FloodSense-VLM: A Lightweight Vision-Language Model for Real-Time Disaster Management

This paper presents a novel fully automated system for real-time flood monitoring and forecasting, utilizing multimodal visual-language model (VLM) approaches. Our system, which uses environmental sensing and state of the art deep learning, takes on UFO questions and challenges in disaster management. The empirical analyses we present, related to the spatiotemporal patterns of flooding, are achieved through collecting, processing, and founding, on water level and precipitation amounts, and the knowledge of the terrain. Our system utilizes a lightweight transformer-based VLM, specifically HuggingFaceTB/SmolVLM-256M-Instruct, which we further fine-tuned via Parameter-Efficient Fine-Tuning (PEFT) with Low-Rank Adaptation (LoRA) to adjust and adapt the model architecture and weights for flood analysis without some knowledge of the flood context. This strategy only requires training a small percentage (2.20%) of the model parameters. The system is able to produce reliable performance results with a final validation loss of 0. 1 4 4 4, and test loss of 0. 1 3 8 9 on our custom Flood-VQA dataset. In addition, on binary flood detection tasks, the model achieved an impressive 80.0% precision, and 0.615 F1-Score. Additionally, the model demonstrates strong cross-domain generalisation capacities, having achieved 73.3% on the MMStar benchmark dataset. The lightweight nature of the model enables deployment on edge devices with constrained resources and within the urban, and remote areas of flooding.