Wheat is the most extensively produced cereal in the world and one of the main direct sources of sustenance for humans. But wheat is always under threat from many diseases such as leaf rust, stem rust, powdery mildew, Septoria blotch and blight. These diseases are important hazards to productivity and grain quality with major financial losses and threats to food security if not adequately controlled (
Mahlein, 2016;
Terentev et al., 2023).
Traditional disease identification in wheat is primarily based on visual inspection of the field by the human eye. These are fast, on-the-spot analyses, but they are also inherently slow, inefficient and subjective or inconsistent. Furthermore, they are not scalable enough for large-scale agricultural applications, particularly in resource-limited regions (
Mahlein, 2016;
Wahabzada et al., 2016).
However, hyperspectral imaging (HSI) has been widely regarded as a robust, non-invasive remote sensing technology to monitor delicate plant physiological and biochemical responses
(Hilton et al., 2021; Mahlein, 2016;
Gogoi et al., 2018). Unlike conventional red-green-blue (RGB) photography that acquires information in just three spectral bands, HSI acquires data in hundreds of closely spaced, adjacent wavelength intervals across the image to enable the extraction of pixel-level spectra
(Hilton et al., 2021; Gogoi et al., 2018). The high resolution spectral data permits early and precise diagnosis of plant stress signals, sometimes before they can be seen by the naked eye (
Mahlein, 2016;
Xie et al., 2017; Terentev et al., 2023).
Several reviews on the applications of remote sensing and machine learning for plant disease detection have been conducted, but they have mainly considered broad crop categories or general methodological overviews, rather than specifically focusing on the wheat crop or systematically comparing classical and deep learning paradigms in an HSI framework
(Behmann et al., 2015; García-Vera et al., 2024;
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Gogoi et al., 2018; Gupta and Soni, 2022). There is no thorough and focused evaluation in the literature that consolidates HSI-based methodologies, benchmark datasets, sensor platforms and deployment problems for wheat disease detection.
The present review attempts to fill these gaps by targeting the following specific objectives: (i) to present the basic principles of hyperspectral imagery and pre-processing techniques, (ii) to review feature extraction strategies and machine/deep learning-based classification methods for wheat disease detection, (iii) to discuss available datasets and sensor platforms and (iv) to evaluate challenges in practical use and future prospects for scalable HSI deployments in field-scale wheat monitoring. To the best of authors’ knowledge, this is the first review that thoroughly integrates the above features especially in the context of HSI-based wheat disease detection.
Review methodology
The review was performed through systematic search of peer-reviewed literature from 2015 to 2025. The main databases consulted were Web of Science, Scopus, IEEE Xplore, Google Scholar and PubMed. The search was conducted using a combination of the following keywords: “hyperspectral imaging”, “wheat disease detection”, “precision agriculture”, “spectral imaging”, “deep learning crop disease”, “remote sensing wheat”, “plant disease classification” and “hyperspectral image classification”
(Behmann et al., 2015; García-Vera et al., 2024;
Hilton et al., 2021; Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Mahlein, 2016;
Rasti et al., 2018).
Studies were included in this review if they (i) used hyperspectral imaging as the main sensing modality, (ii) targeted wheat or other cereal crops that are phylogenetically close for disease detection or classification, (iii) provided quantitative performance metrics such as classification accuracy, F1-score, or overall accuracy and (iv) were published in peer-reviewed journals or reputable conference proceedings
(Abdulridha et al., 2023; Chang et al., 2024; Chossegros et al., 2025; Carvalho et al., 2025; Erdilo et al., 2025; Fedotov et al., 2025; Feng et al., 2020; Feng et al., 2022; Goyal et al., 2025; Haagsma et al., 2023; Hilton et al., 2021; Khan et al., 2021; Li et al., 2019; Li et al., 2020; Mahlein, 2016;
Mortensen et al., 2020; Rasti et al., 2018; Terentev et al., 2023; Ualiyeva et al., 2025; Wang et al., 2019; Xie et al., 2017; Xu et al., 2018; Zhang et al., 2019; Zhang et al., 2016; Zhong et al., 2022; Gogoi et al., 2018). Studies were excluded if they: (i) were based on RGB or conventional multispectral imaging only and did not include hyperspectral data; (ii) included other crops with no direct methodological implications for HSI-based disease detection; (iii) were conference abstracts, editorials or short communications that did not provide sufficient methodological detail; or (iv) were published before 2015, given the rapid pace of development of deep learning techniques in this area (
Kamilaris and Prenafeta-Boldú, 2018;
Krizhevsky et al., 2012; Li et al., 2019; Gupta and Soni, 2022).
After the application of these criteria, 36 studies were found and included in the final synthesis [1-36]. The studies were systematically classified and analysed along the following dimensions: (i) HSI acquisition platform (UAV-based, ground-based, or satellite-based); (ii) pre-processing and feature extraction strategies; (iii) classification methodology (classical machine learning vs. deep learning); (iv) wheat disease type investigated; and (v) dataset characteristics including spectral range, spatial resolution and availability
(Abdulridha et al., 2023; Chossegros et al., 2025; Feng et al., 2020; Feng et al., 2022; Haagsma et al., 2023; Li et al., 2019; Rasti et al., 2018; Zhang et al., 2019; Zhong et al., 2022; Gogoi et al., 2018; Gupta and Soni, 2022). This analytical approach sets the basis for the comparative discussions in later parts.
Fundamentals of hyperspectral imaging
Hyperspectral imaging (HSI) is an advanced technique of remote sensing and it involves the collection of rich spectral information across a broad range of electromagnetic spectrum
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018). Contrary to classical imaging systems that collect data in a few distinct bands (red, green and blue RGB), HSI captures hundreds of contiguous and narrow spectral bands for each spatial position within a scene
(Hilton et al., 2021; Rasti et al., 2018). This enables distinguishing between minor variations in material composition, structure and physiology associated with their respective specific spectral fingerprints (
Mahlein, 2016;
Wahabzada et al., 2016).
