Comparative Analysis of Learning-Based Approaches for Change Detection in Satellite Images

Abstract

Satellite image change detection, where two images of the same area from different times are compared, is crucial for earth sensing and monitoring applications. Many learning-based detection methods have been proposed for this task, with different performance characteristics. Since these detection methods have been tested under different settings, comparing their performance across a variety of situations is difficult. The goal of this article is therefore to comprehensively compare the state-of-the-art detection methods from the literature, across a variety of dataset parameters. To that end, we analyze the impact of image resolution, training set size, and noise on learning performance. A first set of experiments, using a large set of high-resolution images, reveals that training set resolution should match the resolution of the images the model will be applied to, that larger training sets are beneficial, and that adding Gaussian noise improves performance. A second set of experiments, using a smaller set of low-resolution images, confirms that the training set should also be of the same low resolution, but shows that adding noise does not improve performance in this case. The results also indicate that BiasUNet is the most effective method for detecting changes between image pairs.