PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey for Hyperspectral Image Classification: The Evolution from Conventional to Transformers and Mamba Models

Muhammad Ahmad, Salvatore Distifano, Adil Mehmood Khan, Manuel Mazzara, Chenyu Li, Hao Li, Jagannath Aryal, Yao Ding, Gemine Vivone, Danfeng Hong

arXiv:2404.14955v4cs.CV

TL;DR

HSC must classify complex, high-dimensional hyperspectral data, while existing surveys vary in coverage, currency, depth, and perspective. This survey reviews conventional and deep-learning methods through Transformer and Mamba models, alongside related trends and challenges. Its experimental discussion reports that Transformer and graph-based classification yields significantly better results than convolution-based methods, especially with limited labeled training samples.

  • Problem

    HSC addresses classification of complex hyperspectral data, while existing surveys may be limited in coverage, currency, depth, or perspective.

    Source: §1 · p. 3

  • Method

    The survey reviews HSC methods from traditional and deep learning to Transformer and Mamba models, and discusses emerging topics, challenges, and experiments.

    Source: §17 · p. 32

  • Results

    Transformer and graph-based classification yields significantly better results than convolution-based methods, especially with limited labeled training samples.

    Source: §16 · p. 32

  • Takeaways & Limitations

    The survey identifies open challenges and future research directions across HSC methods and emerging models.

    Source: §17 · p. 32

    Deep-learning HSC faces challenges from scarce labeled data, computational requirements, and potential overfitting.

    Source: §10 · p. 15, §11 · p. 18

Abstract

from arXiv · show

Hyperspectral Image Classification (HSC) presents significant challenges owing to the high dimensionality and intricate nature of Hyperspectral (HS) data. While traditional Machine Learning (TML) approaches have demonstrated effectiveness, they often encounter substantial obstacles in real-world applications, including the variability of optimal feature sets, subjectivity in human-driven design, inherent biases, and methodological limitations. Specifically, TML suffers from the curse of dimensionality, difficulties in feature selection and extraction, insufficient consideration of spatial information, limited robustness against noise, scalability issues, and inadequate adaptability to complex data distributions. In recent years, Deep Learning (DL) techniques have emerged as robust solutions to address these challenges. This survey offers a comprehensive overview of current trends and future prospects in HSC, emphasizing advancements from DL models to the increasing adoption of Transformer and Mamba Model architectures. We systematically review key concepts, methodologies, and state-of-the-art approaches in DL for HSC. Furthermore, we investigate the potential of Transformer-based models and the Mamba Model in HSC, detailing their advantages and challenges. Emerging trends in HSC are explored, including in-depth discussions on Explainable AI and Interoperability concepts, alongside Diffusion Models for image denoising, feature extraction, and image fusion. Comprehensive experimental results were conducted on three HS datasets to substantiate the efficacy of various conventional DL models and Transformers. Additionally, we identify several open challenges and pertinent research questions in the field of HSC. Finally, we outline future research directions and potential applications aimed at enhancing the accuracy and efficiency of HSC.

1. Introduction

Hyperspectral imaging supports applications across agriculture, environmental monitoring, mineralogy, healthcare, and security, while hyperspectral image classification has developed from traditional methods toward deep learning and emerging architectures. This survey reviews these approaches, their challenges, experimental comparisons, and future directions.

1. Introduction

  • Hyperspectral imaging is used to distinguish materials with similar visual appearances but different spectral properties across applications including agriculture, environmental monitoring, mineral exploration, and healthcare.The introduction also describes uses in surveillance, security, and other domains.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Existing HSC surveys may have limited coverage, outdated information, insufficient depth, or a narrow application or methodological perspective.The survey frames its review as addressing these shortcomings for researchers and practitioners.Source: §1 · p. 3
  • The survey covers traditional and deep learning methods, emerging Transformer-based models, Explainable AI, Interoperability, Diffusion Models, and open research challenges.It also discusses Mamba and reports experiments on three hyperspectral datasets.Source: §1 · p. 3
  • Accurate HSC supports land-cover mapping, land management, resource planning, and environmental assessment.The introduction also discusses applications in agriculture, mineral exploration, and urban planning.Source: §1 · pp. 3–4
  • The survey describes CNNs, RNNs, autoencoders, and Transformers as deep-learning approaches that learn representations or capture spatial, spectral, or temporal relationships in hyperspectral data.It describes Transformers as using self-attention to model long-range dependencies and non-local spectral relationships.Source: §1 · p. 4
  • The survey is organized from hyperspectral data and learning categories through conventional methods, deep models, emerging architectures, experimental results, and future research directions.Its outlined topics include Explainable AI and Interoperability, Diffusion Models, and Mamba.Source: §1 · pp. 4–5

