This paper presents an intelligent system for hyperspectral image classification based on an enhanced generative adversarial network with embedded label conditioning. The proposed architecture enables effective augmentation of limited training datasets with class-specific synthetic spectral samples. Key components include label embeddings, spectral regularization, and tailored loss functions designed to improve class separability in the feature space. Experiments on benchmark hyperspectral datasets demonstrate improved classification accuracy, even under scarce supervision. The approach shows strong potential for precision agriculture and vegetation monitoring applications.
Hyperspectral imaging (HSI) captures detailed reflectance information by collecting hundreds of
narrow, contiguous spectral bands, spanning visible to short-wave infrared (SWIR) wavelengths. This
enables detection of subtle physiological variations in crops, such as water stress, chlorophyll
deficiency, or early disease presence [
By contrast, multispectral imaging (MSI) acquires data in a limited number of wide spectral bands
(typically 3 15), often corresponding to red, green, blue (RGB), near-infrared (NIR), and red-edge
ranges. While MSI supports widely used vegetation indices such as NDVI and EVI, it lacks the spectral
resolution needed to distinguish between spectrally similar vegetation types [
This rich spectral-spatial structure makes HSI suitable for classification, target detection, and anomaly identification tasks [18]. However, its high dimensionality also introduces computational and statistical challenges. Specifically, it increases memory consumption and training time, and exacerbates overfitting in machine learning models.
To mitigate these issues, dimensionality reduction methods such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), and deep autoencoders are often applied prior to classification [7, 15].
Modern approaches to HSI classification rely on deep learning architectures such as 2D/3D CNNs and hybrid transformers that jointly capture spatial and spectral features [9, 8]. Still, they face difficulties stemming from spectral redundancy, class imbalance, and limited labeled data.
These limitations motivate the use of generative models to augment the training set with synthetic but realistic hyperspectral samples, particularly in low-resource settings.
Hyperspectral image (HSI) classification is often hindered by two critical limitations: the scarcity of labeled data and the significant imbalance between common and rare classes. These issues are especially problematic in agricultural monitoring, where data collection is costly and class boundaries are often spectrally ambiguous.
To alleviate these challenges, Generative Adversarial Networks (GANs) have emerged as a viable solution. A classical GAN consists of a generator that synthesizes realistic data samples and a discriminator that distinguishes real from synthetic inputs. When extended with a class-conditioning mechanism via an auxiliary classifier, this framework forms the Auxiliary Classifier GAN (AC-GAN) [11], capable of producing labeled synthetic spectra.
However, training GANs in high-dimensional spectral domains is prone to instability and mode collapse. The Wasserstein GAN with Gradient Penalty (WGAN-GP) introduces improvements by replacing the Jensen Shannon divergence with the Wasserstein-1 distance, while enforcing the Lipschitz constraint through gradient penalty [12]. This modification enhances convergence and training robustness.
By combining both approaches, the AC-WGAN-GP model provides a promising foundation for hyperspectral classification. It enables stable conditional generation by feeding the generator with Gaussian noise, PCA-reduced spectral vectors, and one-hot encoded class labels. The discriminator learns to distinguish real from fake samples, while the auxiliary classifier encourages semantic consistency in generated outputs.
Nonetheless, several shortcomings remain unresolved. One-hot labels do not capture inter-class generator is often underutilized, particularly in class-overlapping or data-sparse regions. Moreover, the architecture does not explicitly mitigate class imbalance, resulting in under-representation of minority categories.
To address these limitations, we propose a series of architectural and training refinements to the AC-WGAN-GP model. These include replacing one-hot labels with learnable class embeddings to reflect semantic proximity, incorporating residual deconvolution and cross-attention in the generator for enhanced spectral fidelity, and upgrading the discriminator with Layer Normalization and minibatch discrimination for better generalization. The classifier is restructured to jointly process both spectral features and label embeddings, enabling more discriminative and robust feature learning.