Hyperspectral image (HSI) The primary product of an HSI system is the hyperspectral datacube, or spectral image cube. This 3D image comprises two spatial dimensions (x, y) and one spectral dimension (λ), where each pixel in the lateral plane has a continuous reflectance spectrum over a defined wavelength region, which are usually found invisible-near infrared or shortwave infrared regions
(Hilton et al., 2021; Rasti et al., 2018; Zhang et al., 2019).
This high-spectral information offers accurate measurement of plant properties, including chlorophyll content, water status, pigments concentration and impaired disease physiologic activity-decades before visible symptoms occur (
Mahlein, 2016;
Wahabzada et al., 2016). In agriculture, especially crop disease detection, HSI has the ability to discriminate between healthy and diseased tissues at an early stage very efficiently used as a diagnostic tool
(Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023; Xie et al., 2017).
Spectral data cube
The hyperspectral data cube is an object whose organizational or structural concept is a three-dimensional array with two spatial dimensions (usually in reference to corresponding x and y coordinates of an image) and one spectral dimension, defined over hundreds of consecutive wavelength bands
(Hilton et al., 2021; Rasti et al., 2018). This setup enables hyperspectral imaging to acquire spatial patterns and full spectral peaks for each pixel of a scene
(Hilton et al., 2021; Rasti et al., 2018; Zhang et al., 2019). Unlike conventional imaging methods, which contain only surface color information, the spectral data contained in the cube can be used to identify subtle biochemical and physiological changes found in plant tissues (
Mahlein, 2016;
Wahabzada et al., 2016).
Certain areas of the electromagnetic spectrum are especially useful for plant health diagnosis
(Hilton et al., 2021; Mahlein, 2016). The visible (VIS: 400-700 nm) range records pigmentation changes, in particular of chlorophyll and carotenoids
(Hilton et al., 2021; Mahlein, 2016). In the visible to the near-infrared (vis-NIR: 400-700 nm) spectral region, the response is related to internal leaf structure and cell integrity, whereas in a longer-range of wavelengths (the shortwave infrared: SWIR 1100-2500 nm), it responds to water content, lignin and other biochemical components
(Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016). By measuring reflectance changes in these bands, hyperspectral systems allow stress-related physiological traits to be observed (for example origin of the stress) so that disease resistance can be assessed at anearlier stage than visual assessment
(Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023; Xie et al., 2017). This specific capacity to combine the spatial and spectral information makes hyperspectral imaging a powerful tool for accurate and early diagnosis of diseases in plants, which are necessary for precision agriculture techniques and sustainable crop management practices
(Behmann et al., 2015; Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016).
Imaging modalities
• Pushbroom or line-scan sensors are commonly used for airborne and satellite platforms
(Hilton et al., 2021; Mortensen et al., 2020). One spatial-line is recorded at a time by these sensors while the platform advances, thus incrementally composing the hyperspectral data cube
(Hilton et al., 2021). This technique has high spectral resolution and is particularly appropriate for applications demanding detailed mapping across large areas
(Hilton et al., 2021; Rasti et al., 2018). Pushbroom sensors are effective andcompatible with Unmanned Aerial Vehicles (UAVs) and thus have been widely used for the application of agricultural inspection, such as crop disease detection
(Abdulridha et al., 2023; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023). The sensor selection has to be done depending on the application scenarios including aerial borne, satellite-based or in a controlled laboratory setting
(Hilton et al., 2021; Mahlein, 2016).
• Whiskbroom (or point-scan) sensors move a rotating mirror to scan across the scene on a per-point or per-strip basis
(Hilton et al., 2021). Slower at capturing data compared to the pushbroom (whisk is slower than broom!), whiskbroom sensors can have a higher signal-to-noise ratio and are commonly found in satellite imaging
(Hilton et al., 2021; Rasti et al., 2018). But their mechanical intricacy has limited their applicability in UAV deployment
(Hilton et al., 2021; Mortensen et al., 2020).
• Snapshot (or area-scan) sensors record the complete hyperspectral data cube in a single exposure with no scanning movement
(Hilton et al., 2021). Such systems are extremely useful when operating in a dynamically system or real time system environment,
e.g. laboratory plant phenotyping or robotics system setting
(Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016). They are usually limited by lower spatial and spectral resolution as well as restricted to smaller wavelength range because of the sensor or hardware issues
(Hilton et al., 2021; Rasti et al., 2018).
For the design of hyperspectral imaging systems for concrete agricultural uses, these imaging modalities must be understood, as each has its consequences on the type of data acquired, processing requirements and deployment potential in field conditions (Fig 1).
HSI systems are generally based on one of three most common EPI acquisition modalities: pushbroom, whiskbroom, or snapshot mode
(Hilton et al., 2021; Rasti et al., 2018). Features of all these sensors, the pushbroom ones particularly, make them useful for agricultural monitoring due to their ability to efficiently collect high-resolution spectral information in aerial campaigns
(Hilton et al., 2021; Mortensen et al., 2020). When coupled with UAV platforms, pushbroom systems allow for fast and large area data acquisition of crop fields; thus they are highly effective in perceiving the spatial patterns of plant stress and disease at a high spectral resolution
(Abdulridha et al., 2023; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023).
Spectral signatures of wheat diseases
Wheat diseases induce a range of physiological and bio chemical changes in the plant, including chlorophyll content, water content, cell structure and pigment constitution alterations. These internal perturbations appear as small but measurable changes in the plant spectral reflectance profile within certain wavebands (
Mahlein, 2016;
Wahabzada et al., 2016). Hyperspectral imaging (HSI) systems can be utilized to acquire these spectral signatures accurately and provide a means for non-destructive detection, classification and mapping of different wheat diseases in both the latent as well as distinguishable stage(s) of infection
(Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023; Xie et al., 2017). By examining reflectance anomalies in targeted spectral bands, especially over the visible, near-infrared and shortwave-infrared regions of the spectrum, HSI allows for early warning and improved crop health management strategies
(Hilton et al., 2021; Mahlein, 2016;
Terentev et al., 2023; Ualiyeva et al., 2025).
Common diseases
a. Leaf rust (Puccinia triticina) - Reduction in chlorophyll gives rise to Causes red-orange pustules and hinder in photosynthesis
(Terentev et al., 2023; Xie et al., 2017).
b. Stem rust (P. graminis) - Affects tissue results in More severe, affects vascular tissues
(Abdulridha et al., 2023; Fedotov et al., 2025).
c. Powdery mildew (B. graminis) - Ends up as White powder on leaf surfaces
(Feng et al., 2022).
d. Septoria blotch (Z. tritici) - It results in necrotic lesions and reduced green leaf area
(Chossegros et al., 2025).
