Network and Dataset for Multiscale Remote Sensing Image Change Detection
Shenbo Liu, Dongxue Zhao, Yuheng Zhou, Tan Ying, He Huang, Zhao Zhang, Lijun Tang
Abstract
Remote sensing image change detection (RSCD) aims to identify differences between remote sensing images of the same location at different times. However, due to the significant variations in the size and appearance of objects in real-world scenes, existing RSCD algorithms often lack strong capabilities in extracting multiscale features, thereby failing to fully capture the characteristics of changes. To address this issue, a multiscale remote sensing change detection network (MSNet) and a multiscale RSCD dataset (MSRS-CD) are proposed. A multiscale convolution module (MSCM) is investigated, and combined with MSCM, an encoder capable of capturing features of different sizes is designed to efficiently extract multiscale semantic change features. A global multiscale feature fusion module is designed to achieve global multiscale feature fusion and obtain multiscale high-level semantic change features. As existing RSCD datasets lack rich scale information and often focus on change targets of specific sizes, a new dataset, MSRS-CD, is constructed. This dataset consists of 842 pairs of images with a resolution of 1024 × 1024 pixels, featuring uniformly distributed change detection target sizes. Comparative experiments are conducted with 10 other state-of-the-art algorithms on the MSRS-CD dataset and another public dataset, LEVIR-CD. Experimental results demonstrate that MSNet achieves the best performance on both datasets, with an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</i>1 score of 75.74% on the MSRS-CD dataset and 91.41% on the LEVIR-CD dataset.