The following section presents the full specification of the improved model, along with a training strategy designed to avoid data leakage and promote effective representation of rare or spectrally ambiguous classes. 4. Problem Formulation of Hyperspectral Image Classification Using
In hyperspectral image classification tasks involving generative models, evaluation metrics play a crucial role in objectively comparing model performance and quantifying improvements resulting from architectural modifications. In this study, we employ three widely adopted metrics: Overall hyperspectral classification research [21, 20, 18].
While global metrics like OA, AA, and assess overall performance, per-class metrics Precision, Recall, and F1-score reveal how well individual classes are classified. correctly predicted samples among all test samples:
Overall Accuracy (OA) is a standard metric in HSI classification that measures the proportion of ), which are standard in
, where is the total number of test samples, ̅is the number of classes, and ℎ represents correctly classified samples of class (confusion matrix diagonal). of class size. It is calculated as the mean of per-class accuracies:
Average Accuracy (AA) evaluates classification performance across all classes equally, regardless where ̅ is the number of classes, ℎ the correctly classified samples for class , and the total test samples in class .
= ̅ ∑ =1 ℎ − ∑ 2 − ∑̅ ̅ =1(ℎ + ∙ ℎ+ ), =1(ℎ + ∙ ℎ+ ) = ℎ ℎ+ , actual counts, and ℎ+ predicted counts per class. model has predicted as belonging to that class. The metric is defined as: where is the total number of test samples, ̅ the number of classes, ℎ correct predictions, ℎ + Precision is the proportion of correctly classified pixels of a given class among all pixels that the samples predicted as class (regardless of their true class). that class. The metric is defined as: number of samples of class correctly classified as class ; ℎ+ samples that truly belong to class (ground truth labels). class metric is defined as: number of samples of class correctly classified as class ; ℎ + ,
The proposed
model builds upon the AC-WGAN-GP framework by introducing targeted improvements aimed at enhancing class-aware sample generation, increasing spectral diversity, and stabilizing training. The network still comprises three main components generator (G), discriminator (D), and classifier (C)
but their internal architectures are modified to address key challenges such as class imbalance, spectral overlap, and limited supervision.
The overall structure of the proposed architecture is illustrated in Figure 1, highlighting the internal connections between the generator, discriminator, and classifier modules. Embedding (CS+LE) module. Labels are encoded as dense vectors and concatenated with PCAtransformed spectral features and Gaussian noise. This allows the generator to operate in a more structured latent space, facilitating the creation of class-specific and spectrally consistent samples.
The generator architecture is extended with ResNet-style Deconv1D blocks and a cross-attention mechanism that aligns spectral and label embeddings with intermediate features. Spectral Dropout is applied to improve generalization by randomly zeroing entire spectral bands, mimicking sensor noise or occlusion.
In the discriminator, Batch Normalization is replaced with LayerNorm, ensuring stable gradients under gradient penalty regularization. To encourage diversity and mitigate mode collapse, a Minibatch Discrimination layer is included, allowing the model to detect and penalize overly similar samples.
The auxiliary classifier is redesigned with a deeper convolutional stack and uses embedded labels to better reflect inter-class relationships. It outputs both classification scores and internal features used in contrastive and alignment-based losses, promoting more compact and discriminative feature spaces.
To guide training, the loss functions for each module are expanded beyond standard adversarial objectives. The generator incorporates cosine similarity with class PCA centers, categorical loss, and alignment between features and label embeddings. The classifier combines class-weighted crossentropy, contrastive separation, cosine alignment, and embedding regularization. These terms are weighted by tuned hyperparameters to ensure balanced optimization across classes and objectives.
All synthetic samples are generated in online mode and selected via spectral clustering from real training data. Only the most representative samples are used in training, maintaining separation from test data and ensuring reliable evaluation.
Tables 1 and 2 summarize the architectural differences between the baseline and improved ACWGANGP models. They detail the input/output dimensions, layers, normalization, and activation functions for each module (G: Generator, D: Discriminator, C: Classifier).
