Source-linked AI summary
Deep Learning Techniques for Hyperspectral Image Analysis in Agriculture: A Review
Mohamed Fadhlallah Guerri, Cosimo Distante, Paolo Spagnolo, Fares Bougourzi, Abdelmalik Taleb-Ahmed
TL;DR
Agricultural HSI analysis must address redundant spectral bands, limited labeled samples, and complex spatial-spectral relationships. This review synthesizes HSI fundamentals, datasets, deep-learning approaches, agricultural applications, and associated limitations. It concludes that HSI and deep learning support a broad range of agricultural analysis and quality, yield, and safety applications, while challenges remain in data redundancy, sampling, acquisition, and task-specific strategy selection.
Problem
Agricultural HSI analysis is constrained by redundant spectral bands, limited labeled samples, complex spatial-spectral relationships, and practical acquisition challenges.
Method
The paper reviews HSI technology, public agricultural datasets, deep-learning techniques, and agricultural applications, including active learning and transfer learning.
Results
The review comprehensively synthesizes HSI methods, research gaps, achievements, classification research, and agricultural applications.
Takeaways & Limitations
HSI combined with deep learning provides a basis for agricultural smart systems and applications involving crop yield, quality, and safety.
Takeaways & Limitations
HSI analysis remains bounded by redundant spectral information, limited labeled samples, acquisition constraints, and task-dependent active-learning query strategies.
Abstract
from arXiv · showhide
In the recent years, hyperspectral imaging (HSI) has gained considerably popularity among computer vision researchers for its potential in solving remote sensing problems, especially in agriculture field. However, HSI classification is a complex task due to the high redundancy of spectral bands, limited training samples, and non-linear relationship between spatial position and spectral bands. Fortunately, deep learning techniques have shown promising results in HSI analysis. This literature review explores recent applications of deep learning approaches such as Autoencoders, Convolutional Neural Networks (1D, 2D, and 3D), Recurrent Neural Networks, Deep Belief Networks, and Generative Adversarial Networks in agriculture. The performance of these approaches has been evaluated and discussed on well-known land cover datasets including Indian Pines, Salinas Valley, and Pavia University.
1 Introduction
The review positions HSI and deep learning as tools for faster, non-destructive agricultural analysis, while synthesizing datasets, methods, and applications. It covers HSI fundamentals, public datasets, deep-learning techniques, and agricultural uses.
- Motivation: HSI offers a rapid, non-destructive alternative to time-consuming, labor-intensive, and expensive traditional crop classification approaches.It captures information across multiple wavelengths for monitoring crop growth and agricultural conditions.
- Motivation: Deep learning can automatically extract relevant characteristics and has shown promise for agricultural HSI applications.The review discusses deep learning alongside HSI for crop classification and other agricultural tasks.
- Scope and contribution: The review examines HSI applications, compares HSI with RGB cameras, and discusses how deep learning can support product quality management.HSI acquires substantially more data in a single image than RGB cameras, including information outside human visual perception.
- Scope and contribution: The paper provides foundational HSI information, analyzes publicly available agricultural datasets, and reviews deep-learning techniques with their strengths and limitations.These contributions are intended to support understanding of HSI technology and its agricultural uses.
- Scope and contribution: The review presents agricultural applications of HSI, including uses intended to improve crop yield, quality, and safety.The paper also surveys prior reviews and organizes its discussion across HSI, datasets, deep learning, applications, limitations, and future research.
2 Hyperspectral Imaging
Hyperspectral imaging combines spatial imaging with spectroscopy to record spectral information for each pixel. Its analysis typically proceeds from acquisition and preprocessing through segmentation and model selection.
- HSI fundamentals: HSI simultaneously acquires spatial and spectral data by merging conventional imaging with spectroscopy.The technique is also described as chemical or spectroscopic imaging and is used for technical inspection of fruits and vegetables.
- HSI fundamentals: An HSI system uses an illumination source, objective lens, ROI detector, and spectroscopic imager.The illumination and lens choices affect reliable performance, focusing, and formation of output-image pixels.
- HSI fundamentals: HSI records spectral variation across pixels to support substance identification, target detection, and agricultural analysis.It combines spectroscopic methods with two-dimensional spatial information and one-dimensional spectral detail.
- Agricultural use: Hyperspectral sensors can assess crop health from distinctive spectral signatures associated with different growth stages.A material’s spectral signature reflects how its structure emits, absorbs, or reflects light.
- Analysis workflow: HSI analysis proceeds through image acquisition, spectral correction and calibration, segmentation, and selection of regression or classification models.Acquisition settings include sensor and platform choice, spectral and spatial resolution, illumination, scan rate, frame rate, and exposure time.
3 Acquisition Modes
Hyperspectral systems acquire spatial and spectral information through four modes: whisk-broom, push broom, staring, and snapshot. These modes differ in whether sensors scan line by line or capture the scene at once.
- Acquisition modes: HSI acquisition systems are classified as whisk-broom, push broom, staring, and snapshot methods.The classification is based on how each method acquires spectral and spatial information.
- Whisk-broom: Whisk-broom imaging scans a target one line at a time with a moving sensor and combines spectral images into a 3D hyperspectral cube.Its parallel scanning strips can provide high spatial resolution.
- Push broom: Push broom imaging captures successive lines with a stationary hyperspectral camera and combines them into a 3D hyperspectral cube.It commonly uses a linear array sensor and avoids mechanical sensor movement.
- Staring: Staring imaging captures the complete target scene at once with a stationary camera or array sensor.Unlike push broom imaging, it does not scan the scene line by line.
- Snapshot: Snapshot imaging captures the entire field of view in a single exposure using a specialized focal plane or detector array.The method is designed for single-snapshot acquisition.
