MLCNet: Multitask Level-Specific Constraint Network for Building Change Detection

Abstract

Change detection is an essential fundamental task in remote sensing image analysis. Owing to powerful deep abstract feature extraction ability, many deep learning-based change detection methods have emerged recently. Although previous works have recognized the characteristics and advantages of multilevel deep features and attempted to integrate them, the utilization process lacks a clear emphasis on the advantages aspect of different-level features. One-size-fits-all fusion treats all levels equally, neglecting the spatial advantage and semantic advantage of low and high-level features, respectively. This leads to deficiencies in the integrity and edge accuracy of the final change predictions. To address these issues, we propose a multitask level-specific constraint network, named MLCNet, which addresses the issues by optimizing the advantages of features at different levels. MLCNet comprises a Siamese encoder, fusion modules tailored for features at different levels, and a decoding mechanism that effectively combines semantic and spatial information of features at adjacent levels. In addition, by reconstructing the original ground truths (GTs) into the semantic, binary, and edge GTs, multitask learning constraints are established during network training, compelling the network to enhance targeted emphasis on the characteristics of features at different levels. Experimental results on the three building change detection datasets validate the practicality of MLCNet.

Keywords

• Change detection (CD)
• deep supervision
• level-specific constraint
• multilevel features
• multitask learning