Deep Learning on Hyperspectral Image Classification

Classification of Hyperspectral images is one of the main problems in the research field of Remote Sensing and other applications developed through computer vision. With the advantage of spectral and spatial information, it is possible to distinguish effectively different materials on the surface. Since last decade, the intensive employment of Convolutional Neural Networks (CNN) for classification and segmentation tasks led to high-quality results in the field of Hyperspectral Imagery Classification. However, these works are not able to perform satisfactorily on data acquired from various Hyperspectral Imaging Sensors. In this thesis, we propose a novel CNN architecture for HSI pixel-wise classification to improve the robustness and stability of the model to the data obtained from various sensors, thus giving state-of-the-art results. The proposed approach focuses on featureplayedtion through Dilated Convolution and Transposed Convolution. Moreover, the ELU activation function also played an essential role by activating the neurons with negative input values. Since, to face dataset imbalance problem, we adopt an oversampling strategy that increases the samples in minority classes. To prove the validity of the proposed framework, we tested it on five different HSI datasets and compared the performance with the most successful previous works. Training of the neural network has been performed on various ratios of the train, validation, and test data distribution. The evaluation of the model has been done by Three and Five-Fold cross-validation, and the performances have proven that our approach is competitive with the state-of-art and exhibits the best results on all the employed datasets, which prove that the proposed model is very robust under various Hyperspectral datasets irrespective of their characteristics.