Lightweight Cross-Temporal Gating Network for Remote Sensing Change Detection

Abstract

Accurate remote sensing change detection requires separating genuine land-cover changes from appearance variations while retaining small objects and boundaries. This paper presents a Lightweight Cross-Temporal Gating Network (LCTGNet), a compact Siamese convolutional model for bi-temporal images. A single shared MobileNetV2 extracts four feature levels, and scale-specific projections map them to a common 64-channel space. Coordinate attention strengthens direction-aware spatial responses. The proposed Cross-Temporal Gating (CTG) module combines an explicit absolute difference with a learned gate and a complementary bi-temporal context branch, thereby filtering unreliable difference responses without discarding contextual evidence. A Lightweight Multi-Scale Context Enhancement (LMCE) module further refines only the two deepest change features through parallel dilated depthwise convolutions. A top-down decoder produces four deeply supervised predictions. The complete model contains 2.636 million registered parameters and requires approximately 2.34 GFLOPs for a 256 × 256 image pair under a multiply-and-add-as-two-operations convention. Across seeds 0, 42, and 3407, it obtains F1-scores of 91.44±0.07% on LEVIR-CD, $93.96 \pm 0.18 {\%}$ on WHU-CD, and $84.51 \pm 0.13 {\%}$ on SYSU-CD. The results show competitive accuracy with a compact executed computation graph, while the WHU experiment reveals a stable recall-precision trade-off.