LTS-ProtoNet: Local–Temporal–Semantic Prototype Network for Remote Sensing Change Detection

Abstract

Remote sensing change detection still faces several challenges in complex scenes, including boundary ambiguity, cross-temporal appearance interference, and large object-scale variations. To address these issues, this article proposes LTS-ProtoNet, a local–temporal–semantic prototype network driven by a learnable change prototype basis (CPB). The proposed method learns sample-shared latent prototype bases through the CPB mechanism and introduces dimension-matched prototype representations into different stages of the hierarchical reasoning process for change detection (CD). Built upon this design, three complementary modules are developed. The local prototype refinement module enhances shallow boundary and local structural details. The temporal prototype interaction (TPI) module performs prototype-mediated bitemporal semantic interaction to alleviate pseudochange interference. The semantic prototype enhancement module adaptively aggregates multiscale semantic information. Extensive experiments on three public datasets, including LEVIR-CD, WHU-CD, and GZ-CD, demonstrate that LTS-ProtoNet effectively improves the completeness of changed regions, boundary localization accuracy, and multiscale target detection stability in complex scenes, achieving competitive detection performance.