Generative Adversarial Networks Under CutMix Transformations for Multimodal Change Detection
Anamaria Rădoi
Abstract
The current technological developments lead to increased heterogeneity and variability in remote sensing imagery. In this context, unsupervised multimodal change detection techniques are mandatory to perform a continuous monitoring and rapid damage assessment by means of heterogeneous remote sensing data. Taking advantage of the latest advances in deep learning, we address multimodal change detection from an inter-modality image translation perspective. Inter-modality translation is achieved by means of generative adversarial networks built over U-Net architectures at both generator and discriminator levels and trained under CutMix transformations. A change prior is used to guide the learning process of the neural network framework and to reduce the impact of changed locations over the learned model. The change prior is derived in an unsupervised manner from comparisons between the post-event locations and <i>k</i> nearest neighbor locations determined in the pre-event image. The experiments were conducted over several pairs of heterogeneous remote sensing images, and the comparisons with current state-of-the-art approaches show the effectiveness of the proposed multimodal change detection framework.