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5
Input Size Z+PCA+embed 1/8H×1×256 1/8H×1×320 1/4H×1×256 1/2H×1×128 H×1×64
H×1×1 1/2H×1×64 1/4H×1×128 1/8H×1×256 1/16H×1×512 1/16H×1×512 H×1×1
H×1×64 1/2H×1×128 1/4H×1×256 1/4H×256
Dense + Reshape Cross-Attn + Concat ResNet Deconv1d (3×1×320×256) ResNet Deconv1d (3×1×256×128) ResNet Deconv1d (3×1×128×64) Upsampling+Conv1
D (3×1×64×1) Conv1d (3×1×1×64)
Conv1d (3×1×64×128)
Conv1d (3×1×128×256)
Conv1d (3×1×256×512)
Minibatch
Discrimination Flatten + Dense (1) Conv1d (5×1×1×64)
Conv1d (3×1×64×128)
Conv1d (3×1×128×256)
Flatten Dense + Label
Embed Dense (C) BN/LN
BN
2×1 2×1 2×1 2×1 2×1 1×1 2×1 2×1
LN LN LN BN BN BN BN
ReLU
G D C 1 2 3 4 5 1 2 3 4 5 1 2
To improve training stability and class balance, we implement a selective sampling approach for synthetic data. For each class, real samples are clustered using KMeans, and synthetic samples are selected based on cosine similarity to cluster centers. Both central and peripheral samples are chosen to ensure diversity. Labels are smoothed using a fixed coefficient .
Input: Real data , synthetic data , classes , smoothing , ratio Output: Filtered samples , smoothed labels for each class = 1 to do
Cluster с into clusters; foreach cluster do
Measure similarity between and center; Select % most and (1 )% least similar samples;
Append to with label ; end end end foreach label i in do
Smooth: ℎ = (1 − ) ∙ + /( − 1);
return , Algorithm 1: Simplified sample selection and smoothing.
The improved AC-WGAN-GP architecture employs a multi-component loss formulation to enable efficient and stable training across all network modules. Each component of the loss not only incorporates core adversarial objectives common to classical GANs but also introduces domainspecific terms tailored to the challenges of hyperspectral classification.
Unlike in traditional GANs, the generator in AC-WGAN-GP is optimized not only through adversarial feedback from the discriminator but also by enforcing alignment with class conditions and spectral context.
The basic Wasserstein loss component for the generator is given by:
= − , [ ( ( , ))], (7) where , is the expectation over all possible combinations of latent noise z and conditional class label c; denotes the latent noise vector; is the conditional class label; ( , ) is the generated spectral sample; ( ( ,
1. Cosine Similarity with PCA Vectors (Cosine PCA Loss) ensures that the generated spectrum aligns with the average PCA vector of its target class: where ̂ denotes the generated spectral sample and is the PCA-transformed class vector.
2. Cosine Alignment Loss the class embedding vector : = [1 − ( ̂,
)], lign = [1 −
( , )], where denotes the feature obtained from the classifier, and is the class embedding. 3. Categorical Cross-Entropy penalizes the generator if the classifier fails to recognize the correct class of a generated sample:
ce = , [−log ( ∣∣ ( , ) )], where denotes the probability predicted by the classifier for class c.
The full generator loss is then defined as: = + ∙
+ lign ∙ lign + ce ∙ ce, where , , and are weighting coefficients that control the contribution of each loss component. These are tuned empirically based on data characteristics, class imbalance, and desired classification performance.
In the AC-WGAN-GP framework, the discriminator functions as a critic that estimates the divergence between real and generated spectral samples. Unlike in classical GANs, where the discriminator performs binary classification, the WGAN formulation approximates the Wasserstein distance between real and synthetic distributions.
The discriminator loss is defined as: = ̃~ (g)[ ( ̃)] − ~ ( )[ ( )] + ∙ gp,
gp = ̂~ ( ̂)[(∥ ∇ ̂ ( ̂) ∥2− 1)2], where is the distribution of generated samples (from the generator); data is the distribution of real training samples; ̃ is a generated spectrum ( , ); is a real spectral sample; ̂ is a linear interpolation between and ̃; is a hyperparameter controlling the weight of the gradient penalty term ℒgp that ensures 1-Lipschitz continuity.
