Forest-Fire Response System using Deep-Learning-based Approaches with CCTV Images and Weather Data

Abstract

Forest fires are an increasing environmental and financial risk and require intelligent and rapid fire detection mechanisms. In this paper, an AI-based IoT system was proposed, which combines thermal, visual, and meteorological information to identify a forest fire at its early-stage development and activate a response based on UAV. The presented system involves a hybrid deep ensemble framework, namely, DeViW-FNet that consists of Swin Transformers, weather models based on BiLSTMs, and multimodal co-attention fusion to determine anomalous fire patterns. The new uncommon methods, such as the Cross-Domain Calibration, the Federated Dynamic Time Warping Autoencoders, and the Quantum-Inspired Edge Ensemble Voting, have a significant impact on the system and enhance its strength in the extreme and dubious environment. Experiments on a wide range of environmental conditions such as fog, smoke, low light, etc. reveal that the detection performance is high with 94.5 mean visual detection accuracy, 95.1 weather-based classification accuracy, and 90 plus anomaly detection F1-score. Swarm reinforcement learning is applied to ensure the response latency of UAV is minimized to a level that the accuracy of the navigation was 93.5%. The study also presents the promise of cross-modal AI fusion in real-time fire detection in complicated environments. The suggested framework is scalable, low-latency, and accommodating to the changes in the environment, which would be applicable to forests in large scale. Some improvements that can be made in the future are thermal drone vision, explainable AI modules, and compatibility with satellite-based wildfire propagation simulators.