2. Hyperspectral Data Representation

Hyperspectral data can be represented spectrally, spatially, or through joint spatial-spectral information. The section also outlines learning strategies, including supervised and unsupervised learning and approaches that use limited labels or knowledge transferred across tasks.

2. Hyperspectral Data Representation

  • A hyperspectral sample is represented as a 3D hypercube containing spectral bands and spatial dimensions.In the notation X ∈ R^B×(N×M), B is the number of bands and N and M are spatial dimensions.Source: §2 · p. 5
  • Spectral representation processes each pixel independently, while spatial representation uses neighboring pixels to exploit spatial correlation.Spectral dimensionality reduction can reduce redundancy and improve class separability; spatial methods include texture and morphological features.Source: §2 · p. 5
  • Spatial-spectral approaches combine spectral features with spatial context by concatenating features or processing the three-dimensional image cube.The section describes both as strategies for using complementary information in HSI analysis and classification.Source: §2 · p. 5

3. Learning Categories

  • Supervised learning trains on labeled examples and requires substantial labeled data, whereas unsupervised learning can uncover patterns without predefined labels.The section gives PCA and k-means as examples of unsupervised techniques.Source: §3 · pp. 5–6
  • Learning categories include supervised, unsupervised, semi-supervised, reinforcement, transfer, online, and self-learning approaches.Unsupervised methods learn patterns from unlabeled data, while semi-supervised learning combines a limited labeled set with more unlabeled data.Source: §3 · p. 6

4. Brief Overview of Conventional Methods

Conventional HSC maps hyperspectral inputs to class labels using handcrafted features and classifiers, but faces challenges with dimensionality, feature engineering, spatial context, noise, scalability, and complex data distributions. Deep learning offers automated feature learning and flexible ways to model spectral and spatial information, while retaining data and compute demands.

4. Brief Overview of Conventional Methods

  • Conventional HSC treats classification as mapping hyperspectral inputs to class labels while minimizing the difference between predicted and actual labels.The section describes handcrafted feature extraction and classifiers including Random Forests, SVM, KNN, and ELM.Source: §4 · pp. 6–7
  • Handcrafted descriptors include global features such as color histograms, GIST, and texture descriptors, and local features such as HOG and SIFT.The section notes that SIFT is computationally intensive, whereas LBP uses simple binary operations on local neighborhoods.Source: §4 · p. 7
  • Traditional HSC approaches face the curse of dimensionality, feature-selection and extraction difficulties, limited spatial information, noise sensitivity, scalability issues, and poor adaptation to complex distributions.Handcrafted features also vary by task and can reflect subjectivity and bias in human-driven design.Source: §4 · p. 7

5. Advantages of Deep Learning

  • Deep learning can learn features without pre-specified information and handle spectral and spatial data individually or jointly.The section describes end-to-end optimization and CNNs’ ability to capture complex data distributions.Source: §5 · p. 7
  • Deep learning can benefit from pretraining and transfer learning, but HSC models still face demands for labeled data, computational resources, and protection against overfitting.The section notes that pretraining on large-scale datasets followed by fine-tuning on smaller HS datasets can enhance classification performance, especially with limited labeled data.Source: §5 · p. 8

6. Convolutional Neural Networks for HSC

CNN approaches for HSC extract features from spectral, spatial, or joint spectral-spatial inputs before classification. Their designs trade complementary information capture against data, preprocessing, generalization, and computational demands.

spatial CNN.

