[Paper Review] RCDT: Relational Remote Sensing Change Detection with Transformer

Kaixuan Lu, Xiao Huang|arXiv (Cornell University)|Dec 9, 2022

Remote-Sensing Image Classification7 citations

TL;DR

RCDT proposes a simplified, high-performance remote sensing change detection framework using a Transformer-based architecture with a weight-sharing Siamese backbone, relational cross-attention for bi-temporal feature correspondence, and a features constraint module for high-resolution refinement. It achieves state-of-the-art performance across four public datasets, demonstrating superior accuracy and robustness through an efficient cross-attention mechanism that unifies semantic, correspondence, and attention functions.

ABSTRACT

Deep learning based change detection methods have received wide attentoion, thanks to their strong capability in obtaining rich features from images. However, existing AI-based CD methods largely rely on three functionality-enhancing modules, i.e., semantic enhancement, attention mechanisms, and correspondence enhancement. The stacking of these modules leads to great model complexity. To unify these three modules into a simple pipeline, we introduce Relational Change Detection Transformer (RCDT), a novel and simple framework for remote sensing change detection tasks. The proposed RCDT consists of three major components, a weight-sharing Siamese Backbone to obtain bi-temporal features, a Relational Cross Attention Module (RCAM) that implements offset cross attention to obtain bi-temporal relation-aware features, and a Features Constrain Module (FCM) to achieve the final refined predictions with high-resolution constraints. Extensive experiments on four different publically available datasets suggest that our proposed RCDT exhibits superior change detection performance compared with other competing methods. The therotical, methodogical, and experimental knowledge of this study is expected to benefit future change detection efforts that involve the cross attention mechanism.

Motivation & Objective

• To simplify complex AI-based change detection pipelines that rely on stacked modules for semantic enhancement, correspondence modeling, and attention mechanisms.
• To improve performance and generalizability in change detection by unifying three key functional components into a single, efficient architecture.
• To address limitations in existing methods related to model complexity, computational cost, and poor generalization in noisy or complex scenes.
• To demonstrate the effectiveness of cross-attention mechanisms in capturing long-range dependencies and bi-temporal feature correspondence in very high-resolution (VHR) images.

Proposed method

• A weight-sharing Siamese backbone is used to extract bi-temporal features from before-and-after images, ensuring parameter efficiency and feature consistency.
• The Relational Cross Attention Module (RCAM) applies offset-aware cross-attention to explicitly model correspondence between bi-temporal features, enhancing relation-aware representation learning.
• The Features Constrain Module (FCM) enforces high-resolution constraints during decoding, preserving spatial details and refining final predictions.
• A multi-scale training strategy is employed to improve feature learning, particularly for small objects and complex scenes.
• The model architecture avoids heavy convolutional stacks, relying instead on self- and cross-attention mechanisms to capture global context and long-range dependencies.
• Feature maps and attention maps are visualized to validate the model’s ability to focus on changed regions and maintain spatial fidelity.

Experimental results

Research questions

• RQ1 Can a unified, lightweight Transformer-based architecture outperform existing complex, multi-module deep learning pipelines in remote sensing change detection?
• RQ2 How effectively can cross-attention mechanisms capture bi-temporal feature correspondence without relying on separate attention or correspondence modules?
• RQ3 To what extent does the proposed RCDT generalize across diverse datasets with varying spatial resolutions and scene complexities?
• RQ4 Can the cross-attention mechanism improve detection accuracy for small or subtle changes in complex urban and natural environments?
• RQ5 How does the FCM module contribute to preserving high-resolution details in the final change detection output?

Key findings

• RCDT achieves second place on LEVIR-CD with a mean Intersection over Union (IoU) of 85.50, outperforming most existing methods.
• On the DSIFN dataset, RCDT attains an IoU of 69.89, demonstrating strong performance in complex urban change detection.
• RCDT achieves the highest performance on CDD with an IoU of 93.79, indicating excellent generalization on diverse and challenging change patterns.
• On SYSU-CD, RCDT achieves an IoU of 67.46, confirming its robustness across different geographical and land cover types.
• Visualizations confirm that the model effectively focuses on changed regions through high attention values in offset cross-attention maps.
• The ablation study validates that the integration of RCAM and FCM significantly improves detection accuracy and spatial precision.