Date of Award
Deep Learning, Machine Learning, Satellite Images, Layer-wise Relevance Propagation, Image classification, Convolutional Neural Networks
The study of satellite images provides a way to monitor changes in the surface of the Earth and the atmosphere. Convolutional Neural Networks (CNN) have shown accurate results in solving practical problems in multiple fields. Some of the more recognized fields using CNNs are satellite imagery processing, medicine, communication, transportation, and computer vision. Despite the success of CNNs, there remains a need to explain the network predictions further and understand what the network is determining as valuable information.
There are several frameworks and methodologies developed to explain how CNNs predict outputs and what their internal representations are [1, 4, 12, 25, 27, 30, 32]. A technique developed by Bash et al.  called Layer-Wise Relevance Propagation (LRP) back propagates the resulting classification scores through the network until it reaches the input layer. The resultant scores at the input layer are called the relevance of the image and represent the contribution per pixel towards the final classification. LRP is used in our work to optimize a proposed network and moreover, analyze the relevance of an image to determine features that are essential for the classification of satellite images.
Given the nature of satellite images, it is not possible to rely on subjective descriptions of relevant regions. A novel technique is developed in this Thesis to overcome the challenge. This technique allows us to describe qualitatively and quantitatively the regions that are shown as relevant by LRP. The technique relies on the analysis of the Near Infra-red (NIR) band and its relevant areas to understand the classification given by the network. Regions of vegetation are found to be relevant, and its significance is measured when classifying categories like Trees, Grasslands, Roads, Buildings, and Barren Lands.
Patino Vargas, Yamile, "Towards Improving Accuracy and Interpretability of Deep Learning based on Satellite Image Classification" (2019). CUNY Academic Works.