This paper addresses the challenge of hyperspectral image classification under conditions of limited labeled data and class imbalance. An improved method based on the AC-WGAN-GP architecture is proposed to enhance classification performance through dataset augmentation with synthetic samples generated via class-aware sampling and label embedding. The generator, discriminator, and classifier were modified accordingly, resulting in high classification accuracy on standard benchmark datasets. The proposed approach demonstrates strong potential for applications in remote sensing and precision agriculture.
Hyperspectral imaging (HSI) captures hundreds of contiguous spectral bands across the visible to
SWIR range, enabling fine-grained detection of crop conditions such as nutrient deficiency, disease,
and water stress [
HSI is valuable in precision agriculture due to its ability to distinguish visually similar crops and
detect early plant stress [
Data collection is supported by platforms like satellites (e.g., Sentinel-2, Landsat), which offer
wide coverage but lower resolution and weather limitations [
Generative adversarial networks (GANs) are effective in generating structured data, including
hyperspectral images, helping reduce dependence on large labeled datasets—a key benefit in
datascarce settings [
The model generates realistic synthetic samples with specified labels, preserving diversity even
with limited training data. Architectural enhancements further address class imbalance and
spectral similarity, improving classification performance in both HSI and MSI contexts—relevant
for precision agriculture [
The AC-WGAN-GP architecture includes a generator (G), discriminator (D), and auxiliary
classifier (C). The generator receives Gaussian noise, spectral features (e.g., PCA), and class labels
(one-hot or embedded) to produce synthetic spectral samples. Its structure comprises 1D
transposed convolutions (Deconv1D) with ReLU activations and a final Tanh layer, using batch
normalization for training stability [
Figure 1 shows the data flow and interaction among components, each optimized within a unified training framework. The discriminator, built from 1D Conv layers with Leaky-ReLU activations, evaluates sample authenticity. It outputs a linear value and uses Wasserstein loss with gradient penalty, ensuring stable training under distribution shifts [26].
The auxiliary classifier performs multi-class classification on both real and generated samples. It
consists of a Conv layer, flattening, and a fully connected Softmax output. Batch normalization is
excluded to preserve spectral sensitivity. It also guides the generator to produce correctly labeled
data [
Together, G, D, and C form a closed feedback loop: the generator is informed by both
discriminator and classifier, promoting realism and class accuracy—crucial for imbalanced or
overlapping spectral classes [
The AC-WGAN-GP model trains the generator (G), discriminator (D), and auxiliary classifier (C) in
an alternating fashion to ensure stable and controlled generation. Batch Normalization (BN)
accelerates convergence, while Gradient Penalty (GP) enforces Lipschitz continuity for stable
training [
The discriminator uses the Wasserstein loss with GP:
L D = ~x ~ pg D (~x )− x ~ p( x) D ( x)+ λx^~ p( x^) [( ‖ ∇ x^ D ( x^) ‖2− 1)2 ].
This balances the Wasserstein distance and gradient regularization to prevent mode collapse. The generator minimizes:
LG = − ~x ~ pg [D (~x ) ]+ [log p ( С = с |~x ) ] combining realism (discriminator) and class consistency (classifier).
The classifier is trained on both real and synthetic data:
LC = [log p ( C = c | x) ]+ [log p ( C = c |~x ) ] ensuring inter-class discrimination even under spectral overlap and imbalance.
HSI processing faces the “curse of dimensionality”: hundreds of spectral channels increase training
complexity and risk overfitting, especially with limited labeled data [
AC-WGAN-GP addresses this by generating synthetic samples that preserve spectral and
semantic class properties. Conditional generation via class labels and classifier guidance improves
data diversity and reduces overfitting [
Spectral overlap between classes causes classification ambiguity. The model maintains semantic
consistency through classifier feedback, enhancing discrimination [
In summary, AC-WGAN-GP effectively processes hyperspectral data under limited-label conditions, enhancing classification through synthetic, spectrally valid augmentation. (1) (2) (3)
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
Accuracy (OA), Average Accuracy (AA), and the Cohen’s Kappa coefficient (), which are standard
in hyperspectral classification research [
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.
Overall Accuracy (OA) is a standard metric in HSI classification that measures the proportion of correctly predicted samples among all test samples:
where is the total number of test samples, C¯ is the number of classes, and ℎ represents correctly classified samples of class (confusion matrix diagonal).
Average Accuracy (AA) evaluates classification performance across all classes equally, regardless of class size. It is calculated as the mean of per-class accuracies:
OA =
¯ 1 C
N ∑i= 1 hii, AA =
¯ 1 ∑C hii ,
C¯ i= 1 N i κ=
C¯ C¯ N ∑ hii− ∑ ( hi+⋅ h+i)
i= 1 i= 1 where C¯ is the number of classes, ℎ the correctly classified samples for class , and the total test samples in class .