  • Spectral CNNs process each pixel’s original or reduced spectral profile as a 1D sequence to capture wavelength-wise intensity patterns.The input is a spectral vector with B bands.Source: spatial CNN · p. 8
  • CNNs use convolution and pooling to build hierarchical features, followed by fully connected layers or global average pooling and a classifier such as Softmax or SVM.Global average pooling summarizes spatial information and reduces parameters.Source: spatial CNN · p. 8
  • Spatial CNNs treat an HSI as a 3D volume and use convolutional filters to capture local spatial patterns from pixels and their neighbors.Pooling can condense feature maps, while PCA-derived components and other strategies have been used to add spatial information or address limited samples and overfitting.Source: spatial CNN · p. 9
  • Spatial CNNs may require dimensionality reduction or feature selection and can neglect useful spectral information, overfit limited labels, and generalize poorly across spatial settings.Spectral CNNs can neglect spatial context and face dimensionality, memory, overfitting, robustness, and preprocessing challenges.Source: spatial CNN · pp. 9–10
  • Spectral-spatial CNNs combine 1D spectral and 2D spatial convolutions, then fuse their features for classification; 3D CNNs can also extract joint information.The two feature types are complementary, but integrating them increases computational demands and can overfit small or imbalanced datasets.Source: spatial CNN · pp. 10–11
  • GCNs represent pixels as nodes with spectral signatures as features, connect spatial neighbors, and aggregate information across the resulting graph.Large-graph construction is computationally demanding; mini-batch training and superpixel segmentation are among proposed efficiency strategies.Source: spatial CNN · p. 11
  • Future CNN research should integrate spatial and spectral information without sacrificing spectral detail while addressing deployment costs, scarce labels, and generalization beyond grid data.The surveyed directions include efficient processing, unsupervised or semi-supervised learning, and combining CNNs with GCNs.Source: spatial CNN · p. 12

7. Deep Belief Network for HSC

DBNs learn hierarchical features through unsupervised, layer-by-layer training with Restricted Boltzmann Machines. HSC work combines spectral and spatial information, while addressing hidden-unit co-adaptation and limited use of DBNs.

7. Deep Belief Network for HSC

  • DBNs are hierarchical networks built from Restricted Boltzmann Machines, whose visible and hidden units support layer-by-layer feature learning from HS data.Their RBMs can be trained greedily to capture progressively abstract features.Source: §7 · p. 12
  • Typical HSC training uses unlabeled samples for unsupervised pre-training, followed by supervised fine-tuning on labeled samples.Hidden-unit co-adaptation and activation sparsity can limit feature diversity; diversified DBN regularization was proposed to address these issues.Source: §7 · p. 12
  • HSC studies use DBNs for texture enhancement and combine spectral and spatial features through logistic-regression or graph-based RBM pipelines.The survey describes DBNs as less used than other DNNs and points to improved pretraining, fine-tuning, and regularization as research directions.Source: §7 · pp. 12–13

8. Recurrent Neural Networks for HSC

RNNs treat HS spectral bands as sequential steps, with LSTM and GRU variants designed to handle long-term dependencies. HSC studies combine recurrent models with spatial features, while long sequences and overfitting remain concerns.

8. Recurrent Neural Networks for HSC

  • RNNs process HS spectral bands as sequential steps, and LSTM and GRU models are designed to address vanishing gradients and capture long-term dependencies.GRUs simplify LSTM gating and offer comparable performance with greater computational efficiency.Source: §8 · p. 13
  • HSC studies combine recurrent processing with convolutional, multiscale, or local spatial features, and apply recurrent models to multi-temporal and multispectral data.Examples include CRNNs that pair convolutional feature extraction with recurrent spectral-context modeling, including a semi-supervised variant using pseudo labels.Source: §8 · p. 13
  • Long input sequences may lead to overfitting, motivating research on sequence reduction, parallel processing, spectral-band grouping, and multi-temporal HSI.Sequence reduction should preserve important spectral and spatial information.Source: §8 · pp. 13–14

9. Autoencoders for HSC

Autoencoders learn compressed representations by reconstructing high-dimensional HS inputs, supporting unsupervised feature learning rather than direct classification. Stacked and denoising variants extend this approach to abstract features and noisy inputs.

