A multi-scale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas

Abstract

Accurate change detection of built-up areas (BAs) fosters a comprehensive understanding of urban development. The post-classification comparison (PCC) is a widely-used change detection method by classification and temporal comparison. For classification, image-level labeling is an efficient alternative to pixel-level one for pixel-wise weakly supervised segmentation, which frequently applies pixel-level pseudo labels generated from class activation map (CAM) to train semantic segmentation networks. CAM can be obtained from classification networks trained with image-level labels and can indicate the spatial location of objects. The existing studies are subject to the following issues: 1) They only rely on the single-scale and low-resolution CAM, but ignore the multi-scale property of BAs; 2) Pixel-level pseudo labels usually contain noises (e.g., omissions and false alarms); 3) The temporal correlation between multi-temporal images is less considered in PCC. To address these limitations, this paper proposed a multi-scale weakly supervised learning method, which utilized a large number of single-temporal high-resolution images and image-level labels to detect BA changes. This method consisted of three modules: 1) multi-scale CAM for BA pseudo label generation; 2) adaptive online noise correction for BA detection; and 3) generation of reliable pseudo labels for BA change detection. Based on ZY-3 images (2.5 m), we constructed the first multi-view datasets for both BA detection and change detection. Each ZY-3 image includes a multi-spectral image with red, green, blue, and near-infrared bands and a multi-view image with nadir-, forward-, and backward-views. The BA detection dataset contained 86,166 image-level samples (256 × 256 pixels for each sample), covering 48 major cities in China, while the BA change detection dataset consisted of ZY-3 bi-temporal images at rapidly urbanizing areas (i.e., Beijing and Shanghai). Experiments showed that the proposed method can detect BA changes and suppress pseudo changes effectively, yielding 88.2% F1-score in BA detection and 79.3% for Shanghai and 78.5% for Beijing in change detection. Further analysis demonstrated the proposed method to be advantageous in the following two fronts: 1) the image-level weak labels can achieve pixel-wise BA change detection at low cost; and 2) the multi-scale CAM and temporal correlation are effective in the scenarios with limited labels. Datasets and codes will be accessed at https://github.com/lauraset/MSWS.