Classification of deep learning based hyperspectral satellite images in remote sensing

Classification in hyperspectral remote sensing images (HRSIs) is a challenging process in image analysis and one of the most popular topics. In recent years, many methods have been proposed to solve the HRSIs classification problem. Compared to traditional machine learning methods, deep learning, especially Convolutional Neural Networks (CNNs), is commonly used in the classification of HRSIs. Deep learning-based methods based on CNNs show remarkable performance in HRSIs classification and greatly support the development of classification technology. In this thesis, seven different deep learning-based methods have been developed for HRSI classification. In the first study, a hybrid method is proposed in which 3D convolution and 2D depthwise separable convolution are used together. In the second study, a new 3D CNN-based method was developed to extract spatial-spectral features. In the third study, a 3D CNN-based LeNet5 method with less trainable parameters was developed. In the fourth study, a hybrid 3D Residual spatial-spectral convolution network is proposed for extraction of deep spatial-spectral features using 3D CNN and ResNet18 architecture. In the fifth study, a hybrid method in which the 3D/2D Complete Inception module and the 3D/2D CNN method are used together is proposed to solve the HRSI classification problem. In the proposed method, multi-level feature extraction is performed by using multiple convolution layers with the Inception module. In the sixth study, LeNet5, AlexNet, VGG16, GoogleNet and ResNet50 architectures, which are among the successful examples of CNN, are used for the HRSI classification problem. A hybrid approach based on 3D CNN is used when using these architectures. In the seventh study, a hybrid 3D-2D depthwise separable convolutional network based deep learning method based on multipath feature fusion is developed. All of the proposed methods have been tested on frequently used IP, PU, SA, BO, HL, UH, KUM, and recently emerged WHU-Hi datasets. The obtained classification results with the proposed methods reveal that the proposed methods provide more successful classification performance than the state-of-the-art methods.