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

This repository contains the official implementation and open-source assets for our paper on deployment-ready, distribution-free uncertainty estimation in remote sensing. By applying Marginal Split Conformal Prediction to the Bitemporal Image Transformer (BIT), we provide pixel-wise prediction sets with statistically rigorous $1-\alpha$ coverage bounds without requiring any model retraining or weight modifications.

---

Key Contributions

• Rigorous Reliability: First framework to deploy Split Conformal Prediction for bi-temporal satellite image change detection.
• Structural Insight: Discovery of a near-binary confidence regime in geospatial transformers, where spatial ambiguity manifests as Empty Prediction Sets (EPS) rather than dual-class uncertainty.
• Performance Boost: Corrected a widely propagated normalization error in public BIT pipelines, achieving F1 = 89.94% and IoU = 81.72% on the LEVIR-CD test set.

---

System Architecture

Our pipeline feeds bi-temporal image pairs through a shared ResNet-18 Siamese encoder and a semantic tokenizer. The resulting softmax probabilities are wrapped by a conformal validation stage to generate reliable prediction sets at inference time.

---

Key Results

Empirical Coverage vs. Theoretical Target

Our implementation demonstrates a tight empirical match ($\pm 0.05$ pp) across all tested error bounds ($\alpha$).

$\alpha$	Target Coverage	Calibrated Threshold ($\hat{q}$)	Empirical Coverage	Uncertain Pixels
0.05	95.00%	0.0118	94.99%	0.00%
0.10	90.00%	0.0021	90.02%	0.00%
0.15	85.00%	0.0011	85.02%	0.00%
0.20	80.00%	0.0008	79.95%	0.00%

---

Quick Start (Google Colab)

You can run our inference pipeline directly in your browser without any local setup. Click the link below to open our verified Jupyter Notebook:

---

Authors & Contributors

• Ankona Mukherjee
• • Aman Kumar
• Latchan Chhetri & more..

---

Citation

If you find this work useful in your research, please consider citing our paper:

@inproceedings{chhetri2026risk,
  title={Risk-Controlled Urban Change Detection: Conformal Prediction Wrappers for Provable Reliability in High-Resolution Satellite Imagery},
  author={Chhetri, Latchan and Mukherjee, Ankona and Kumar, Aman and Sarma, Gaurav},
  booktitle={International Conference on Computational Intelligence: Machine Learning for Geospatial Analytics (ICCI)},
  year={2026}
}