Bridging Supervision Gaps: A Unified Framework for Remote Sensing Change Detection

TL;DR

UniCD proposes a unified, supervision-adaptive framework for remote sensing change detection that jointly supports supervised, weakly-supervised, and unsupervised tasks through a shared encoder and three specialized branches. STAM enables efficient spatial-temporal fusion in the supervised path, CRR imposes spatial coherency and contrastive constraints to refine weakly-supervised change representations, and SPCI leverages CLIP and FastSAM priors to generate pseudo-labels for unsupervised CD. Across LEVIR-CD, WHU-CD, and CLCD, UniCD achieves state-of-the-art results in all three settings, with substantial gains in weakly- and unsupervised scenarios (e.g., improvements of 12.72% and 12.37% on LEVIR-CD). The framework demonstrates strong generalization by transforming unsupervised signals into controlled weakly-supervised optimization, offering practical impact for scalable change monitoring under annotation scarcity.

Abstract

Change detection (CD) aims to identify surface changes from multi-temporal remote sensing imagery. In real-world scenarios, Pixel-level change labels are expensive to acquire, and existing models struggle to adapt to scenarios with diverse annotation availability. To tackle this challenge, we propose a unified change detection framework (UniCD), which collaboratively handles supervised, weakly-supervised, and unsupervised tasks through a coupled architecture. UniCD eliminates architectural barriers through a shared encoder and multi-branch collaborative learning mechanism, achieving deep coupling of heterogeneous supervision signals. Specifically, UniCD consists of three supervision-specific branches. In the supervision branch, UniCD introduces the spatial-temporal awareness module (STAM), achieving efficient synergistic fusion of bi-temporal features. In the weakly-supervised branch, we construct change representation regularization (CRR), which steers model convergence from coarse-grained activations toward coherent and separable change modeling. In the unsupervised branch, we propose semantic prior-driven change inference (SPCI), which transforms unsupervised tasks into controlled weakly-supervised path optimization. Experiments on mainstream datasets demonstrate that UniCD achieves optimal performance across three tasks. It exhibits significant accuracy improvements in weakly and unsupervised scenarios, surpassing current state-of-the-art by 12.72% and 12.37% on LEVIR-CD, respectively.

Figures (4)

• Figure 1: Our proposed UniCD is a supervision-adaptive framework that supports supervised, weakly-supervised, and unsupervised CD tasks in a unified manner.
• Figure 2: Framework of UniCD. It consists of three branches. The supervised branch employs STAM for pixel-level prediction. In the weakly-supervised branch, UniCD is refined by SCR and CFR. SCR enforces geometric consistency under view perturbations to learn stable, invariant features. CFR uses CAMs as semantic anchors to align unchanged features and separate changed features. In the unsupervised branch, SPCI leverages semantic priors from foundation models and integrates feature responses to extract pseudo-labels, transforming the unsupervised task into CRR-constrained optimization within the weakly-supervised branch.
• Figure 3: The overall structure of STAM, which can effectively model spatial structures and temporal variations.
• Figure 4: Qualitative visual analysis of LEVIR-CD under three supervision patterns. False positive (erroneously changed) pixels are marked in red, while false negative (erroneously unchanged) pixels are marked in blue.