4 Datasets for HSI in Agriculture
Agricultural HSI research relies on accessible datasets and performance criteria for training and comparing deep learning models. Dataset usefulness depends on data quantity, spatial resolution, spectral channels, and scenario diversity.
- Dataset resources: Open-source HSI datasets support the training and validation of deep learning networks and the comparison of algorithm performance.The review introduces evaluation criteria before describing publicly available datasets.
- Dataset characteristics: Agricultural HSI datasets differ in data quantity, spatial resolution, spectral channels, and variety of scenarios.These characteristics are identified as critical for evaluating available datasets.
- Dataset resources: Table 2 presents selected hyperspectral agricultural datasets for research use.The table is framed as a collection of suggested datasets for agriculture.
- Representative datasets: Salinas contains agricultural classes and was captured by the same sensor used for Indian Pines, while University of Pavia imagery has 103 spectral bands from the ROSIS sensor.Both Indian Pines and Salinas are airborne hyperspectral datasets.
5 Deep Learning-Based Approaches for HSI Classification
The review surveys CNNs, autoencoders, DBNs, RNNs, GANs, active learning, and transfer learning for HSI classification. These approaches extract spectral-spatial features or address limited labels, while their application remains constrained by noise, variability, computational cost, and domain shift.
- Overview: The review covers CNN, SAE, RNN, GAN, DBN, transfer learning, and active learning approaches for hyperspectral classification.These methods are presented as recent deep learning techniques for classifying hyperspectral data.
- CNN: CNNs extract spectral-spatial discriminative features and have outperformed Random Forest, Support Vector Machine, and k-Nearest Neighbors in multiple studies.CNNs are widely used for hyperspectral image classification, detection, segmentation, and land-use or land-cover classification.
- CNN: ResNet has been used for hyperspectral classification, denoising, spatial-resolution enhancement, and unsupervised spectral-spatial feature learning.The review identifies ResNet as a major development for training deep CNN models.
- Autoencoders: SAE models reconstruct inputs through stacked encoder-decoder layers to extract abstract features during unsupervised pretraining.The review describes this as an early application of deep learning to HSI analysis.
- DBN: DBNs stack Restricted Boltzmann Machines with interlayer connections and initially pretrain the layers unsupervised using unlabeled samples.The output of each layer serves as the input to the next layer.
- DBN limitations: DBNs are limited by poor transferability, model over-complexity, sensitivity to noisy HSI data, and high training cost.The review specifically links noise sensitivity to environmental factors and computational cost to large hyperspectral images.
- RNN: RNNs model hyperspectral pixels as sequential data through recurrent hidden states that depend on previous steps.Reported challenges include structural changes, agricultural data variability, and training difficulty.
- RNN limitations: RNN application to HSI is challenged by atmospheric, illumination, and viewpoint changes, as well as weather, soil, and plant-growth variability.The sequential nature of HSI data also complicates training.
6 Applications of HSI technology in agriculture
HSI supports agricultural applications including soil assessment, crop yield prediction, crop classification, nutrient and contaminant estimation, disease monitoring, and plastic detection. These applications use detailed spectral information, while sensing platforms and spatial resolution constrain some detection tasks.
- Agricultural applications: HSI applications span soil assessment, crop yield prediction, crop classification, nutrient estimation, disease monitoring, and invasive-species detection.The section organizes these uses across agricultural management and monitoring tasks.
- Yield prediction: HSI estimates crop yield using indicators of crop health and vigor, including chlorophyll, water content, nutrient levels, and vegetation changes.These measurements are linked to yield-estimation applications in precision agriculture.
- Crop classification: HSI classifies crops by capturing distinctive spectral information in visible and near-infrared wavelengths across crop types and growth stages.These spectral signatures provide information for classification purposes.
- Disease and species monitoring: HSI detects plant stress and disease symptoms by analyzing narrow-band reflectance changes, including chlorophyll and leaf-water changes not visible to the human eye.The same spectral analysis supports invasive plant-species detection.
- Plastic pollution: Plastic detection depends on sensor resolution: microplastics require laboratory-scale centimeter distances, whereas macroplastics can be detected remotely but with satellite-data constraints.Satellite approaches have low spatial resolution, fixed sampling times, rigid access protocols, and limited acquisition customization; UAV HSI can improve spatial resolution for small plastics.
7 Discussion
HSI provides sample data that can be combined with machine-learning algorithms to determine quality while reducing unpredictability associated with manual detection procedures.
- Quality assessment: HSI extracts hyperspectral sample data that can be merged with machine-learning algorithms to determine sample quality.The workflow is presented as an alternative to classical detection procedures involving manual processes, instruments, and reagent preparation.
8 Conclusion
The review synthesizes HSI classification technologies, deep-learning methods, datasets, applications, achievements, and research gaps. It identifies data complexity, limited samples, low spatial resolution, target-detection difficulty, and unresolved methodological and dataset needs as continuing challenges.
- Review scope: The review presents individual HSI methods together with their performance, achievements, and research gaps.It also describes recent techniques that may support future research.
- Limitations: HSI analysis remains constrained by data complexity, high cost, low spatial resolution, limited training samples, and unpredictable target and background spectra.These factors complicate interpretation, classification, and efficient target-detection algorithm development.
- Deep-learning analysis: Deep-learning methods are analyzed for HSI classification and evaluated for their effectiveness.The paper frames these methods as improved approaches to HSI analysis.
- Future scope: Future work includes automatic parameter optimization, redundant-band diagnosis, compact mobile HSI systems, and more effective learning from high-dimensional data cubes.The review also identifies meta-learning and newer agricultural datasets as underexplored or needed directions.