These changes cause:
• Decreased NIR reflectance (700-1000 nm) (
Mahlein, 2016;
Wahabzada et al., 2016).
• Increased VIS reflectance (400-700 nm) (
Mahlein, 2016;
Wahabzada et al., 2016).
• Alterations in SWIR water absorption bands (1450, 1950 nm)
(Hilton et al., 2021; Mahlein, 2016).
Spectral changes
The first requirement for classifying and analyzing data is to prepare the images through some preprocessing steps that improve the quality of hyperspectral image and facilitate the processing and interpretation. The raw HSI data is generally corrupted by different kinds of noise and artifacts such as sensor noise, non-uniform illumination, atmospheric distortion or redundant spectral information. These difficulties require pre- processing methods (spectral smoothing, radiometric calibration, geometric correction and illumination normalization) to remove the influence of environmental/environmental changes and instrument characteristics
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhang et al., 2019).
The wheat diseases trigger specific physiological stress responses which change the leaf pigment content and internal water distribution, that are known to arrest importantly on the plant’s spectral reflectance characteristics (
Mahlein, 2016;
Wahabzada et al., 2016). These changes are more significant in the Visible (VIS: 400–700 nm), Near- IR (NIR: 700-1000 nm) and Short-wavelength IR (SWIR: 1000-2500 nm) spectral ranges relevant for remote sensing for crop health monitoring
(Hilton et al., 2021; Mahlein, 2016).
Some of the important spectral variations in the diseased wheat leaves are (Fig 2).
•
Decreased NIR reflectance (700-1000 nm): Damage to the leaf cause SAR in the host by collapsed spongy mesophyll, broken intercellular air spaces leading to resolution of NIR since healthy tissue reflects strongly thanks to predominantly unbroken cell walls and coupling with internal organ tissue (
Mahlein, 2016;
Terentev et al., 2023; Wahabzada et al., 2016).
•
Increased VIS reflectance (400-700 nm): Pathogen-induced degradation of chlorophyll decreases light absorbance in the blue (450 nm) and red (670 nm) spectra, leading to higher reflectance. This coloration response is frequently expressed as yellowing or chlorosis that can be observed with the naked eye in infected leaves (
Mahlein, 2016;
Terentev et al., 2023; Xie et al., 2017).
• Changes to the strong water absorbance bands around1450 nm and 1950nm in SWIR
(Hilton et al., 2021; Mahlein, 2016).
• Infection-mediated changes in leaf water content are reflected as distinct shifts in these absorption features. Cut off values for the water absorption depth may be used as an indication of dehydration when reduced, or tissue necrosis or water retention resulting from exacerbation of a disorder when increased (
Mahlein, 2016;
Wahabzada et al., 2016).
• These wavelength-dependent reflectance anomalies can be exploited by HSI sensors to detect and distinguish healthy and diseased plant regions well before visible symptoms appear
(Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023; Ualiyeva et al., 2025; Xie et al., 2017).
Preprocessing and feature extraction
Moreover, as hyperspectral images are high-dimensional (sometimes with hundreds of spectral bands), many dimensionality reduction methods have been applied to reduce the size of these data while retaining most informative features and avoiding overfitting
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). Common methodologies: The Planetary Data System (PDS) has a list of classification algorithms avail- able for planetary image analysis that include dimensionality reduction methods as PCA and components transformation by Minimum Noise Fraction (MNF), band selection based on statistical differences or obtaining features following the separability rule
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhang et al., 2019).
The preprocessing part is an important aspect in the HSI pipeline as it directly affects the performance of classification algorithm applied to subsequent classified disease detection on wheat
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhang et al., 2019; Zhong et al., 2022).
Radiometric calibration
Radiometric calibration is a mandatory pre-processing step in hyperspectral image analysis, which converts raw sensor digital numbers (DN) to physically relevant reflectance values
(Hilton et al., 2021; Li et al., 2019; Rasti et al., 2018). Raw HSI data is inevitably influenced by sensor-specific responses, dark current noise and non-uniform illumination conditions
(Hilton et al., 2021; Rasti et al., 2018). If not calibrated, these artifacts cause systematic inaccuracies that impair the reliability of future spectral analysis and disease classification
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). In wheat disease diagnosis, precise recovery of reflectance is crucial, since the spectral changes generated by the illness are frequently modest and can easily be disguised by uncorrected sensor errors (
Mahlein, 2016;
Terentev et al., 2023; Ualiyeva et al., 2025; Xie et al., 2017). The calibration is normally done with a white Spectral on reference panel of known reflectance qualities, acquired under the same illumination circumstances as the target picture
(Hilton et al., 2021; Rasti et al., 2018).
Where,
R(λ) = Calibrated reflectance at wavelength λ.
DN(λ) = Raw digital number recorded by the sensor.
D(λ) = Dark current reference acquired with the lens cap on.
W(λ) = White reference panel measurement.
This normalization removes distortions produced by the light and the non-uniformity of the sensor response so that the final reflectance values are similar across different acquisition sessions, platforms and environmental conditions
(Hilton et al., 2021; Li et al., 2019; Rasti et al., 2018).
Dimensionality reduction - Principal component analysis (PCA)
Hyperspectral captures generally have hundreds of spectral bands, of which a large number are highly correlated and contain redundant information
(Li et al., 2019; Rasti et al., 2018). This high dimensionality incurs considerable processing hurdles and may cause the Hughes phenomenon where the performance of the classifier decreases with the increasing number of features with respect to the number of training examples
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). Principal Component Analysis (PCA) is a linear unsupervised dimensionality reduction method, which solves this problem by projecting the original high-dimensional spectral data onto a lower-dimensional subspace spanned by the directions of highest variance
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). PCA is a frequently used preprocessing approach for HSI based wheat disease diagnosis, to retain the most useful spectral components and remove the noise-dominated bands, which reduces the computing cost and improves the downstream classification performance
(Behmann et al., 2015; Fedotov et al., 2025; Li et al., 2019; Rasti et al., 2018; Ualiyeva et al., 2025).
