Optimizing Deep Learning Architecture for Remote Sensing Image Analysis

Abstract

Deep Learning tools have become very efficient in high-resolution image analysis compared to traditional classification models. One such example is the implementation of semantic segmentation using a Convolutional Neural Network (CNN). Unlike image labeling, where images are classified into one label, we can use semantic segmentation to identify the class labels of every pixel in an image. This makes CNN an ideal tool for Land Use Land Cover (LULC) modelling. This is especially true because current land cover classification techniques require a lot of time and resources to complete. This project attempts to create a deep learning architecture for resource-constrained environments by reducing complex mathematical operations that plague the deployment of CNN. The proposed model will be trained using the Potsdam dataset and the Vaihingen dataset obtained from the International Society for Photogrammetry and Remote Sensing (ISPRS). Using the concept of transfer learning, the trained model will then be used to compare and assess the LULC change dynamics for the lower Chippewa Valley watershed region in Wisconsin.