Spectro-TwinNet: Dual-Stream Temporal Deep Learning for Robust Hyperspectral Land-Cover Classification

Abstract- Spectral noise, high dimensionality, and
limited time modeling in hyperspectral data render
adequate prediction of Land Use and Land Cover
(LULC) very difficult. To solve these problems, it
applies Spectro-TwinNet, a spectral-spatial deep
learning architecture that combines adaptive
preprocessing and enhanced temporal representation.
The proposed ANSER module considerably improves
the quality of data, decreasing MSE to 0.061, raising
PSNR to 61.27 dB, and SSIM to 0.944, which is better
than traditional filters. To extract features and classify
3DSpectro-TwinTCN architecture is designed to
combine 3D spectral encoding and dual-stream
temporal learning with maximizing 97.6% accuracy,
94.8% sensitivity, and 95.2% specificity, where the
application of high spectral robustness and temporal
consistency is provided. Experimental analyses attest
to the fact that unified preprocessing and a deep
temporal modeling method provide better
discrimination ability of complex land-cover patterns.