crucial for developing technologies that are applicable beyond a single experimental context.

The experimental setup, including the choice of growth medium and spatial arrangement, is critical for mitigating extraneous environmental influences. For instance, Niu et al. [18] conducted their experiments in a controlled greenhouse environment using specific potting soil mixtures, thereby standardizing the immediate growth conditions.

Light conditions, whether natural or artificial, are among the most challenging environmental factors to control in remote sensing studies. Variability in sunlight intensity, diurnal changes, and the angle of illumination can lead to significant discrepancies in measured spectral indices. Studies indicate that shadows, specular reflections, and sun glint may alter the apparent reflectance of plant canopies, thereby reducing the reliability of data collected in field settings.

To mitigate the impact of environmental variables, research can be conducted within controlled environments such as greenhouses or growth chambers where conditions such as temperature, humidity, and lighting are carefully regulated. Additionally, improvements in imaging hardware can also help reduce environmental effects; for example, incorporating self-lighting sources within imaging setups can eliminate the variability introduced by external light sources like sunlight or artificial greenhouse lights. This approach ensures a consistent imaging environment that is critical for acquiring reproducible phenotypic data.

The complexity and variability of plant responses to herbicide application, reflected in the spectral data, present a major challenge. The field currently lacks standardized data collection protocols, and there is a pressing need for a publicly accessible standard herbicide dataset. The inherently high throughput of phenotyping technologies complements the capabilities of machine learning, which can efficiently process and analyze large datasets to detect subtle patterns that may not be evident through traditional analysis methods.

Exploring the potential of Large Language Models (LLMs), such as vision transformers, could revolutionize herbicide analysis through imaging phenotyping. Traditional machine learning models, while highly explanatory, often lack stability; in contrast, deep learning models typically offer greater stability, but at the cost of reduced explainability. Vision transformers provide a promising middle ground by allowing for the integration of diverse inputs, including prior knowledge and image data, into a comprehensive analytical model. These models, trained with extensive resources, can be fine-tuned with relatively modest herbicide-specific datasets to address the challenges posed by limited data availability in the field.

Addressing these limitations requires ongoing research in advanced data processing methods, integration with other sensing technologies, the development of robust spectral indices tailored to detect herbicide effects, and advancements in sensor technology to improve spatial and spectral resolution while reducing costs. For instance, combining HSI with complementary technologies like chlorophyll fluorescence imaging can provide a more comprehensive understanding of herbicides’ effects on plants [1,16,17]. Developing and validating spectral indices that minimize sensitivity to environmental factors can enhance the robustness of assessments.

In conclusion, hyperspectral imaging represents a powerful tool in herbicide research within weed science, offering detailed spectral information that enhances the detection and analysis of herbicide effects. Its applications in monitoring herbicide drift, analyzing herbicide resistance, and aiding the development of new herbicides with novel modes of action contribute significantly to sustainable and effective weed management strategies. By integrating HSI with advanced analytical methods and machine learning algorithms, its potential can be further expanded, supporting the agricultural sector in overcoming existing herbicide challenges.