TRSANet: A Remote Sensing Deep Learning Model for Water Body Change Detection Based on Time-Reversal Semantic Asymmetry

The goal of change detection is to identify modifications in dual-temporal remote sensing (RS) images, which reveal dynamic changes in specific areas of interest. Specifically, the detection of water body changes via RS is critical for monitoring environmental shifts, such as droughts and floods. However, RS methods that utilize deep learning (DL) to detect water changes face two challenges: capturing the unique spatial-temporal features of changes in water bodies and effectively modeling temporal features and spatiotemporal coupling. Drawing inspiration from the theory of time-reversal asymmetry in physics and considering the temporal characteristics of RS water bodies, we propose the time-reversal semantic asymmetry module to extract essential temporal-spatial features more effectively. Consequently, this article presents the design of a model specifically tailored for detecting changes in water bodies, termed time-reversal semantic asymmetry network (TRSANet). TRSANet outperforms other state-of-the-art methods by not only having fewer parameters but also achieving the highest
$F1$
score and IOU (91.71% and 84.69%, respectively) on the Water-CD dataset for water body change detection. Additionally, TRSANet demonstrates remarkable cross-domain adaptability by attaining the highest accuracy on the LEVIR-CD dataset for building change detection. In the future, TRSANet will also offer methodological support for urban planning and the establishment of water body conservation areas.