Risk-Controlled Urban Change Detection: Conformal Prediction Wrappers for Provable Reliability in High-Resolution Satellite Imagery

Abstract

Satellite-based urban change detection enables an emerging generation of geospatial intelligence applications, from disaster damage assessments to spatial expansion monitoring; however, the deep learning models powering these systems lack formal reliability metrics for individual predictions. This paper introduces a rigorous statistical approach to transformer-based satellite change detection using Marginal Split Conformal Prediction (CP). Working on top of the pre-trained Bitemporal Image Transformer (BIT) and the well-established LEVIR-CD benchmark, we calibrate a non-conformity threshold on the validation split and generate pixel-wise prediction sets. This achieves distribution-free, statistically rigorous 1-alpha coverage bounds without requiring any re-training of the model weights. Our optimized implementation attains F1=89.94% and IoU=81.72% on the LEVIR-CD test set, outperforming all published baselines. Furthermore, conformal calibration reveals a novel structural property specific to geospatial transformers: BIT operates in a near-binary confidence regime where spatial ambiguity emerges as Empty Prediction Sets rather than dual-class uncertainty.

Key Methodologies & Contributions

• Provable Statistical Reliability: Introduced a rigorous statistical reliability framework for transformer-based change detection by deploying Split Conformal Prediction, empirically validating that the 1-alpha coverage guarantee holds on real-world satellite imagery.
• Bimodal Confidence Discovery: Exposed a previously undescribed bimodal confidence structure inside BIT through conformal calibration, where uncertain pixels produce Empty Prediction Sets, carrying direct implications for human-in-the-loop deployment pipelines.
• Optimized BIT Implementation: Corrected a widely propagated normalization error prevalent in BIT reimplementations (which incorrectly fed color-distorted inputs using ImageNet statistics), cleanly recovering full network accuracy without architectural changes.
• SOTA Benchmark Performance: Achieved state-of-the-art results on the LEVIR-CD benchmark (F1=89.94% and IoU=81.72%), exceeding the originally reported numbers and outperforming the strongest purely convolutional baselines like FC-Siam-Diff.

Code & Resources

• The official PyTorch implementation and pre-trained weights will be made publicly available on GitHub following the conclusion of the double-blind review process.