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. Even a model achieving an impressive 89% aggregate F1-score can still be systematically wrong exactly in the regions where accuracy is most critical: building boundaries, shadow margins, and active construction sites. 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 − α 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. This constitutes the first formal characterisation of confidence distribution structure in remote-sensing change detection.

Key Methodologies & Contributions

  • Conformal Prediction Wrapper: Deployed Marginal Split Conformal Prediction on top of the pre-trained Bitemporal Image Transformer (BIT) to generate distribution-free, statistically rigorous 1-α coverage bounds for individual pixel-level predictions without retraining model weights.
  • Optimized BIT Implementation: Corrected a widespread normalization error in the BIT training pipeline to recover full network accuracy, leading to state-of-the-art performance of F1=89.94% and IoU=81.72% on the LEVIR-CD benchmark dataset.
  • Bimodal Confidence Characterization: Exposed a previously undescribed bimodal confidence structure in geospatial transformers through conformal calibration, where spatial ambiguity in boundary zones and active construction sites emerges as Empty Prediction Sets (EPS) rather than dual-class uncertainty.
  • Operational Reliability Metrics: Provided a formally justified, computationally lightweight system enabling automated flagging of geometrically ambiguous pixels (the 10% EPS tail) for human-in-the-loop disaster assessment and geospatial intelligence workflows.


Code & Resources

Status: Under Review at ICCI, 2026. Authors: A. Mukherjee, A. Kumar, S. Bala, S. R. Verma, A. Anand, H. Das, L. Chhetri