A lightweight Dual-Stream Attention Network for real-time landslide monitoring in multi-modal remote sensing imagery

Rapid delineation of landslides from post-event remote-sensing imagery demands models that can be deployed in the field, often on battery-powered edge devices with strict memory and latency budgets. Most state-of-the-art detectors break those constraints, shipping tens of millions of parameters to gain marginal accuracy. We therefore introduce a Dual-Stream Attention Network whose total trainable footprint is just 2.03 M parameters – roughly the size of a single JPEG image – yet still exploits the complementary physics of co-registered RGB orthomosaics and digital-elevation models. Two light encoders process the modalities independently; a provably isometric late-fusion operator merges their hierarchies without information loss, and a $(\le 1)$-Lipschitz spatial-channel gate discards redundant features while preserving gradient stability. Despite its size, the network attains an IoU of 0.835 and an mIoU of 0.818 on the high-resolution Bijie benchmark, coming within 4–5 percentage points of heavyweight baselines such as DeepLabv3^{\protect \relax \special {t4ht=＋}}(R-101) while using $\approx 95\text{\%}$ fewer weights. Ablation studies show that each architectural choice – dual-stream processing, late fusion, and attentional gating – contributes at least $+1.5$ pp IoU. A single Jetson Xavier AGX segments a 256 × 256 tile in 1.6 s ($<10$ W envelope), confirming real-time suitability for rapid landslide mapping missions. By reconciling DEM-derived information with extreme parameter efficiency, the proposed architecture offers a practical foundation for next-generation, on-device geohazard monitoring systems.