Explainable Convolutional Autoencoders for Enhanced Representation learning in hyper spectral images

Abstract

Abstract The categorization of pixels in hyperspectral pictures into distinct classifications, such as kinds of land cover, satellite, crops, and urban areas, is a crucial task in remote sensing and image analysis. Due to its capacity to extract spatial and spectral information from hyperspectral data, convolutional autoencoders (CAEs) are a class of neural network design that may be used for this job. In this abstract a technique for hyperspectral image categorization using CAEs is discussed. A CAE is trained as an unsupervised feature extractor, where the encoder learns to extract useful spectral and spatial features from the hyperspectral data and the decoder learns to reconstruct the original data from the lower-dimensional latent representation. When the CAE has been trained, features from the hyperspectral pictures are extracted using the encoder, and a classifier is trained on top of these features for supervised classification.