Y = XW ...(2)
Let X ∈ Rn×d be the original hyperspectral data matrix with n pixels and d spectral bands.
W ∈ Rd×x be the matrix of k primary eigen vectors obtained from the covariance matrix:
and Y ∈ R
n×k is the dimensionality-reduced representation containing the k components that explain the greatest share of overall spectral variation. The choice of k is usually chosen by the cumulative explained variance criterion, often set at 95-99%
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018).
Dimensionality reduction - Linear discriminant analysis (LDA)
PCA is an unsupervised method which discovers directions of maximum variance, while LDA is a supervised dimensionality reduction method which looks for projection paths that maximize the separability among a set of pre-defined classes
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). LDA is advantageous over PCA for hyperspectral wheat disease classification that aims to differentiate between healthy tissue and various disease categories (e.g., leaf rust, stem rust and powdery mildew), because it uses the class label information explicitly to guide the projection
(Behmann et al., 2015; Li et al., 2019;
Rasti et al., 2018). It has been proven that LDA is able to generate more compact and discriminative feature representation, particularly in the case where the number of spectral bands is substantially bigger than the number of training samples, which is usually the case for agricultural HSI datasets
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhong et al., 2022).
Where,
SB is the between-class scatter matrix, defined as:
and is the within-class scatter matrix, defined as:
μ
c = Mean of class c.
μ = Global mean.
n
c = Number of samples in class c.
X
c = Set of samples belonging to class c.
The ideal projection maximizes the between-class separability and minimizes the within-class scatter simultaneously, which results in a highly discriminative low-dimensional subspace for multi-class wheat disease classification
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhong et al., 2022).
Feature extraction-minimum noise fraction (MNF)
The Minimum Noise Fraction (MNF) transform is a two-stage orthogonal transformation developed for hyperspectral data, which separates spectrally relevant signal components from noise-dominated band
(Li et al., 2019; Rasti et al., 2018). MNF differs from PCA in that it first estimates and removes the noise covariance structure and then determines the principal components of the underlying signal, rather than ordering components only on the basis of variance
(Li et al., 2019; Rasti et al., 2018). This makes MNF especially suited for airborne and UAV based HSI data in agricultural applications where sensor noise, atmospheric scattering and platform induced vibration can dramatically impact spectrum quality
(Abdulridha et al., 2023; Feng et al., 2020; Haagsma et al., 2023; Li et al., 2019; Mortensen et al., 2020; Rasti et al., 2018). In wheat disease detection workflows, MNF is usually applied as the initial preprocessing step and after that, only the leading MNF components, which have the highest signal-to-noise ratio, are kept for further classification
(Fedotov et al., 2025; Li et al., 2019; Rasti et al., 2018; Ualiyeva et al., 2025).
∑
singal x = Covariance matrix of the underlying spectral noise.
∑
noise x = Calculated noise covariance matrix.
The MNF transform consists of two successive PCA steps: the first PCA decorrelates the noise components and normalizes the noise to unit variance; the second PCA is performed on the noise-normalized data
(Li et al., 2019; Rasti et al., 2018). The MNF discards the components with eigenvalues near to one as being dominated by noise. The components with much greater eigenvalues include useful spectral information for discriminating between diseases
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018).
Vegetation index - normalized difference vegetation index (NDVI)
One of the most commonly used spectral indices in remote sensing to evaluate vegetation vigour and detect plant stress is the Normalized Difference Vegetation Index (NDVI)
(Hilton et al., 2021; Mahlein, 2016). NDVI utilizes the high contrast between the high reflectance of healthy vegetation in the near-infrared region (due to the interior structure of the leaf cells) and the strong absorption in the red area (due to photosynthesis powered by chlorophyll)
(Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016). In sick wheat, pathogen-induced chlorophyll degradation reduces the absorption of red light, while damaging cellular integrity and reducing NIR reflectance, hence causing a detectable reduction in NDVI values (
Mahlein, 2016;
Terentev et al., 2023; Xie et al., 2017). This renders NDVI a sensitive and computationally effective early indication of disease-induced stress, notably for leaf rust, Septoria blotch and powdery mildew where chlorophyll loss is a key physiological consequence
(Chossegros et al., 2025; Feng et al., 2022; Mahlein, 2016;
Terentev et al., 2023; Xie et al., 2017).
Where,
R
NIR and R
red = Reflectance values at near-infrared (about 800 nm) and red (around 670 nm) wavelengths, respectively.
NDVI values vary from -1 to +1 where values around +1 indicate dense healthy vegetation and decreasing values imply increasing degrees of stress or disease severity
(Hilton et al., 2021; Mahlein, 2016). In hyperspectral data sets, NDVI is more accurately derived than in multispectral systems due to the presence of narrowly defined bands at ideal wavelengths
(Hilton et al., 2021; Rasti et al., 2018).
Vegetation index - Photochemical reflectance index (PRI)
The photochemical reflectance index (PRI) is a narrow-band spectral index developed to follow dynamic changes in carotenoid pigment composition, specifically the xanthophyll cycle, which is directly related to photosynthetic light-use efficiency
(Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016). The PRI is extremely responsive to early-stage physiological stress prior to the development of observable symptoms, suggesting that the PRI is a useful early warning indication in HSI-based wheat disease monitoring
(Feng et al., 2020; Haagsma et al., 2023; Mahlein, 2016;
Mortensen et al., 2020; Terentev et al., 2023). As the ratio of carotenoids to chlorophyll changes rapidly with pathogen infection, PRI can identify disease-induced metabolic disruption before any structural damage to the leaf is visible, therefore providing a diagnostic advantage over morphologically based indices such as NDVI (
Mahlein, 2016;
Wahabzada et al., 2016).
Where,
R
531 and R
570 = Reflectance at 531 nm and 570 nm, respectively.
Reduction of PRI values from healthy baseline measurements are indicative of compromised photosynthetic efficiency and carotenoid pool alterations associated with early illness stress (
Mahlein, 2016;
Wahabzada et al., 2016). PRI has been shown sensitive to stripe rust and powdery mildew infection in wheat at latent stages, a few days before apparent chlorotic or necrotic signs developed
(Feng et al., 2022; Feng et al., 2020; Xie et al., 2017).
