FCDAE: Fully Connected Deep Autoencoders for snow and glacier features classification from hyperspectral data

Deep Learning (DL) has revolutionized the analysis of hyperspectral data, provided benefits in noise handling and reduced time complexity by eliminating the need for explicit feature extraction. This paper focuses on snow and glacial feature classification using DL based approach and hyperspectral data. Based on EO-1 hyperion data, the classification accuracy of pixel-based data has been analyzed. The study comprises of three main steps: (1) Labeling of EO-1 hyperion data, obtained from field captured spectral signature; (2) Development of a novel DL framework where Fully Connected Deep Autoencoders (FCDAE) are employed for classifying seven snow and glacier features. Each model is trained on class-specific training data, allowing it to capture the unique characteristics and patterns associated with the corresponding class; (3) Performing comparative analysis of FCDAE framework with established Machine Learning (ML) techniques like Support Vector Machines (SVM) and Random Forests (RF). The results indicate superior performance of FCDAE, with exceptional overall accuracy of 97.12% and a kappa coefficient of 0.98. Our experiments on labeled dataset demonstrate that the proposed FCDAE framework achieves state-of–the-art performance against traditional ML frameworks.