Beyond 2D and more:|bInterpreting remote sensing image classification methods via explainable artificial intelligence

Official URL: https://risc01.sabanciuniv.edu/record=b3395411

Within the hyperspectral remote sensing image classification research area, this thesis delves into the challenges of explaining the decision-making process of deeplearning models. The focus is on the integration of three prominent explainable artificial intelligence methods, namely Grad-CAM, Grad-CAM++, and Guided Backpropagation. These methods have been employed in order to comprehend the decision-making process of a typical convolutional neural network model during spatial-spectral hyperspectral image classification. The conducted experiments investigate the impact of varying pixel patch sizes on spatial attention and the significance of individual spectral bands in the classification process. This thesis sheds light on the behavior of convolutional neural networks in the spatial-spectral context, providing a deeper understanding of how these models respond to changes in hyperspectral data. Furthermore, the study analyzes the relative advantages and limitations of the employed explainability techniques —Grad-CAM, Grad-CAM++, and Guided Backpropagation— in explaining the decision-making processes of the convolutional neural network model. In conclusion, the results provide both deeper interpretations of the behavior of convolutional neural networks as well as a comparative performance analysis of explainability techniques.