Multi-Temporal Attention U-Net For Seasonally Variable Subsurface Feature Detection

The detection of subsurface features in optical satellite imagery is particularly challenging due to their faint and seasonally variable patterns caused by fluctuations in vegetation and soil moisture. In this study, we propose a deep learning framework that leverages multi-temporal observations to enhance the detection of subtle surface indicators (i.e. soil and vegetation marks) associated with subsurface features. This approach employs an attention-augmented U-Net architecture (TempNet3D) coupled with a temporal squeeze-and-excitation mechanism to dynamically reweight multiple seasonal inputs. The model processes three temporally distinct satellite images, representing different times of the year, and fuses them into a unified feature space that highlights the most informative cues from each observation. To further refine segmentation masks of narrow and fragmented features, we incorporated distance-based weight maps into the loss function. Compared to a standard U-Net trained on single-date inputs, our method shows consistently improved segmentation accuracy and more coherent delineation of target features across varied seasonal conditions. These findings underscore the value of incorporating temporal diversity into feature learning for earth observation tasks.