DEVELOPMENT OF DEEP LEARNING APPROACHES TO EXTRACT INFORMATION FROM SATELLITE IMAGES

Abstract

This thesis titled ‘Development of Deep Learning Approaches to Extract Information from Satellite Images’ incorporates the study and critical analysis of the performance of conventional convolutional neural network (CNN) and recurrent neural network (RNN) models, and further development of novel CNN and RNN models to exploit the spectral and/or spatial and/or temporal information present in satellite images for applications such as land cover classification, disaster monitoring and mitigation, and agriculture crop monitoring. The main objective of this thesis is to pursue these studies with unimodal or multimodal satellite images. Special focus during these studies is to develop novel deep neural network (DNN) models (common term for CNN and RNN based models) that can achieve better classification accuracy and generalization performance in mixed land cover class scenarios. Four tasks are set to pursue these studies. The tasks set are based on: type of satellite image used, application for which the task is realized, multi-modality images involvement, and presence of time series information. The tasks to be carried out are; i) Study and assessment of polarization signatures as potential features for PolSAR image based land cover classification and further development of novel polarization signatures and CNNs based approaches to achieve improved classification performances with PolSAR satellite images, ii) Development of a novel approach using polarization signatures and CNNs for detection of human settlements affected during floods, iii) Development of novel spectral and spectral-spatial CNN models for multisensor (PolSAR and multispectral) images based land cover classification, and iv) Development of novel CNN, RNN, and CNN-RNN models for multisensor (PolSAR and multispectral) satellite time series images based crop classification.