UNDERSTANDING MULTISPECTRAL SATELLITE IMAGE LAND COVER CLASSIFICATION WITH THE DEEP LEARNING

As satellite images provide periodical observations of a large area, Remote Sensing (RS) data is important for analyzing Land Use and Land Cover (LULC). In recent years, with the advancement of Deep Learning, better LULC classification and prediction are achieved via artificial neural networks (ANNs), such as convolutional neural networks (CNNs). As land cover classifications usually apply spatial and spectral properties, the objective of this research is to design a deep learning network to retrieve spatial and spectral features from remote sensing images and perform the land cover classification. To be specific, a CNN for extracting textural information is combined with a network extracting cross-band spectral relationships for the final classification. The thirteen bands of EuroSAT dataset are applied in this research, where more than 90% accuracy can be achieved. An ongoing work is to identify important factors for a better understanding to the remote sensing classification, and compare with current neural network achieves state-of-the-art results on EuroSAT dataset with 98.65% accuracy.