9. Autoencoders for HSC

  • An HSC autoencoder encodes high-dimensional HS input into a latent representation and decodes it to reconstruct the original, minimizing reconstruction error.Its purpose is compressed feature learning rather than direct classification.Source: §9 · p. 14
  • Stacked autoencoders stack layers to learn increasingly abstract features, while denoising autoencoders corrupt inputs and reconstruct the original signal to handle noisy data.The survey describes AE, SAE, and DAE as tools for compressed feature learning and handling noisy inputs.Source: §9 · p. 14
  • One cited approach combined multilayer autoencoders with maximum noise fraction to learn high-level HS representations.The passage also notes a multi-manifold learning framework combined with a Counteractive AE.Source: §9 · p. 14

supervised learning.

AE-based HSC methods combine spectral and spatial representations through varied feature-learning and classifier designs. The survey notes that spatial information remains underused in some approaches, while intra-class variability and inter-class similarity persist.

supervised learning.

  • Early frameworks paired recursive AEs, two-stream networks, PCA, guided filters, sparse AEs, and Random Forests to extract or classify spectral-spatial information.Examples include a spectral-similarity-based RAE, an SAE-CNN two-stream model, and a PCA/guided-filter/sparse-AE pipeline.Source: supervised learning · p. 14
  • Later AE-based approaches combined diverse representations and classifiers, including SAE-ELM systems, Fisher-discriminant regularization, denoising AEs, and spectral-spatial features.Other designs used mutual information to segment spectral components or jointly exploited spectral-spatial features for oil-slick classification.Source: supervised learning · p. 14
  • Many reviewed techniques underused spatial information, while high intra-class variability and inter-class similarity remained challenges to classification accuracy.The survey identifies pre-training, co-training, and adaptive neural networks as possible strategies for improving accuracy and robustness.Source: supervised learning · p. 14

10. Limitations of Traditional Deep Learning

Traditional deep learning for HSC faces data, modeling, and computational constraints. The survey highlights limited labels and high-dimensional spectral data alongside interpretability, optimization, and depth-related training challenges.

10. Limitations of Traditional Deep Learning

  • High-dimensional, narrow-band HS data, low spatial resolution, scarce labels, noise, redundant bands, and spectral mixing complicate deep-learning-based HSC.Limited labels can contribute to the Hughes phenomena, overfitting, and poor generalization; the survey notes feature selection, dimensionality reduction, and data augmentation as responses.Source: §10 · p. 15
  • Deep models can be difficult to train and scale, with convergence challenges, high computational and storage burdens, and limited interpretability.The survey lists optimization methods for training efficiency and parallel or distributed computing for processing large datasets.Source: §10 · p. 15
  • Increasing network depth does not always improve accuracy, as exploding or vanishing gradients may adversely affect convergence.The survey cautions against assuming deeper networks yield higher accuracy.Source: §10 · p. 15

11. Transformers for HSC

This section surveys Transformer approaches for HSC, including attention-based modeling of spectral-spatial relationships and strategies such as transfer learning and domain adaptation to address limited labels and dataset differences. It also covers training-data expansion and emerging priorities in interpretability, multimodal integration, and deployment, along with benefits and practical constraints.

11. Transformers for HSC

  • SAT Net, SST, masked-image pre-training, contrastive learning, NEHT, and DBFFT combine attention with spectral, spatial, or CNN-based feature extraction.The approaches use mechanisms including residual multi-head attention, masked reconstruction, associative contrastive learning, dimensionality reduction, and dynamically weighted spectral and pixel features.Source: §11 · p. 16
  • Other HSC designs use EMP with TNT modules, spectral embedding and cross-layer fusion, separate spatial and spectral Transformer branches, or morphology-aware attention.HiT, MorphFormer, LSFAT, DATE, SSTN with factorized architecture search, and MAT with active learning add convolutional projections, multi-scale aggregation, or local-global attention; ASSAL reduces labeled-sample needs while maintaining high accuracy.Source: §11 · pp. 16–17
  • Swin Transformers use hierarchical windowed attention to capture multi-scale features with reduced computational complexity, but can be less suited to spectral sequences and remain resource-intensive and difficult to interpret.SSTs learn global context from raw pixels and can process high-resolution images without complex pooling, but quadratic attention, fixed-size patches, limited translation invariance, and data demands constrain scalability and fine detail.Source: §11 · pp. 17–18

