3DSPECSN: Adaptive 3D spatial patch based siamese network for robust hyperspectral image analysis

Abstract

The classification of hyperspectral image (HSI) remains a challenging task due to high dimensionality, spatial correlation of the features and variability of the HSI data sources. This research introduces 3D Spatial Patch Extraction (3DSPE) technology which uses a Siamese Network framework to successfully capture multi-dimensional spectral-spatial data patterns for enhanced unmixing outcomes. The 3DSPE technique generates superior multiple spectral mixture signature detection due to its fusion of spectral and spatial features. The performance increases for endmember recovery and spatial distribution analysis when spectral patterns combine with spatial dependencies through the implementation of a Siamese Network architecture. A trainable image stratification approach for hyperspectral data increases speed of convergence and limits overfitting and builds generalized performance through adaptive optimization techniques at lower processing times for large datasets. The proposed framework shows strong performance in terms of Overall Accuracy (OA), Average Accuracy (AA), and Kappa Coefficient (κ) measurements from Indian Pines, Pavia University, and Salinas tests which attains nearly 99 % accuracies. The methodology not only achieved high accuracy, also enabling stable development of improved hyperspectral unmixing models that deliver better remote sensing results through improved precision along with enhanced flexibility and scalability. The approach provides a scalable efficient solution which can be applied to multiple remote sensing tasks, land-cover analyses and resource monitoring operations with potential seamless integration of other spectral analysis tools.