Artifact detection in SAR images with AI methods

Abstract:
The increasing number of Earth observation data necessitates for advanced automated evaluation. Autoencoders (AE), which are deep neural networks, have been successfully applied to change detection on optical images. Here, we present an investigation of the applicability of three different convolutional AE methods for change detection on time series of SAR images. During the evaluation, the so-called joint AE approach is proved to be more precise and less sensitive to changes in brightness, thus designating less false positives. Moreover, the joint AE method indicates three noticeable and conspicuous regions.