11.a. Benefits and Challenges of Transformers

  • Transformer challenges include quadratic computational demands, limited labeled data, hyperparameter sensitivity, and interpretability concerns; hybrid models, alternative attention, and multimodal inputs are proposed directions.Hybrids can combine SST global context with CNN local features but add complexity; sparse attention may trade efficiency for performance, while multimodal fusion and fine-grained tokenization remain challenges.Source: §11 · pp. 18–19

11. Transformers for HSC

  • Transformers can model global and long-range spectral-spatial relationships, adaptively emphasize informative bands, and learn contextual features for HSC.Attention maps can also indicate influential features or image regions, offering some insight into classification decisions.Source: §11 · pp. 18–19

11.b.i. Domain Adaptation and Transfer Learning

  • Domain adaptation and transfer learning reuse knowledge from related datasets or modalities to improve HSC with limited labels, including when sensor dimensions or data distributions differ.Methods include transferring lower-layer CNN features and fine-tuning on target data, heterogeneous transfer across HSIs, band-selection-based pre-training, and using natural-image networks.Source: §11 · p. 20

11.b.ii. Data Augmentation

  • A data-augmentation CNN improved HSC accuracy by 10% over a PCA-based CNN; another study used flips, rotations, and translations to expand training data.Augmentation also includes transformations that preserve labels and oversampling that synthesizes samples to address class imbalance.Source: §11 · pp. 20–21

11.b.iii. Generative Adversarial Networks

  • GAN-based HSC studies used spectral-feature and spectral-spatial frameworks (Zhan et al., 2017; He et al., 2017b); other work targeted scarce examples with GAN-generated samples and class-wise imbalance through minority oversampling (Alipour-Fard and Arefi, 2020; Roy et al., 2021a).Source: §11 · p. 21

11.b.iv. Active Learning (AL)

  • Active learning selects informative unlabeled samples for annotation, using stream- or pool-based querying and heterogeneity, performance, or representativeness criteria to guide selection.Heterogeneity criteria include uncertainty, expected model change, and query-by-committee; representative sampling can be combined with heterogeneity in hybrid methods.Source: §11 · pp. 21–22

11.b.v. Explainable AI for HS Analysis

  • Explainable AI is increasingly sought for HS classification in critical decision-making, including environmental monitoring, agriculture, and defense.Source: §11 · p. 22

11.b.vi. Integration of Multimodal Data

  • Fusing HS data with modalities such as LiDAR or SAR can provide complementary information for classifying complex land-cover classes, but requires addressing alignment, dimensionality, model complexity, and data availability.Fusion approaches include early feature concatenation, late combination of classification outputs, and hybrid strategies.Source: §11 · pp. 22–23

11.b.vii. Edge Computing for Real-time Processing; 11.b.viii. Development of Benchmark Datasets

  • Lightweight edge models are explored for real-time, on-site HSC, while standardized benchmark datasets support fair algorithm comparisons and reproducible evaluation of generalization.Source: §11 · p. 23

12. Explainable AI and Interoperability

The section presents explainability and interpretability as ways to make HSC decisions more transparent and assessable, while noting that complex deep-learning models are difficult to interpret and interpretability can trade off with performance.

Importance and role

  • Explainability can help users assess model reliability, detect biases, identify influential features, and diagnose potential errors or limitations.The section also links interpretability to transparency, trust, and informed decisions in applications where misclassifications may matter.Source: §12 · p. 24

Interpretability techniques

  • Feature importance, saliency maps, attention maps, rule extraction, and model visualizations are presented as ways to expose influential features or model behavior.These approaches can highlight spectral bands or spatial locations, show attention weights, extract human-readable rules, or visualize internal representations.Source: §12 · p. 24

Expert collaboration and trust

  • Domain experts can validate and refine AI explanations against domain-specific knowledge, while explanations can support transparency and trust.The section also notes that interpretability may help meet regulatory or ethical requirements and aid model validation and verification.Source: §12 · p. 24

Challenges

  • CNNs and Transformers can be difficult to interpret because of their hierarchical and nonlinear nature, and high interpretability may trade off against model performance.Explanations should also align with human understanding and be meaningful to domain experts and end-users.Source: §12 · pp. 24–25

13. Diffusion Models for Hyperspectral Imaging

The section surveys diffusion approaches for hyperspectral imaging, including denoising, image fusion, and feature extraction. It describes their potential benefits alongside risks of over-smoothing and high computational cost.

