A Change Detection Reality Check

TL;DR

This work interrogates progress in bi-temporal change detection by re-evaluating state-of-the-art methods under a fair, controlled setup. It demonstrates that a simple U‑Net baseline with pretrained backbones can outperform many recent architectures on LEVIR‑CD and WHU‑CD, even when newer models claim improvements. The authors also reassess WHU‑CD using original train/test splits and multiple seeds, highlighting variability and potential data leakage in some benchmarks. They argue for standardized benchmarking libraries (OpenCD, GEO‑Bench, TorchGeo) to ensure reliable progress and guide future research in change detection.

Abstract

In recent years, there has been an explosion of proposed change detection deep learning architectures in the remote sensing literature. These approaches claim to offer state-of-the-art performance on different standard benchmark datasets. However, has the field truly made significant progress? In this paper we perform experiments which conclude a simple U-Net segmentation baseline without training tricks or complicated architectural changes is still a top performer for the task of change detection.