Abstract
Hyperspectral image (HSI) classification plays a critical role in a range of remote sensing applications, including agriculture, urban planning, and environmental monitoring. However, its performance is often constrained by two major challenges: low spatial resolution and limited availability of labeled training data. To address these limitations, we introduce RES-P-berry-GAN, a lightweight yet powerful Residual Pyramid Generative Adversarial Network (GAN) specifically designed to enhance both spatial detail and classification accuracy in HSIs. The proposed architecture integrates three key innovations: Residual Dense Block Networks (RDBN) to ensure effective feature preservation and deep representation; a Pyramid structure that captures multi-scale contextual information; and a Pixel Attention (PA) mechanism that emphasizes important spectral signatures across spatial dimensions. Additionally, a ResNet-50-based feature extractor is employed to guide perceptual quality during training, ensuring the generated outputs retain high-fidelity spectral and spatial information. We evaluate RES-P-berry-GAN on three widely used benchmark datasets–Indian Pines, Pavia University, and Pavia Center. Our results demonstrate that the model achieves state-of-the-art classification accuracy, reaching up to 97.15%, while maintaining remarkable efficiency with only 2.90 million parameters. Compared to baseline methods such as ESRGAN and InfoGAN, RES-P-berry-GAN shows consistent improvements in both visual and quantitative metrics. The combination of lightweight design and high performance makes RES-P-berry-GAN highly applicable to real-world scenarios where computational resources and labeled data are limited. Its robustness across diverse landscapes highlights its potential for scalable deployment in hyperspectral analysis tasks.