Sensor-Adaptive Cross-Temporal Difference Modulation for Building Damage Detection in Bi-Temporal Remote Sensing Imagery

Post-event building damage assessment from bi-temporal satellite imagery is challenging because damaged pixels are sparse, structural changes can resemble radiometric variation, and sensor changes can dominate direct feature subtraction. This study evaluates SiameseCTDM with a Sensor-Adaptive Cross-Temporal Difference Module (SA-CTDM). At each encoder stage, SA-CTDM combines absolute and magnitude-normalized relative discrepancies through a learned per-channel gate, followed by channel recalibration. The gate provides adaptive interpolation between two discrepancy definitions; it is not treated as an independently superior accuracy component. A source-preserving class-conditional objective is also evaluated for optical-to-SAR adaptation by combining supervised BRIGHT learning, continued xBD supervision, and multi-stage alignment of intact and damaged prototypes. The adaptation terms are training-only and add no inference-time parameters. On the xBD hold set, complete SA-CTDM obtains 66.68 ± 0.54% F1, compared with 66.16 ± 1.12% for original CTDM, representing a modest 0.52 percentage-point mean increase with lower variance. A normalized-only ablation is the strongest single run, so an independent gating advantage is not claimed. A unified batch-one benchmark at 512×512 input resolution reports measured latency and throughput in addition to parameter counts and GFLOPs. On a fixed seed-42 sample-level BRIGHT split, the complete configuration using 611 labeled pairs reaches 34.33% F1, whereas the aligned 305-pair configuration does not outperform vanilla fine-tuning. The results therefore characterize split-specific and label-budget-dependent behavior rather than uniform few-shot superiority.