Vegetation index - Soil-adjusted vegetation index (SAVI)
The Soil-Adjusted Vegetation Index (SAVI) is a variation of NDVI that tries to reduce the influence of soil background light reflectance in the spectral assessment of vegetation
(Hilton et al., 2021; Mahlein, 2016). Bare soil reflectance in wheat fields, particularly during the early growth stages when canopy cover is incomplete, can dramatically distort spectral indices and impair the accuracy of disease classification models
(Hilton et al., 2021; Mahlein, 2016;
Mortensen et al., 2020). However, SAVI overcomes this problem by introducing a soil brightness adjustment factor, which compensates for the differences in soil reflectance under sparse canopy situation, thereby making it more robust than NDVI for disease detection in partially formed or heterogeneous wheat fields
(Hilton et al., 2021; Mahlein, 2016;
Mortensen et al., 2020).
Where,
R
NIR and R
Red = Reflectance values in the near-infrared and red spectral bands respectively.
L is the soil brightness correction factor which is empirically calibrated to 0.5 for moderate vegetation densities
(Hilton et al., 2021; Mahlein, 2016). When L=0, SAVI becomes NDVI and for larger values of L the contribution of soil background is more and more suppressed
(Hilton et al., 2021; Mahlein, 2016). SAVI is particularly useful in semi-arid wheat producing regions and early season monitoring, where soil exposure is high and traditional NDVI estimations often overestimate vegetation stress
(Hilton et al., 2021; Mahlein, 2016;
Mortensen et al., 2020).
Vegetation index - Disease water stress index (DSWI)
The Disease Water Stress Index (DSWI) is a multi-band spectral index developed to characterize simultaneous changes of the leaf water content and chlorophyll concentration, both affected by the pathogen infection in wheat (
Mahlein, 2016;
Terentev et al., 2023; Wahabzada et al., 2016). Unlike single-parameter indices, DSWI combines reflectance information from four different wavelength regions that are simultaneously sensitive to water absorption, chlorophyll activity and cell structure integrity, thereby providing a more comprehensive characterization of disease-induced physiological stress
(Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016). The DSWI is especially useful in separating disease-related stress from abiotic stressors such as drought, as the spectral band combination uniquely captures pathogen-related tissue degradation patterns that differ from water-deficit responses (
Mahlein, 2016;
Mortensen et al., 2020; Terentev et al., 2023; Wahabzada et al., 2016).
Where,
R
802, R
547, R
1657 and R
682 = Reflectance values at 802 nm, 547 nm, 1657 nm and 682 nm respectively.
The numerator includes NIR reflectance related to the integrity of leaf cell structure and green reflectance related to carotenoid activity and the denominator includes SWIR water absorption and red chlorophyll absorption properties
(Hilton et al., 2021; Mahlein, 2016;
Wahabzada et al., 2016). Lower DSWI scores are indicative of concurrent stress on water and chlorophyll, a specific spectral signature of active fungal infection in wheat leaves
(Feng et al., 2022; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023; Ualiyeva et al., 2025).
Machine learning and deep learning approaches
Classical classifiers, such as Support Vector Machines (SVM), Random Forests (RF) and k-Nearest Neighbors (k- NN), are extensively used in his work have been adopted for hyperspectral image classification due to their robustness and handling of high dimensional data, which need of hand-crafted feature extraction process and domain specific preprocessing steps
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). Although such methods have proven useful to some extent for certain cases, they often face challenges in dealing with the intrinsic high-dimensional, spectral redundancy and spatial variation properties of HS image
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhang et al., 2019).
However, deep learning methods and especially Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Transformer-based architectures have demonstrated superior classification accuracy and scalability in recent years (
Kamilaris and Prenafeta-Boldú, 2018;
Krizhevsky et al., 2012; Li et al., 2019; Xu et al., 2018; Zhong et al., 2022). Such models can learn hierarchical spectral–spatial features from raw hyperspectral data automatically and thereby eliminate manually engineered features
(Guo et al., 2019; Hang et al., 2019; Li et al., 2019; Li et al., 2020; Wang et al., 2019; Xu et al., 2018; Zhang et al., 2016; Zhong et al., 2022). Besides, deep learning models are more efficient for the generalization of different field conditions, sensor noises and complicated disease shapes could be better adapted to large-scale agricultural monitoring frameworks and real time decision support systems
(Chang et al., 2024; Erdilo et al., 2025; Fedotov et al., 2025; García-Vera et al., 2024;
Goyal et al., 2025; Khan et al., 2021; Li et al., 2019).
Classical models
Among the classical machine learning methods, SVM, RF, k-NN and DT have been widely applied to analyze hyper- spectral imagery classification tasks (especially in agriculture areas such as disease detection for wheat crops)
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). The interpretations and the computational demands are lower with these approaches, which yield competitive performance for small to medium sized datasets
(Behmann et al., 2015; Li et al., 2019). However, most of these models rely on manual feature extraction.
Feature extraction
This is also known as point cloud sub-sampling a process of choosing or synthesizing appropriate spectral or spatial features from raw HSIs based on domain knowledge
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). Also, the curse of dimensionality is critical problem since hyperspectral images have hundreds of spectral bands, most of which are correlated and noisy
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018). Consequently, These algorithms are usually optimal only when reducing the dimensionality of data, since using either PCA, LDA considered the version feature selection approaches will reduce the complexity of the feature-space and improve class discrimination
(Behmann et al., 2015; Li et al., 2019; Rasti et al., 2018; Zhong et al., 2022).
Despite these limitations, conventional machine learning models are still applicable when the labeled datasets are insufficient or interpretability and model explain ability matter
(Behmann et al., 2015; García-Vera et al., 2024;
Li et al., 2019).
