Building Extraction in High Spatial Resolution Images Using Deep Learning Techniques

Abstract

High spatial resolution images are processed in the object domain as the traditional pixel-based method processes individual pixels (layer by layer) and classifies them, thus ignoring the neighborhood or contextual features. Analysis in object domain includes three steps: segmenting the image into homogeneous regions/objects, extracting features and assigning class labels to each of these regions based on the extracted features. Object-based analysis of an image faced challenges such as identifying the appropriate scale for segmentation and incapability to capture complex features that a high resolution image entails. This paper aims to solve this challenge by using a deep learning technique called Region-based Convolutional Neural Networks (R-CNN). Faster R-CNN was used here for the extraction of buildings in satellite images. The dataset used for training and testing was WorldView-2 with spatial resolution of 0.46 m. The results obtained using faster R-CNN had classification accuracy of 99% with 2000 epochs whereas building extraction using support vector machine showed 88.3%. The results obtained clearly indicate that convolutional neural networks are better at extracting features and detecting objects in high resolution images.