Unified building change detection pre-training method with masked semantic annotations

• DOI: 10.1016/j.jag.2023.103346
• Publication Date: 2023-05-15T20:26:11Z
• Authors: Yujun Quan; Anzhu Yu; Wenyue Guo; Xuanbei Lu; Bingchun Jiang; Shulei Zheng; Peipei He
• Keywords: Sample (material)

Building change detection (CD) using remote sensing images plays a vital role in urban development, and deep learning models attracted attention for their potential to accomplish CD tasks automatically. However, most methods are still facing challenges, such as the costly time-consuming process of constructing datasets severely imbalanced distribution positive negative samples preventing loss functions from functioning desired training process. Inspired by weak supervision have demonstrated excellent performance solving above-mentioned problems, unified pre-training paradigm is proposed task small number improve inference accuracy building detection. The keys this method follows. First, detects changes pseudo-labels them generated highly available semantic segmentation datasets. Second, balanced sample ensured masked areas controlling proportion areas. Third, multi-task networks simultaneous extraction used paradigm, owing information can be employed an effective signal assist with solve problem that adversely affect ability algorithm converge. In particular, experiments were performed on three challenging For aerial WHU-CD satellite Gaofen Challenge-CD datasets, our pre-trained weights pseudo-bitemporal applied subsets containing different proportions ground truth fine-tuning, respectively. Notably, 10% fine-tuning obtained intersection over union (IoU) comparable 100% without weights, whereas even greater IoU was achieved 30% weights. Experiments based superior conventional supervised methods. When conducting truth, results showed use yields substantially exceeds conducted LEVIR-CD dataset proved transferable method. greatly reduce need high-quality labels alleviate bottlenecks caused distribution. Moreover, has generalization capabilities.