The discriminant function of a linear SVM is:
SVM Decision Funtion: f(x) = w{T} X + b
B. Deep learning models
The advent of deep learning models including, among others, Convolutional Neural Networks (CNNs), Recurrent Neural Network (RNNs) and Transformer-based models has greatly boosted hyperspectral image classification by enabling feature hierarchies to be automatically learned from raw input data (
Kamilaris and Prenafeta-Boldú, 2018;
Krizhevsky et al., 2012; Li et al., 2019; Xu et al., 2018; Zhong et al., 2022). These methods alleviate the heavy dependence of manual feature engineering, since these end-to-end approaches harness large training datasets to learn discriminative spectral–spatial features
(Guo et al., 2019; Li et al., 2019; Li et al., 2020; Wang et al., 2019; Xu et al., 2018; Zhang et al., 2016; Zhong et al., 2022; Metagar and Walikar, 2024).
3D Convolutional Neural Networks (3D-CNNs) have shown remarkable performance in hyperspectral image analysis due to their capability of capturing both spectral and spatial information when extracting the features
(Li et al., 2019; Xu et al., 2018; Zhong et al., 2022) (Fig 3).
3D-CNNs, by utilization of convolutional filters along the three dimensions of the hyperspectral cube, capture local spectral signature and spatial context simultaneously and can achieve significantly better classification accuracy in particular for fine-grained disease discrimination in crops like wheat
(Fedotov et al., 2025; Guo et al., 2019; Khan et al., 2021; Li et al., 2019; Xu et al., 2018; Zhong et al., 2022; Metagar and Walikar, 2024).
Furthermore, RNNs and its variants such as Long-short term memory (LSTM) networks have also been employed on hyperspectral data by considering spectral bands as sequence input, which allows to learn the context relationship between wavelength
(Li et al., 2019; Li et al., 2020; Zhong et al., 2022). The latter are particularly useful, when spectral signatures has inherent order due to physiological differences in plant health
(Li et al., 2019; Li et al., 2020).
In recent years, Vision Transformers (ViTs) and attention has been applied to hyperspectral classification. ViTs being taught to model global relationships over the spectral–spatial domain are capable of providing better context-awareness and feature discrimination
(Li et al., 2019; Zhong et al., 2022). Attention mechanisms allow the model to focus selectively to those parts of input instances that are more informative, which is useful, in general, for complex problems such as a disease having subtle or uneven spread of features
(Li et al., 2019; Zhong et al., 2022).
Together, these deep learning models present the most advanced hyper-spectral image classification results with improved accuracy, robustness and scalability compared to traditional ML algorithms.
Applications and case studies
There are several cases of real-world hyperspectral imaging (HSI) applications in wheat disease detection that include different systems and scales, which are designed to address the underpinning agricultural monitoring requirements
(Abdulridha et al., 2023; Chossegros et al., 2025; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023). These platforms vary on spatial resolution, spectral accuracy, flexibility in operations and cost of deployment
(Hilton et al., 2021; Mahlein, 2016).
•
UAV-mounted systems: Unmanned Aerial Vehicles (UAVs) carrying push broom or snapshot hyperspectral imaging sensors are increasingly used for field scale disease monitoring
(Abdulridha et al., 2023; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Terentev et al., 2023). These systems collect multispectral-spatial data at high resolution over specific areas of interest for early disease detection and site-specific application of agrochemicals
(Abdulridha et al., 2023; Feng et al., 2020; Mortensen et al., 2020). The UAV-based HSI method is particularly useful for surfacing disease hotspots and guiding site-specific chemical spraying by pesticides or a localized irrigation
(Abdulridha et al., 2023; Feng et al., 2020; Mortensen et al., 2020; Terentev et al., 2023) (Fig 4).
•
Ground-based phenotyping platforms: Operated by tractors or stationery rigs, such platforms are generally utilized in a non-field validated experimental and research environment
(Chossegros et al., 2025; Haagsma et al., 2023; Mahlein, 2016;
Wahabzada et al., 2016). Provide high-fidelity, short-range hyper- spectral measurements and thus are well suited for examining subtle dynamic shifts in physiology accompanying disease evolution
(Chossegros et al., 2025; Haagsma et al., 2023; Mahlein, 2016;
Wahabzada et al., 2016). Such systems are essential in validating remote sensing algorithms and for the development of wheat varieties with resistance to disease
(Chossegros et al., 2025; Haagsma et al., 2023; Mahlein, 2016).
•
Satellite-based hyperspectral sensors: Hyperspectral satellite missions (
e.g., PRISMA, EnMAP ) are at the moment spectrally limited with respect to UAV- and ground- based instruments, but they could potentially revolutionize large-scale agricultural monitoring
(Hilton et al., 2021; Rasti et al., 2018). Such systems are useful for regional or national surveillance operations and allow the monitoring of disease events across large-scale geographical areas but at a cost in time frequency of detection (and loss of sensor quality)
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018).
In recent times, it has been demonstrated that the joint use of hyperspectral imaging and deep learning models like 3D Convolutional Neural Networks (3D-CNNs) or Vision Transformer models contribute to an early diagnosis of wheat diseases
(Audebert et al., 2018; Chang et al., 2024; Chossegros et al., 2025; Erdilo et al., 2025; Fedotov et al., 2025; Khan et al., 2021; Li et al., 2019; Ualiyeva et al., 2025; Zhong et al., 2022). In a number of studies, such methods have reached diagnostic accuracy rates above 90 per cent even when symptoms were not yet evident by eye
(Chang et al., 2024; Chossegros et al., 2025; Erdilo et al., 2025; Fedotov et al., 2025; Khan et al., 2021; Li et al., 2019; Ualiyeva et al., 2025; Zhong et al., 2022). This highlights the revolutionary capability of HSI as a digital- agriculture-based sustainable crop protection tool (
García-Vera et al., 2024;
Hilton et al., 2021; Mahlein, 2016).
Comparison tables
The following tables summarizes the important aspects of hyperspectral imaging technique for wheat disease detection. Table 1 lists key wheat diseases, associated pathogens, the most informative spectral bands and their spectral effects. Table 2 compares commonly used classification models in terms of input type, accuracy and implementation characteristics across classical and deep learning methods.
A comparative summary of representative hyperspectral imaging and machine-learning approaches for wheat disease detection, covering the diseases investigated, sensing platforms, spectral ranges, modelling methods, reported performance, dataset characteristics, and key limitations, is presented in Table 3.