The Kappa coefficient () measures agreement between predicted and true labels while
accounting for chance. Unlike OA, it reflects class distribution, making it suitable for imbalanced
datasets [
C N 2− ∑ ( hi+⋅ h+i)
i= 1 where is the total number of test samples, C¯ the number of classes, ℎ correct predictions, ℎ+ actual counts, and ℎ+ predicted counts per class.
The improved AC-WGAN-GP retains the classical conditional GAN structure comprising a generator (G), a discriminator (D), and an auxiliary classifier (C), but introduces targeted modifications to address class imbalance, spectral overlap, mode collapse, and training instability (Figure 2).
A revised training strategy complements the architecture. Conditional samples are generated using class embeddings and PCA vectors, followed by clustering in spectral space—applied only to real training data. Synthetic samples closest to the cluster centers (measured by cosine similarity) are selected and merged with real data for classifier training.
Crucially, all processing is confined to training data, avoiding test set leakage. This improves evaluation rigor and reproducibility, ensuring fair performance assessment under realistic constraints.
In the improved AC-WGAN-GP, the generator synthesizes conditional hyperspectral samples by combining class and spectral information. The baseline design with Deconv1D layers, ReLU, and batch normalization suffered from mode collapse and weak control via one-hot labels (Figure 3).
The updated architecture incorporates a Class-aware Sampling and Label Embedding (CS+LE) module, encoding labels into dense vectors and concatenating them with PCA features and Gaussian noise. This richer input better captures class identity and spectral variation.
To improve training stability and representation of minority classes, ResNet-style Deconv1D blocks with skip connections were added. A cross-attention mechanism aligns label and spectral embeddings with generator features, enhancing semantic coherence. Spectral Dropout is applied in intermediate layers to zero out entire spectral bands, improving robustness. The output is generated via UpSampling1D and a Conv1D layer with Tanh activation, ensuring normalized spectra. This architecture produces more diverse, class-consistent, and spectrally realistic samples.
In AC-WGAN-GP, the discriminator assesses how closely generated spectra resemble real hyperspectral data and provides feedback to the generator. The initial version used Conv1D layers with LeakyReLU and batch normalization, but the latter conflicts withWGAN-GP’s requirement for sample independence, causing instability (Figure 4).
To address this, batch normalization was replaced with LayerNorm, which operates per sample and ensures stable training with gradient penalty. Each Conv1D layer is followed by LeakyReLU and LayerNorm for consistent processing. To mitigate mode collapse, a Minibatch Discrimination layer was added to detect similarity across samples, encouraging output diversity. The final output is a scalar critic score from Flatten and Dense(1), as required by the WGAN formulation. Overall, these changes improve training stability, prevent mode collapse, and enhance the model’s ability to distinguish synthetic from real spectra.
The auxiliary classifier C in AC-WGAN-GP predicts class labels for real and synthetic samples and guides the generator. The initial design with a single Conv1D layer and Softmax output was too shallow to handle spectral similarity and rare classes (Figure 5). To improve performance, the classifier was deepened with three Conv1D layers followed by ReLU and batch normalization. Class labels are passed as dense embeddings, encoding semantic relationships. Extracted features are flattened and concatenated with the label embedding, then processed by a dense layer (256 units) used for both classification and contrastive loss. This enhances class separation while maintaining intra-class compactness.
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: where: denotes the latent noise vector; is the conditional class label; (, ) is the generated spectral sample; ((, )) is the “realism” score assigned by the discriminator.
1. Cosine Similarity with PCA Vectors (Cosine PCA Loss)—ensures that the generated spectrum aligns with the average PCA vector of its target class:
LWGAN = − z,c [D ( G ( z , c )) ], L PCA= [1 − cos ( x^, xPCA ) ].
Lalign= [1 − cos ( f , e ) ]. 2. Cosine Alignment Loss—enforces the classifier’s internal feature representation to align with the class embedding vector :
3. Categorical Cross-Entropy—penalizes the generator if the classifier fails to recognize the correct class of a generated sample:
Lce= z,c [− log P cls ( c | G ( z , c )) ].
The full generator loss is then defined as:
LG = LWGAN + PCA⋅ L PCA + align⋅ Lalign+ ce⋅ Lce, where PCA, align, and ce 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. (7) (8) (9) (10) (11) 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.
1. Categorical Cross-Entropy (Class-Weighted) This standard classification loss is weighted to compensate for class imbalance:
Lce= − x,y [ω y⋅ log P cls ( y | x) ], (14) where is the spectral sample, is the true class label, cls( | ) is the predicted probability, and is the inverse class frequency weight.
2. Contrastive Loss This term promotes closeness of features from the same class and separation between features from different classes:
Lcontrast = ,{max ( 0 ,‖ f i− f j‖ − δ )2 ‖ f i− f j‖ 2 , if yi= y j , , if yi ≠ y j where , are feature vectors and is a margin parameter.
