Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ViLaCD-R1: A Vision-Language Framework for Semantic Change Detection in Remote Sensing

Xingwei Ma, Shiyang Feng, Bo Zhang, Bin Wang

TL;DR

RSCD faces challenges in achieving semantic-level change understanding while maintaining pixel-precise localization. ViLaCD-R1 addresses this with a two-stage approach: a fine-tuned Vision–Language Model (MIR) performs semantic, patch-level reasoning to produce a coarse change mask, which then guides a Mask-Guided Decoder (MGD) to refine to a precise binary change map $M \in \{0,1\}^{H\times W}$. The MIR is trained via supervised fine-tuning (SFT) and reinforcement learning with a GRPO-based objective to maximize cross-temporal consistency, while the MGD employs a local-window Transformer and soft mask guidance to enhance boundary accuracy. Comprehensive experiments on LEVIR-CD, LEVIR-CD+, and SYSU-CD show state-of-the-art performance with robust localization and reduced non-semantic perturbations, demonstrating strong cross-scene generalization and practical impact for high-resolution remote sensing change detection.

Abstract

Remote sensing change detection (RSCD), a complex multi-image inference task, traditionally uses pixel-based operators or encoder-decoder networks that inadequately capture high-level semantics and are vulnerable to non-semantic perturbations. Although recent multimodal and vision-language model (VLM)-based approaches enhance semantic understanding of change regions by incorporating textual descriptions, they still suffer from challenges such as inaccurate spatial localization, imprecise pixel-level boundary delineation, and limited interpretability. To address these issues, we propose ViLaCD-R1, a two-stage framework comprising a Multi-Image Reasoner (MIR) and a Mask-Guided Decoder (MGD). Specifically, the VLM is trained through supervised fine-tuning (SFT) and reinforcement learning (RL) on block-level dual-temporal inference tasks, taking dual-temporal image patches as input and outputting a coarse change mask. Then, the decoder integrates dual-temporal image features with this coarse mask to predict a precise binary change map. Comprehensive evaluations on multiple RSCD benchmarks demonstrate that ViLaCD-R1 substantially improves true semantic change recognition and localization, robustly suppresses non-semantic variations, and achieves state-of-the-art accuracy in complex real-world scenarios.

ViLaCD-R1: A Vision-Language Framework for Semantic Change Detection in Remote Sensing

TL;DR

RSCD faces challenges in achieving semantic-level change understanding while maintaining pixel-precise localization. ViLaCD-R1 addresses this with a two-stage approach: a fine-tuned Vision–Language Model (MIR) performs semantic, patch-level reasoning to produce a coarse change mask, which then guides a Mask-Guided Decoder (MGD) to refine to a precise binary change map . The MIR is trained via supervised fine-tuning (SFT) and reinforcement learning with a GRPO-based objective to maximize cross-temporal consistency, while the MGD employs a local-window Transformer and soft mask guidance to enhance boundary accuracy. Comprehensive experiments on LEVIR-CD, LEVIR-CD+, and SYSU-CD show state-of-the-art performance with robust localization and reduced non-semantic perturbations, demonstrating strong cross-scene generalization and practical impact for high-resolution remote sensing change detection.

Abstract

Remote sensing change detection (RSCD), a complex multi-image inference task, traditionally uses pixel-based operators or encoder-decoder networks that inadequately capture high-level semantics and are vulnerable to non-semantic perturbations. Although recent multimodal and vision-language model (VLM)-based approaches enhance semantic understanding of change regions by incorporating textual descriptions, they still suffer from challenges such as inaccurate spatial localization, imprecise pixel-level boundary delineation, and limited interpretability. To address these issues, we propose ViLaCD-R1, a two-stage framework comprising a Multi-Image Reasoner (MIR) and a Mask-Guided Decoder (MGD). Specifically, the VLM is trained through supervised fine-tuning (SFT) and reinforcement learning (RL) on block-level dual-temporal inference tasks, taking dual-temporal image patches as input and outputting a coarse change mask. Then, the decoder integrates dual-temporal image features with this coarse mask to predict a precise binary change map. Comprehensive evaluations on multiple RSCD benchmarks demonstrate that ViLaCD-R1 substantially improves true semantic change recognition and localization, robustly suppresses non-semantic variations, and achieves state-of-the-art accuracy in complex real-world scenarios.
Paper Structure (21 sections, 12 equations, 4 figures, 5 tables)

This paper contains 21 sections, 12 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Feature-based Change Detection Methods
  4. VLM-based Change Detection Methods
  5. VLM for multi-image reasoning
  6. Method
  7. Preliminary
  8. Problem Definition
  9. RL for fine-tuning VLM
  10. Architecture
  11. Overview
  12. MIR Training
  13. MGD Structure
  14. Reward Design
  15. Decoder Design
  16. ...and 6 more sections

Figures (4)

  • Figure 1: (a) illustrates the feature-based paradigm for remote sensing change detection; (b) shows the VLM-based paradigm; and (c) presents our proposed ViLaCD-R1 architecture, which provides a complete pipeline from coarse region prompting to fine-grained pixel-level segmentation.
  • Figure 2: The overall architecture of ViLaCD-R1. The MIR module first generates coarse change masks for the bi-temporal images. These masks are then used to guide MGD-Net, enabling fine-grained, pixel-level boundary refinement.
  • Figure 3: GRPO training workflow
  • Figure 4: Visulization results of the proposed method on the LEVIR-CD dataset