Major shortcomings of existing hsi systems
Spectral-spatial resolution trade-off
This inverse link between spectral resolution, spatial resolution and acquisition speed is a fundamental engineering limitation in HSI system design
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018). Pushbroom sensors on UAV platforms have excellent spectral resolution but are limited in the ability to sustain high spatial resolution over broad field regions concurrently
(Abdulridha et al., 2023; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020; Rasti et al., 2018). Snapshot sensors increase acquisition speed, albeit at the cost of spectral and spatial resolution
(Hilton et al., 2021; Rasti et al., 2018). None of the commercial HSI platforms available on the market today fully solve this three-way trade-off and the best sensor combination is still application-specific and depends on the disease type, the stage of crop growth and the field scale
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018).
Calibration instability under field conditions
Compared to controlled laboratory settings, radiometric calibration is far more difficult in the dynamic field environment due to fluctuations in sun irradiation, solar zenith angle and crop movement generated by wind during acquisition
(Hilton et al., 2021; Li et al., 2019; Rasti et al., 2018). The majority of the studies that were evaluated execute the calibration using a single white reference panel acquired at the beginning of the flight mission and do not account for the change in illumination during the data collection
(Abdulridha et al., 2023; Feng et al., 2020; Haagsma et al., 2023; Mortensen et al., 2020). This calibration instability causes systematic spectral errors that could lower classification accuracy by 3-8 percentage points in realistic field situations relative to laboratory benchmarks
(Haagsma et al., 2023; Mahlein, 2016;
Rasti et al., 2018). Although the performance of the model is significantly affected by calibration uncertainty, it is rarely quantified or explicitly mentioned in the reviewed literature
(Hilton et al., 2021; Li et al., 2019; Rasti et al., 2018).
Lack of standardised benchmark datasets
The present HSI wheat disease research is critically limited by the lack of large-scale, publicly available and standardized benchmark datasets that would allow for a credible cross-study comparison
(Behmann et al., 2015; García-Vera et al., 2024;
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019). Most of the evaluated research create their own datasets for specific local conditions, wheat varieties, disease strains and growth phases
(Abdulridha et al., 2023; Chossegros et al., 2025; Fedotov et al., 2025; Feng et al., 2020; Feng et al., 2022; Goyal et al., 2025; Haagsma et al., 2023; Khan et al., 2021; Mortensen et al., 2020; Terentev et al., 2023; Ualiyeva et al., 2025). The dataset sizes range from less than 500 to several thousand tagged samples
(Abdulridha et al., 2023; Chossegros et al., 2025; Fedotov et al., 2025; Khan et al., 2021; Mortensen et al., 2020; Terentev et al., 2023). This lack of diversity greatly limits the generalisability of trained models and makes it impossible to know if accuracy improvements claimed are due to actual methodological progress or benign dataset factors (
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Zhong et al., 2022). Perhaps the most important contribution that might be made to speed real progress in this sector is the construction of a community-standard HSI wheat disease benchmark dataset
(Behmann et al., 2015; Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019).
Computational infrastructure requirements
The computational needs of deep learning-based hyperspectral image (HSI) processing pipelines are a major practical challenge for field deployment, since the processing of a single UAV flight mission producing several terabytes of hyperspectral data calls for large GPU computing resources
(Li et al., 2019; Xu et al., 2018; Zhong et al., 2022). Such processing delays are fundamentally at odds with the real-time or near-real-time decision support needs of operational agriculture (
García-Vera et al., 2024;
Li et al., 2019; Mahlein, 2016). Although considerable research on lightweight model architectures and edge computing solutions is underway, there is no commercial HSI-based wheat disease detection system available today at the speed, cost and reliability necessary for routine adoption at the farm scale
(Chang et al., 2024; Li et al., 2019; Xu et al., 2018; Zhong et al., 2022). This gap between research-grade system performance and commercial deployment readiness remains a major open challenge in the area (
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Zhong et al., 2022).
A rigorous comparison of deep learning models
2D CNN vs. 3D CNN
While 2D CNNs are computationally efficient, they cannot explicitly describe the sequential spectral correlations between consecutive wavelength bands. 3D-CNNs employ convolutional filters over both spatial and spectral dimensions simultaneously
(Li et al., 2019; Xu et al., 2018; Zhong et al., 2022). The evaluated papers consistently report that 3D-CNNs outperform 2D-CNNs by 3-6 percentage points in terms of overall accuracy, but at a computational cost of around 4-8 times more parameters, rendering them not suitable for edge deployment
(Fedotov et al., 2025; Khan et al., 2021; Li et al., 2019; Xu et al., 2018; Zhong et al., 2022).
CNN or vision transformer (ViT)
CNNs are good at modeling the local spectral-spatial linkages through hierarchical convolution
(Li et al., 2019; Xu et al., 2018; Zhong et al., 2022). Vision Transformers explore the long-range global dependencies over the full spectral-spatial domain using self-attention processes at the same time
(Chang et al., 2024; Li et al., 2019; Zhong et al., 2022). ViT-based architectures produce the greatest reported accuracies of 93-98%, exceeding 3D-CNNs by 2-4 percentage points, but require considerably larger labelled training datasets and may perform worse on the short agricultural HSI datasets that are typical of most wheat disease investigations
(Chang et al., 2024; Li et al., 2019; Zhong et al., 2022).
RNN and LSTM in spectral sequence modeling
RNNs and LSTMs model hyperspectral data as ordered spectral sequences and utilize the continuity of reflectance across neighbouring wavelength bands to capture slow spectral changes associated with disease development
(Li et al., 2019; Li et al., 2020; Zhong et al., 2022). However, their incapability to include spatial context restricts the performance [20,21]. Hybrid designs of LSTM spectral modelling and CNN spatial feature extraction have shown a more practical accuracy range of 90-95%
(Khan et al., 2021; Li et al., 2019; Li et al., 2020; Xu et al., 2018).
Hybrid and ensemble architectures
The latest hybrid architectures that merge CNN spatial feature extraction with Transformer based global spectrum attention have achieved the highest reported accuracies, but at a considerable expense of architectural complexity, training data needs and inference time
(Chang et al., 2024; Goyal et al., 2025; Li et al., 2019; Zhong et al., 2022). Ensemble approaches integrating many independently trained classifiers provide a pragmatic alternative with competitive accuracy and less need on huge labelled datasets making them well suited for real world agricultural deployment
(Abdulridha et al., 2023; Feng et al., 2022; Terentev et al., 2023).