Overview

  • Diffusion models use spatial and spectral information to improve image quality, denoise images, and enhance spectral features by propagating information across neighboring pixels.The section describes neighboring pixels as likely to have similar spectral characteristics.Source: §13 · p. 25

Denoising models

  • The Perona-Malik model adapts diffusion of image gradients to local image characteristics to reduce noise while preserving meaningful structures and spectral features.The section describes it as based on the heat equation and applied to hyperspectral denoising.Source: §13 · p. 25
  • The Total Variation model denoises hyperspectral images by minimizing image total variation, promoting piecewise smoothness while preserving sharp edges and fine details.Total variation is described as the sum of absolute differences between neighboring pixel values.Source: §13 · p. 25

Fusion, feature extraction, and limitations

  • Diffusion methods can fuse information from multiple hyperspectral images or modalities and enhance discriminative spectral features for classification or target detection.The section also notes that excessive diffusion may erase fine details and subtle spectral variation, while large-scale processing can be computationally intensive.Source: §13 · p. 25

Relevance to classification

  • Diffusion models are discussed as preprocessing tools for reducing noise before classification, with the aim of improving class separability and subsequent classification performance.The section connects these potential benefits to noise from sources such as atmospheric interference or sensor limitations.Source: §13 · p. 25

14. State Space Model (Mamba)

The section presents Mamba as a state-space approach for modeling hyperspectral spectral-spatial information with compact representations and lower computational demands than self-attention. It also reviews model variants, claimed benefits, and remaining limitations.

Mamba architecture

  • Mamba represents hyperspectral pixels as spectral-band sequences and uses state-space transitions to encode dependencies before producing classification outputs.Its components include spectral-spatial token generation, a state transition mechanism, and an output layer.Source: §14 · p. 26

Advantages

  • Mamba processes sequences linearly rather than with Transformers’ quadratic self-attention complexity, using fewer parameters and less memory while modeling long-range dependencies.The section links these properties to handling long spectral sequences and scaling to high-resolution hyperspectral datasets.Source: §14 · p. 26

Model variants

  • MHSSMamba adds multi-head self-attention and token enhancement, WaveMamba combines wavelet transformation with Spatial-Spectral Mamba, and Morp-Mamba integrates morphological operations with an SSM.The described goals include capturing spectral-spatial relationships, local textures and global context, or structural and spectral information.Source: §14 · p. 26
  • MambaHSI is reported to outperform in classification accuracy across four HSI datasets, while MiM uses cross-scan transformation, tokenized feature learning, and multi-scale loss.The section describes MiM as suited to remote-sensing tasks requiring high precision with limited training data.Source: §14 · p. 26
  • HyperMamba uses spatial-neighborhood and spectral-adaptive scanning to capture spatial-spectral dependencies, while HLMamba combines gradient-based edge extraction with multimodal Mamba fusion.The section also describes a CNN-Mamba fusion network for resolution enhancement using mutual guidance between hyperspectral and multispectral imagery.Source: §14 · pp. 26–27

Efficiency and applications

  • The section reports comparable or superior classification accuracy to Transformers and CNNs at lower computational cost, positioning Mamba for resource-constrained applications.Examples include onboard drone processing, real-time monitoring, and edge-device use in agriculture, environmental monitoring, and disaster response.Source: §14 · p. 27

Limitations

  • Mamba may face degraded performance as spectral dimensions grow, and its original design is limited in incorporating temporal dynamics in HSI sequences.Temporal integration is identified as relevant to applications involving environmental change, such as agriculture or urban development.Source: §14 · p. 27

Limitations and future directions

  • The section flags possible weaknesses in noisy-data robustness, feature interpretability, and heterogeneous data fusion, and proposes research on noise-resistant training, interpretable designs, and adaptive fusion.It also proposes enhanced feature representation and temporal Mamba extensions for dynamic monitoring.Source: §14 · p. 27

15. Open Challenges and Research Questions

The section identifies data, modeling, and deployment challenges in HSC and poses research questions about improving accuracy, robustness, interpretability, and applicability.

