A Dual-Branches Multiscale Dynamic Partial Convolutional Attention Network for Remote Sensing Change Detection

Wenbin Tang, Shuli Cheng, Anyu Du

Abstract: Remote sensing change detection aims to accurately detect changes in buildings, roads, and other features in a pair of dual-temporal remote sensing images. In recent years, convolutional neural networks have achieved significant results in this task, but they lack the ability to model global features, leading to suboptimal performance in complex scenarios. To address this limitation, the Transformer, with its powerful global perception capability, has been introduced. However, directly incorporating the Transformer into the model often results in a significant increase in computational resource consumption. To address these issues, this article proposes a dual-branches multiscale dynamic partial convolution attention network, aimed at balancing both accuracy and computational efficiency. We introduce dynamic multiscale convolution attention module, which consists of multiscale context aggregation (MCA) module and dynamic partial convolution attention (DPCATT) module. The MCA module integrates features from different levels, while the DPCATT module enables global interaction between dual-temporal features, thereby enhancing the global modeling capability of the dual-branch features, while reducing the computing resources. Furthermore, we propose a parallel decoder aggregator module, which enhances spatial semantic information through a cascading and feature subtraction dual-branch strategy. In addition, a bottom-up training strategy is applied to strengthen the completeness of change detection. Extensive experiments on three publicly available datasets demonstrate that, compared to other methods, our proposed approach achieves superior performance and detection results, while reducing parameters and computational complexity.