Practical use for farmers and commercialization
Economic availability
The cost of commercial UAV-mounted HSI sensors ranges from USD 20,000 to USD 80,000, which is prohibitive for most wheat producers globally, especially smallholder farmers in underdeveloped nations
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018). Technology development or service-based deployment methods for farmers to receive disease detection via agricultural cooperatives or precision agriculture providers will be necessary to greatly reduce hardware costs and achieve commercial viability (
García-Vera et al., 2024;
Mahlein, 2016).
Ease of use and operation by farmers
In addition to cost issues, effective HSI data collecting needs technical skills in flight planning, sensor calibration and data quality control that cannot be assumed of normal farm operators
(Haagsma et al., 2023; Khan et al., 2021; Mahlein, 2016). Widespread adoption of end to end HSI systems requires automated calibration and one-click processing pipelines, with simple map-based outputs that can be directly used in spray prescription or harvesting decision making workflows without the need for specialist expertise (
García-Vera et al., 2024;
Khan et al., 2021; Li et al., 2019).
Integration of decision support
The practical significance of disease detection using HSI is not only in classification accuracy, but also in how the results are transformed into economically effective and timely management decisions, coupled to yield loss models, treatment cost thresholds and weather forecast data
(Hilton et al., 2021; Li et al., 2019; Rasti et al., 2018). Without this decision support integration layer, even very accurate disease maps have limited operational utility for farmers who need to turn spectral data into specific field management activities under real-world constraints (
García-Vera et al., 2024;
Li et al., 2019; Mahlein, 2016).
Validation and regulatory requirements
Before HSI-based disease detection systems can be used commercially, independent validation across a range of geographic regions, wheat varieties, disease strains and seasons is required, a requirement that has not been comprehensively addressed in the literature to date
(Behmann et al., 2015; García-Vera et al., 2024;
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019). The current research landscape is mainly absent this regulatory dimension which is a crucial precondition for responsible commercial implementation, defined by internationally accepted validation processes and minimum performance criteria (
García-Vera et al., 2024;
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019).
Scalability with data as a service models
A viable route to economic feasibility is through federated, data-as-a-service models where HSI disease detection is provided by satellite or aerial survey services combined with cloud-based AI processing platforms, removing the need for farm-level hardware investment (
García-Vera et al., 2024;
Hilton et al., 2021; Mahlein, 2016). Satellite-based HSI platforms, such as PRISMA and EnMAP, together with automated deep learning pipelines, can provide cost-effective regional disease surveillance at the requisite scale for national food security monitoring
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018; Zhong et al., 2022).
Challenges and future work
The application of hyperspectral imaging (HSI) technology in real field environments of agriculture is facing a number of technical or practical challenges especially wheat disease detection, which has many potential applications
(Behmann et al., 2015; García-Vera et al., 2024;
Hilton et al., 2021; Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Mahlein, 2016).
•
High dimensionality: Hyperspectral data usually have several hundreds of contiguous, very narrow spectral bands that can cause the Hughes phenomenon/the curse of dimensionality, which is known to impact classifier performance and burden the computational cost with increasing dimensions in the absence of sufficient training samples (
Behmann et al., 2015;
Li et al., 2019; Rasti et al., 2018; Zhang et al., 2019; Zhong et al., 2022).
•
High sensor cost: Hyperspectral cameras are still expensive, especially for smallholder farmers or if the system is to be deployed in many fields. Furthermore, there is also an additional cost in requiring a systems solution for stable calibration, high spectral fidelity and ruggedization
(Hilton et al., 2021; Mahlein, 2016;
Rasti et al., 2018).
•
Environmental variability: Factors such as differences in illumination, canopy structure, soil reflectance and atmospheric conditions add to spectral noise and alter reflectance values leading to inconsistent disease classification under realistic field based environment
(Haagsma et al., 2023; Hilton et al., 2021; Li et al., 2019; Mahlein, 2016;
Rasti et al., 2018).
•
Limited labeled datasets: To train models accurately, large annotated datasets of healthy and diseased wheat under a variety of environments at different stages need to be developed. This lack of open and standardized hyperspectral datasets is seriously hampering the scalability and transferability of machine learning-based models
(Behmann et al., 2015; García-Vera et al., 2024;
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Ualiyeva et al., 2025; Zhong et al., 2022; Gupta and Soni, 2022).
In order to manage these challenges, the research is looking at few of many interesting future directions:
•
Edge computing and iot integration: These lightweight, energy-efficient models can be deployed on edge devices (
e.g., drones, field sensors) to enable real- time disease detection so that reliance on cloud-based processing is minimized and responsiveness in precision agriculture is improved (
García-Vera et al., 2024;
Li et al., 2019; Mahlein, 2016;
Xu et al., 2018; Zhong et al., 2022).
•
Federated learning: Federated learning as an approach to decentralized model training using data from multiple farms or regions without having to centralize the raw data and emerging solutions for privacy-preserving and scalability in hyperspectral (HS) data modeling under distributed setting, particularly when data sharing is limited (
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Zhong et al., 2022).
•
Explainable AI (XAI): Because DL models are frequently referred to as black boxes, XAI techniques are currently being developed to explain model decisions, find important spectral bands or regions of interest (ROIs) and improve the trustworthiness of the method among agronomists and decision makers (
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Li et al., 2020; Zhong et al., 2022).
•
Multisensor fusion: Combining HSI with other modalities such as LiDAR, thermal imaging, or RGB cam eras also offers more complex data representation and increases the robustness of disease detection models to various field conditions
(Chang et al., 2024; García-Vera et al., 2024;
Li et al., 2019; Mahlein, 2016;
Mortensen et al., 2020).
These guidelines making hyperspectral sensing more understandable, interpretable and field-deployable towards accelerated adoption in smart agriculture and early disease warning systems (
García-Vera et al., 2024;
Kamilaris and Prenafeta-Boldú, 2018;
Li et al., 2019; Mahlein, 2016).
