Lightweight multiscale feature fusion and edge inference for remote sensing image change detection

Abstract

Remote sensing change detection (CD) supports urban expansion, disaster assessment, and land-use monitoring. Transformer models are accurate, but quadratic attention and memory demand hinder near-sensor deployment. LCSF-Net is a lightweight Siamese CD network with a multi-level lightweight Cross-Scale Fusion (LCSF) design. LCSF replaces quadratic self-attention with shared-weight bidirectional difference-guided spatial attention of linear complexity O(N), improving change-region recall. The edge-device prototype uses BCE-Dice training and TensorRT INT8 sliding-window inference on NVIDIA Jetson Orin Nano. On LEVIR-CD, LCSF-Net attains 90.5% F1 with 1.82 M parameters and 4.61 GFLOPs, surpassing BIT by 0.6 F1 percentage points and 2.6 recall percentage points under executable deployment configurations, while running at 28.9 FPS compared with BIT’s 7.2 FPS. Controlled degradation, zero-shot transfer, and large-scene tests support this trade-off.