Neural reinforcement-oriented hyperspectral image compression: Adaptive approaches for enhanced quality

Abstract

Hyperspectral images (HSIs) typically contain hundreds or even thousands of bands that cover a wide range of wavelengths, each containing the material's spectral and spatial properties. New developments in remote sensing (RST) have enabled hyperspectral images (HSIs) with higher spectral and spatial resolution. However, these images' huge dimensionality and computational complexity provide difficulties for researchers. Broad spectral bands and redundancies contribute to huge dimensionality. In contrast, high spectral resolution, various sample ratios, and data dimensionality are the causes of computational complexity, which lowers precision and increases processing complexity. To address the challenge, a novel Neural Reinforcement with Adaptable Compression (NRAC) approach is proposed for dimension reduction in HSIs. The proposed technique involves initial computation of the mean for each instance in the input image and generating a matrix to select relevant bands from the HSI. After that, the dilation process takes place to mitigate intensity fluctuations of the HSI image. Then, to extract the spatial and spectral features, Convolutional CodeMapper is introduced for pixel value localization, thereby reducing computational overhead and overfitting. Thus, the NRAC approach accurately reconstructed the original HSI image and reduced the dimensionality issue. The evaluation utilized the Indian Pines, Jasper Ridge, Cuprite, and Pavia University Dataset, which attained high Peak Signal-to-Noise Ratio (PSNR), Mean Spectral Reconstruction Error (MSRE), Structural Similarity Index (SSIM), Feature Similarity Index (FSIM), and compression ratio metrics, demonstrating the efficacy of the proposed technique based on prior methods. The enhanced real-world implications of NRAC's capacity to interpret hyperspectral data include better remote sensing, agriculture, and environmental monitoring applications.