MULTI-TEMPORAL HYPERSPECTRAL ANALYSIS FOR EARLY FOREST DEGRADATION DETECTION USING DEEP LEARNING

Abstract

Forest degradation can begin before complete canopy removal. A forest area may still appear visually green while vegetation moisture, canopy structure, pigment response, or disturbance condition has already changed. Hyperspectral imagery is suitable for this problem because it records hundreds of narrow spectral bands and can therefore preserve subtle vegetation information that may not be visible in RGB or broad-band multispectral imagery. This thesis presents ST-HyperDegradeNet, a spectral-temporal hyperspectral deep learning framework for early forest degradation detection and hyperspectral change analysis. The proposed framework operates on two hyperspectral images acquired over the same region at two different times. It performs invalid value handling, reflectance normalization, robust clipping, PCA-based dimensionality reduction, patch extraction, class-imbalance-aware training, full-scene inference, and vegetation-constrained refinement for the forest case study. The main architectural idea is to use a true bi-temporal three-dimensional input representation. Patches from Time 1 and Time 2 are stacked along an explicit temporal dimension so that 3D convolution can learn spectral, spatial, and temporal interactions jointly. Squeeze-and-Excitation attention is used to recalibrate informative channels, while focal loss and class weighting support changed-class learning under imbalance. Two labelled hyperspectral change-detection benchmarks, Bay Area AVIRIS and Hermiston Hyperion, are used for quantitative validation. NEON SOAP Forest is used as a weakly supervised forest-domain case study because manual pixel-level degradation labels are not available. In the final test-split results, the proposed model achieved 98.46% accuracy and 96.53% changed-class F1 on Bay Area, and 98.98% accuracy and 97.97% changed-class F1 on Hermiston. For NEON, the proposed model achieved 87.10% pseudo-reference agreement, and the vegetation-constrained final map identified 1.62% early degradation and 1.91% strong degradation within the selected crop. The results show that ST-HyperDegradeNet is a competitive and interpretable framework for benchmark hyperspectral change detection and a practical candidate screening approach for forest degradation mapping under limited-label conditions. The thesis emphasizes cautious interpretation of pseudo-reference results and identifies future work involving field labels, multi-site validation, LiDAR integration, and longer multi-temporal time series.