3D-Petln: Fine-Tuning 3D Patch Extraction in Hyperspectral Remote Sensing Data With Transfer Learning

Hyperspectral imaging faces many classification challenges due to high dimensionality, similar spectral pattern, and data diversity. The proposed 3D Hyperspectral Spatial Patch Extraction (3DPE) algorithm integrates into 3DPETLN framework to solve these problems by extracting important features and reducing computational cost over spectral correlation. The framework employs IPCA (Incremental PCA) to reduce data dimensions in real time while preserving vital spectral information to assist reliable performance and adjust to changing spectral pattern. A spectral-spatial attention mechanism in 3D CNN layers helps the network learn which spectrum ranges and spatial areas most significant. Zero-padded 3D spatial patches help retain features, save information and avoid boundary information loss over spectral-spatial dependencies. Our technique demonstrated significant performance enhancement of 99.9% against best existing methods when tested on Indian Pines and Pavia University datasets through remote sensing measures like class-wise accuracy and the OA, AA, and Kappa coefficient. The findings contributions are efficient, more precise, and better adaptable.