The auxiliary classifier in AC-WGAN-GP is responsible for both class prediction and learning discriminative features for regularization. Its loss function comprises several components aimed at maximizing classification accuracy while structuring the feature space. to align with (8) (9) (10) (11) (12) (13) 1. Categorical Cross-Entropy (Class-Weighted) This standard classification loss is weighted to compensate for class imbalance:
ce = − , [ ∙ log ( ∣ )], is the inverse class frequency weight. between features from different classes: where is the spectral sample, is the true class label, ( | ) is the predicted probability, and 2. Contrastive Loss This term promotes closeness of features from the same class and separation = , { ‖ − ‖ ,
2 max(0, ‖ − ‖ − )2, where , are feature vectors and is a margin parameter. 3. Cosine Alignment Loss Aligns the feature vector with the corresponding class embedding: where ( ) is the feature vector from the classifier and is the embedding of class . 4. Embedding Divergence Loss Regularizes class embeddings to prevent their collapse in latent lign = , [1 −
( ( ), )], = ∑ ≠ ‖ − ‖ + 1 2 , (14) (15) (16) (17) (18) where , are embeddings of different classes, and is a small positive constant to avoid division С = + ∙ + ∙ + div ∙ , where ,
, div are hyperparameters controlling the contribution of each regularization component. These are selected empirically based on task complexity and class space: by zero.
All experiments were conducted using PyCharm Community Edition 2024.3.4 with Python 3.9 and the TensorFlow 2.19.0 framework. The development environment ran on Windows 10 and local machine specifications were as follows: • •
RAM: 8 GB
Synthetic samples were generated in online mode without being saved to disk, reducing memory usage and preventing data duplication. Spectral vectors were reduced to = 30 components using Principal Component Analysis (PCA). The training and test splits were performed with strict class separation, eliminating potential data leakage.
hyperspectral image classification quality. Each modification step (from the baseline model to the and loss design. inclusion of cosine alignment and embedding divergence loss functions) progressively improves OA,
The most significant improvement is observed on the challenging Indian Pines dataset, where the introduction of the CSLE module (Step 2) raises OA to 67.74%. Adding ResNet-style deconvolutions training stability, particularly on KSC, where accuracy increases to 90.12%. Pines. This confirms the effectiveness of the proposed systemic enhancements to both architecture Impact of stepwise improvements on classification performance (Training ratio: 5%)
2 3 4 5 6
OA 89.73 89.76 90.09 90.44 90.55
AA 94.46 94.78 94.83 95.01 95.13 95.20 88.49 88.46 88.54 89.05 89.45 89.50
OA 66.30 67.74 67.89 68.08 68.10 68.60 AA 54.48 56.61 56.82 56.25 56.31 58.09 61.11 63.10 63.28 63.36 63.56 63.63 OA 89.76 89.81 90.05 90.12 90.16 90.62 KSC AA 84.61 84.26 85.53 85.57 86.18 86.30 88.50 88.64 89.18 89.27 89.42 89.55
increases from 88.70% to 91.27%, AA from 92.90% to 95.36%, and from 87.41 to 90.07. The highest F1-scores (above 99%) were achieved for classes with well-defined spectral structures Stubble, Broccoli, Vineyard soil. In contrast, classes with high spectral variability, such as Untrained vineyard and Untreated vineyard, show lower results (F1 = 78.63% and 70.49%, respectively).
As shown in Figure 2, the classification results for the Salinas dataset visually confirm the effectiveness of the model, especially on classes with clear spectral signatures.
The Indian Pines dataset is one of the most challenging due to strong spectral overlap between classes and significant class imbalance. Table 6 shows a steady improvement in accuracy as the training set size increases: overall accuracy (OA) rises from 54.15% to 77.54%, average accuracy (AA) from 43.24% to 71.15%, and the kappa coefficient ( ) from 46.78 to 74.35.
Table 7 shows per-class classification performance for Indian Pines with 5% training data. High F1scores were achieved for classes with well-defined spectral profiles: Hay-windrowed (92.05%), Wheat (85.97%), Woods (88.80%), and Grass-trees (83.52%). In contrast, classes with a low number of training samples, such as Oats, Alfalfa, and Grass-pasture-mowed, showed lower F1 performance, ranging from 15% to 65%. Figure 3 demonstrates the prediction performance on the Indian Pines dataset, where spectral overlap and class imbalance make classification particularly challenging.
