PSTNet: Progressive Sampling Transformer Network for Remote Sensing Image Change Detection

Abstract

Remote sensing change detection (CD) is to use multitemporal remote sensing data to extract change information by using a variety of image processing and pattern recognition methods, and quantitatively analyze and determine the characteristics and processes of surface changes. In recent research on CD, how to more accurately segment objects and how to extract and effectively link spatiotemporal information are important parts. To achieve this, we propose a progressive sampling (PS) transformer network for remote sensing image CD, which continuously extracts and optimizes feature information in an iterative manner, so that pixels can establish better connections in the spatial domain to model the context. Our intuition is that, through this iterative sampling method, the parts of interest in the image can be gradually extracted. This allows subsequent processing to be more focused on useful areas, which in turn reduces interference from uninteresting parts, and the information after PS will be generalize into several tokens containing rich semantic information. Using the excellent modeling ability of the transformer, the optimized tokens are mapped back to the original image features to achieve the purpose of segmenting accurate difference images. We conducted extensive experiments on three CD datasets, LEVIR-CD, DSIFN-CD, and WHU-CD, and achieved evaluation scores of 90.73/84.11, 80.10/68.93, and 91.67/85.15 on F1-score and IoU metrics, respectively. Notably, the convolutional neural network (CNN) backbone of our network uses only a simplified ResNet model, without using structurally complex frameworks, such as FPN and Unet, but our model uses PS module and transformer to achieve better performance than the recent advanced CD models.

Keywords

• Attention mechanisms
• change detection (CD)
• deep convolutional network
• dual-time satellite remote sensing images
• progressive sampling (PS)
• transformers