Building Change Detection Using Modified Siamese Neural Networks

Abstract

In recent years, deep learning has emerged as an effective tool for analyzing remote sensing imagery. One such application is change detection, where locations that have changed over time must be identified. This work improves upon previous deep learning methods of change detection in buildings by adding a simple yet effective module after concatenation of feature maps in Siamese Neural Networks. Named the Post Concatenation Module (PCM), the module comprises of convolution, activation, and normalization layers that increase AUC values up to 1.5 points for the task at hand. Furthermore, experiments demonstrate the addition of operations before concatenation of feature maps decreases AUC values, signifying the importance of introducing operations after combing the features of the inputs in Siamese Neural Networks.