FloodNet-Lite: A Lightweight Deep Learning for Flood Mapping Using Remote Sensing Data with Optimized UNet and Edge Deployment Approach in 6G

Flood mapping using remote sensing data is critical to disaster response, especially in real-time monitoring and edge deployment. However, existing deep learning models often face challenges related to computational complexity, latency, and limited scalability in dynamic and resource-constrained environments. This paper proposes FloodNet-Lite, a lightweight and scalable deep learning framework for real-time flood mapping in 6G-enabled environments. The system integrates an optimized UNet architecture with MobileNetV3 as a backbone, enhanced by depthwise separable convolutions, attention mechanisms, and advanced model compression techniques such as quantization-aware training and structured pruning. Knowledge distillation transfers learning from a high-capacity teacher model to a compact student model, while self-supervised learning further reduces dependency on labeled data to boost efficiency and robustness. Crucially, this framework is designed for deployment in 6G edge computing infrastructures, leveraging 6G capabilities such as ultra-reliable low-latency communication (URLLC), AI-native edge nodes, and space-air-ground integrated networks (SAGIN). These features enable real-time flood detection via UAVs, ground sensors, and satellite imagery with minimal latency and maximal interoperability. The model achieves a mean IoU of 0.85 while maintaining an ultra-low model size of 18 MB and an inference speed of 14.3 FPS on Jetson Nano. This work provides a robust, scalable, and edge-ready flood mapping solution aligned with the vision and capabilities of future 6G-enabled intelligent disaster management systems.