Particularly difficult were the Corn and Soybean-clean classes, where F1 did not exceed 40 50% due to spectral similarity with nearby crop types. At 5% of training data, the improved model achieves OA = 68.60%, AA = 58.09%, demonstrating stable classification for well-separated classes but with limitations on spectrally overlapping ones. spectral channel, ground truth (GT), classification result.
The KSC dataset is characterized by clearly defined spectral differences between classes, which facilitates high classification performance. As shown in Table 8, even with only 5 % of training samples, the overall accuracy (OA) reaches 90.62%, the average accuracy coefficient ( ) is 89.55. scores are observed for classes with stable spectral characteristics: Salt Marsh (98.02%), Bare Soil (90.52%), Reed Swamp (94.61%), and Water (98.44%).
Lower classification performance is observed for Hardwood Forest (66.19%) and Oak Forest (70.46%), which is partly due to the limited number of training samples (8 11) and the spectral similarity to other forest types.
The results demonstrate that the improved model is capable of accurately classifying most classes even with minimal training data, achieving F1-scores above 85% for 10 out of 13 classes.
As shown in Figure 4, the KSC dataset classification map illustrates high accuracy for classes with well-separated spectral features. channel, ground truth (GT), and classification result.
In this work, we proposed an intelligent system for hyperspectral image classification based on the AC-WGAN-GP architecture, aimed at improving classification under conditions of data imbalance, spectral overlap between classes, and unstable training. The proposed improvements encompass all major components of the model: generator, discriminator, and classifier.
The generator was enhanced through the use of class-aware sampling and label embeddings, allowing better representation of minority classes. Its architecture is based on ResNet-inspired deconvolutional blocks with cross-attention mechanisms and spectral dropout, ensuring greater sample diversity and reduced risk of mode collapse. The generator input includes latent noise, a PCAtransformed spectral vector, and a dense class embedding.
The discriminator was adapted for stable WGAN-GP training by replacing Batch Normalization with Layer Normalization and introducing Minibatch Discrimination. This enables the detection of sample duplications and increases robustness against repetitive patterns in synthetic data. Additionally, the gradient penalty mechanism was implemented to enforce the 1-Lipschitz continuity requirement.
The classifier was modernized by replacing one-hot labels with dense embeddings and incorporating several loss functions: weighted categorical cross-entropy, contrastive loss, and embedding divergence. This enhanced the structure of the feature space, improved sensitivity to spectrally similar classes, and enabled better adaptation to rare cases.
Special attention was paid to fair evaluation: clustering and synthetic sample selectionwere performed strictly on the training set without access to test data. This eliminates any possibility of data leakage and guarantees the reliability of the reported metrics, even under stricter conditions than commonly used in related works.
The results confirm the effectiveness of the proposed intelligent system, particularly in classifying rare categories and under conditions of limited training data. Future research will aim to extend the system to 3D hyperspectral objects, segmentation tasks, and multisensor fusion, as well as to explore its potential in practical applications such as vegetation health monitoring and environmental assessment.
The authors have not employed any Generative AI tools. [6] B. G. Ram, P. Oduor, C. Igathinathane, K. Howatt, X. Sun, A systematic review of hyperspectral imaging in precision agriculture: Analysis of its current state and future prospects, Computers and Electronics in Agriculture 222 (2024). doi:10.1016/j.compag.2024.109037. [7] H. Yu, B. Kong, Y. Hou, X. Xu, T. Chen, X. Liu, A critical review on applications of hyperspectral remote sensing in crop monitoring, Experimental Agriculture 58 (2022). doi:10.1017/S0014479722000278. [8] V. Sineglazov, A. Kot, Design of hybrid neural networks of the ensemble structure, Eastern
European Journal of Enterprise Technologies 1 (2021) 31 45. doi:10.15587/1729-4061.2021.225301. [9] M. Zgurovsky, V. Sineglazov, E. Chumachenko, Classification and analysis topologies known artificial neurons and neural networks, in: Artificial Intelligence Systems Based on Hybrid Neural Networks, volume 904 of Studies in Computational Intelligence, Springer, Cham, 2021.doi:10.1007/978-3-030-48453-8_1. [10] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y.