3. Cosine Alignment Loss Aligns the feature vector with the corresponding class embedding: The discriminator loss is defined as:
L D = ~x ~ pg D (~x )− x ~ p( x) D ( x)+ ⋅ Lgp,
Lgp= x^~ p( x^) [( ‖ ∇ x^ D ( x^) ‖2− 1)2 ], (12) (13) (15) (16) (17) (18) Lalign= x,y [1 − cos ( f ( x) , e y) ],
Ldiv= ∑ ( i ≠ j || ei− e j ||2 + ε 1 ), 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 space:
where , are embeddings of different classes, and is a small positive constant to avoid division by zero.
Total Classifier Loss:
LС = Lce+ contrast⋅ Lcontrast + align⋅ Lalign+ div⋅ Ldiv, where contrast, align, div are hyperparameters controlling the contribution of each regularization component. These are selected empirically based on task complexity and class imbalance.
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. Synthetic samples were generated inonline 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.
The classification performance was evaluated on three benchmark HSI datasets—Salinas, Indian Pines, and KSC—at varying training ratios. As shown in Table 2, all datasets demonstrate clear improvement in OA, AA, and with increased training size. Prediction maps for studied datasets are shown in Figures 6–8.
Per-class F1 analysis reveals that classes with stable and well-separated spectral signatures (e.g., Stubble, Woods, Water) achieve F1 scores above 90–95%. In contrast, classes with limited samples or high spectral overlap (e.g., Oats, Corn, Oak Forest) show lower F1 scores (40–70%).
The improved model achieves reliable classification even under extreme label scarcity, achieving over 90% OA on the Salinas and KSC datasets using only 5% of labeled data, and over 68% on Indian Pines—one of the most challenging datasets.
This study presented an enhanced AC-WGAN-GP architecture for conditional hyperspectral sample generation, addressing class imbalance, spectral overlap, and training instability. Improvements include class-aware sampling, label embeddings, PCA-based conditioning, ResNet deconvolutions, crossattention, and spectral dropout in the generator; layer normalization and minibatch discrimination in the discriminator; and weighted categorical, contrastive, and embedding divergence losses in the classifier.
The method demonstrated strong performance under limited data and imbalanced class distributions, with all evaluations performed without test data leakage. Future research will extend this approach to 3D hyperspectral data, semantic segmentation tasks.
While preparing this work, the authors used the AI programs Grammarly Pro to correct text grammar and Strike Plagiarism to search for possible plagiarism. After using this tool, the authors reviewed and edited the content as needed and took full responsibility for the publication’s content. [26] 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 [27] Y. Lu, D. Chen, E. Olaniyi, Y. Huang, Gans for Image Augmentation in Agriculture: A Systematic Review, Comput. Electron. Agriculture 200 (2022). doi:10.1016/j.compag.2022.107208 [28] 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:10.1007/978-3030-48453-8_3 [29] 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 [30] A. V. Iatsyshyn, et al. Application of Augmented Reality Technologies for Education Projects
Preparation, in: Cloud Technologies in Education, 2643, 2020, 134–160. [31] M. Zaliskyi, R. Odarchenko, S. Gnatyuk, Y. Petrova, A. Chaplits, Method of Traffic Monitoring for DDoS Attacks Detection in e-Health Systems and Networks, in: Informatics & Data-Driven Medicine, 2255, 2018, 193–204. [32] V. Dudykevych, H. Mykytyn, K. Ruda, The Concept of a Deepfake Detection System of Biometric Image Modifications based on Neural Networks, in: 2022 IEEE 3rd KhPI Week on Advanced technology: conference proceedings, 2022, 585–588. [33] V. Dudykevych, S. Yevseiev, G. Mykytyn, K. Ruda, H. Hulak, Detecting Deepfake Modifications of Biometric Images using Neural Networks, in: Cybersecurity Providing in Information and Telecommunication Systems, 3654, 2024, 391–397. [34] V. Kharchenko, I. Chyrka, Detection of Airplanes on the Ground using YOLO Neural Network, in: Int. Conf. on Mathematical Methods in Electromagnetic Theory, 2018, 294–297. doi:10.1109/MMET.2018.8460392 [35] O. Zaporozhets, V. Isaienko, K. Synylo, Trends on Current and Forecasted Aircraft Hybrid Electric Architectures and Their Impact on Environment, Energy 211 (2020). doi:10.1016/j.energy.2020.118814 [36] Z. Hu, Y. Khokhlachova, V. Sydorenko, I. Opirskyy, Method for Optimization of Information Security Systems Behavior under Conditions of Influences, Int. J. Intell. Syst. Appl. 9 (12) (2017) 46–58. doi:10.5815/ijisa.2017.12.05