Limited labeled training data

  • Limited labeled training data is expensive and time-consuming to acquire and may hinder accurate, robust classification; proposed questions include transfer learning and active learning.The section asks how to use related domains or datasets and select informative samples when labeled data are scarce.Source: §15 · pp. 27–28

Curse of dimensionality

  • High spectral dimensionality raises the curse of dimensionality, prompting questions about dimensionality reduction and models that preserve or capture intrinsic data structure.The section specifically raises deep-learning architectures and autoencoders as possible approaches.Source: §15 · p. 28

Spectral-spatial complexity and robustness

  • Complex spectral-spatial patterns challenge traditional classifiers, while class imbalance, rare events, and environmental variability raise further questions about model design and generalization.The section asks about deep networks, graph models, attention, generative augmentation, preprocessing, and transfer learning for these challenges.Source: §15 · p. 28

Interpretability and resource constraints

  • Deep models may lack interpretability, and some HSC applications require real-time operation under processing, memory, or energy constraints.The section asks how to align explanations with domain users and balance model complexity against accuracy in resource-constrained settings.Source: §15 · p. 28

Conclusion

  • The section concludes that addressing these challenges is important for advancing HSC and developing more accurate, robust, and interpretable models for applications including agriculture and environmental monitoring.Remote sensing is also listed as an application domain.Source: §15 · p. 28

16. Experimental Results and Discussion

The experiments compare representative models on three HSI datasets under controlled, disjoint sample splits. Results vary with training-sample availability: Hybrid CNN and 3D Inception nets perform strongly with fewer labels, while Transformer and graph-based methods outperform convolution-based methods, particularly with limited labels.

16. Experimental Results and Discussion

  • Fair comparisons require identical sample locations and uniform training-sample counts across cross-validation rounds.Random selection can otherwise give models different samples when they are run at different times.Source: §16 · pp. 28–29
  • The study prevents training–test overlap by selecting training and test samples without overlap.The authors note that testing on the entire dataset after training on some of its samples can inflate accuracy.Source: §16 · p. 29
  • The comparison covers representative CNN, Inception, graph, and Transformer models across the IP, PU, and UH Scene benchmark datasets.Comparative methods can confuse samples with similar spatial structures, including Meadows and Bare Soil in Pavia University.Source: §16 · p. 29
  • The overall accuracy for the Grapes Untrained class is lower than for other classes.Source: §16 · p. 30
  • Hybrid CNN and 3D Inception nets consistently outperform other comparative methods with fewer labels, while Transformer and graph-based classification significantly outperform convolution-based methods, especially with limited labels.The reported rankings vary as training samples increase, when other methods may outperform Hybrid CNN and 3D Inception nets.Source: §16 · p. 32

17. Conclusions and Future Research Directions

The survey traces HSC from traditional machine learning through deep learning to Transformer and Mamba models. It highlights remaining challenges and proposes research directions spanning model efficiency, transfer learning, and real-world applications.

17. Conclusions and Future Research Directions

  • The survey charts HSC’s evolution from traditional machine learning to deep learning, Transformer architectures, and Mamba models.It discusses progress in attention mechanisms, self-supervised learning, and diffusion models for denoising and feature extraction.Source: §17 · p. 32
  • Mamba-based approaches show potential for integrating spatial-spectral information through architectures tailored to HSC.Source: §17 · p. 34
  • Data scarcity, interpretability, and real-time performance remain challenges; future work should address domain-specific advances, efficiency, and cross-domain transfer learning.The survey identifies precision agriculture and environmental monitoring among potential real-world applications.Source: §17 · p. 34
Loading 2404.14955v4…