Bengio, Generative adversarial nets, in: Advances in Neural Information Processing Systems, volume 27, 2014. doi:10.48550/arXiv.1406.2661. [11] A. Odena, C. Olah, J. Shlens, Conditional image synthesis with auxiliary classifier gans, arXiv preprint arXiv:1610.09585 (2016). doi:10.48550/arXiv.1610.09585. [12] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, A. C. Courville, Improved training of wasserstein gans, in: Advances in Neural Information Processing Systems, vol. 30, 2017, pp. 5767 5777. doi:10.48550/arXiv.1704.00028. [13] M. Arjovsky, S. Chintala, L. Bottou, Wasserstein GAN, arXiv preprint arXiv:1701.07875 (2017).
doi:10.48550/arXiv.1701.07875. [14] A. Abbas, S. Jain, M. Gour, S. Vankudothu, Tomato plant disease detection using transfer learnin with C-GAN synthetic images, Computers and Electronics in Agriculture 187 (2021). doi:10.1016/j.compag.2021.106279. [15] C. Sun, X. Zhang, H. Meng, X. Cao, J. Zhang, AC-WGAN-GP: Generating labeled samples for improving hyperspectral image classification with small-samples, Remote Sensing 14 (2022) 4910. doi:10.3390/rs14194910. [16] A. Radford, L. Metz, S. Chintala, Unsupervised representation learning with deep convolutional generative adversarial networks, arXiv preprint arXiv:1511.06434 (2016). doi:10.48550/arXiv. 1511.06434. [17] Y. Zhan, D. Hu, Y. Wang, X. Yu, Semisupervised Hyperspectral image classification based on generative adversarial networks, IEEE Geoscience and Remote Sensing Letters 15 (2018) 212 216. doi:10.1109/LGRS.2017.2780890. [18] Z. He, K. Xia, P. Ghamisi, Y. Hu, S. Fan, B. Zu, Hypervitgan: Semisupervised GAN with transformer for HSI classification, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 15 (2022). doi:10.1109/JSTARS.2022.3192127. [19] M. Zgurovsky, V. Sineglazov, E. Chumachenko, Classification and analysis of multicriteria optimization methods, in: Artificial Intelligence Systems Based on Hybrid Neural Networks, volume 904 of Studies in Computational Intelligence, Springer, Cham, 2021. doi:10.1007/978-3030-48453-8_2. [20] Y. Huang, Z. Chen, J. Liu, Limited agricultural spectral dataset expansion based on generative adversarial networks, Computers and Electronics in Agriculture 215 (2023). doi:10.1016/j.Compag.2023.108385. [21] L. Zhu, Y. Chen, P. Ghamisi, J. A. Benediktsson, Generative adversarial networks for hyperspectral image classification, IEEE Transactions on Geoscience and Remote Sensing 56 (2018) 5046 5063. doi:10.1109/TGRS.2018.2805286. [22] J. Feng, N. Zhao, R. Shang, X. Zhang, L. Jiao, Self-supervised divide-and-conquer generative adversarial network for classification of hyperspectral images, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1 17. doi:10.1109/TGRS.2022.3202908. [23] Y. Lu, D. Chen, E. Olaniyi, Y. Huang, Gans for image augmentation in agriculture: A systematic review, Computers and Electronics in Agriculture 200 (2022). doi:10.1016/j.compag.2022.107208. [24] M. Zgurovsky, V. Sineglazov, E. Chumachenko, Formation of hybrid artificial neural networks topologies, in: Artificial Intelligence Systems Based on Hybrid Neural Networks, volume 904 of Studies in Computational Intelligence, Springer, Cham, 2021. doi:10.1007/978-3-030-48453-8_3. [25] V. M. Sineglazov, K. D. Riazanovskiy, O. I. Chumachenko, Multicriteria conditional optimization based on genetic algorithms, System Research and Information Technologies (2020). doi:10.20535/SRIT.2308-8893.2020.3.07.