Verifying Out-of-Distribution Robustness in Multi-spectral Satellite Change Detection

Swarnava Dey, Pallab Dasgupta, Partha Pratim Chakrabarti

Keywords: change detection, certified robustness, out-of-distribution generalization, multi-spectral remote sensing, encoder–decoder architectures

TL;DR: We show why certified robustness collapses in satellite change detection and introduce a simple head-consistency training plus a diagnostic verifier and CropRot benchmark to achieve practical OOD stability.

Abstract:

Reliable multi-spectral change detection on-board satellites requires robustness under distribution shifts. We address this challenge from both the certification and empirical perspectives.

On the certification side, we adapt neural verification to the unique structure of change detection, accounting for sensor noise, encoder–decoder heads, and semantic evaluation. We introduce a tail-tapped verifier that transports input intervals to the final decoder tap and applies $\alpha$-CROWN solely to the decision head. This yields per-pixel logit-margin lower bounds, which we summarize through task-aligned predicates such as coverage, false positives, and minimum island size.

On the empirical side, we study out-of-distribution robustness across three representative backbones — U-Net style encoder–decoder (FresUNet), lightweight convolutional attention encoder–decoder (FALCONet), and transformer-inspired global attention encoder–decoder (AttU-Net) — on the Onera Satellite Change Detection (OSCD) dataset. We find that existing certificates vanish even for mild perturbations ($\varepsilon \ge 1/255$), while empirical robustness varies widely across architectures.

Our results highlight both the difficulty of certifying change detection and the promise of architecture design for achieving practical robustness. This establishes a foundation for principled verification and stress-tested deployment of satellite-based change detection models

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 2458