Advanced hyperspectral snow and glacier classification using fully connected autoencoders

Abstract

The accurate classification of snow and glacier surface features is essential for glacier monitoring, mass balance estimation, melt-runoff modeling, and climate impact assessment in high-mountain regions. Hyperspectral imagery provides rich spectral information that enables fine discrimination of materials with similar appearances, which multispectral data often cannot achieve due to broader bands. However, challenges such as high dimensionality, class imbalance, limited labeled data, and spectral similarity between transitional facies limit the effectiveness of traditional methods. This study applies a class-specific Fully Connected Deep Autoencoder (FCDAE) framework for pixel-wise classification of seven snow and glacier endmembers such as Clean Snow, Dirty Glacier Ice, Firn, Glacier Ice, Ice Mixed Debris, Waterbody, and Non-Snow using EO-1 Hyperion data from the North-Western Himalayas. The approach integrates field-collected spectral signatures with image-derived endmembers and trains independent Autoencoders per class based on reconstruction error. The proposed FCDAE achieved an overall accuracy of 97.47%, Kappa coefficient of 0.96, and F1-score of 0.97 on the held-out test set, outperforming traditional machine learning baselines (Random Forest (RF): 95.74%, Support Vector Machines (SVM): 88.79%). These results highlight the effectiveness of reconstruction-based deep learning for hyperspectral snow and glacier feature classification in data-scarce, spectrally complex Himalayan environments.