Urban building change detection with satellite imagery: a temporospatial attention Siamese SNUNet-ECAM framework

Abstract

Abstract Urban building change monitoring on bi-temporal Very High Resolution (VHR) satellite imagery is critical for land-use monitoring, post-disaster assessment, and urban planning. We use LEVIR-CD, a public benchmark of 637 bi-temporal 1024 $$\times$$ 1024-pixel image pairs at 0.5 m/pixel resolution, collected from Google Earth imagery of 20 regions in Texas, USA, with a 5–14-year interval between acquisitions and a $$\sim$$ 4.6% building-change pixel ratio. Rather than proposing a new architectural module, this paper presents a systematic integration of a Siamese ResNet50 encoder, the Temporospatial Interactive Attention Module (TIAM), and the SNUNet dense nested decoder with Ensemble Channel Attention (ECAM) — into a single end-to-end trainable pipeline, evaluated on the official LEVIR-CD split. The central empirical finding, based on a controlled ablation, is that full-resolution image tiling (versus resizing) is the dominant driver of performance on this benchmark, improving F1 by 17.11 percentage points (73.24% $$\rightarrow$$ 90.35%) — an effect several times larger than the combined, measured contribution of the three architectural modules (TIAM +6.83 SNUNet +5.31, ECAM +3.87 F1%). We additionally document three previously unreported training instabilities (NaN propagation from Lovász-Hinge loss, silent fp16 overflow in TIAM’s attention computation, and catastrophic forgetting without progressive unfreezing) along with the engineering fixes that resolve them. Using the decision threshold selected on the validation set (0.56, avoiding test-set threshold tuning and test-set augmentation), the system reaches F1 90.35%, IoU 82.40%, Precision 91.70%, Recall 89.04%, and Overall Accuracy 99.03% on the official LEVIR-CD test split—essentially matching, rather than exceeding, a simpler ResNet-50 U-Net baseline (90.38% F1) and remaining below ChangeFormer (90.40% F1). Notably, our model achieves this comparable performance with substantially fewer parameters than ChangeFormer, demonstrating a more parameter-efficient architecture without sacrificing predictive performance. We report this honestly as evidence that, on this benchmark, correct preprocessing can be as important as architectural sophistication.