Computer vision-driven forest wildfire and smoke recognition via IoT drone cameras

Abstract

Forest wildfires often lead to significant casualties and economic losses, making early detection crucial for prevention and control. Internet of Things connected cameras mounted on drone provide wide monitoring coverage and flexibility, while computer vision technology enhances the accuracy and response time of forest wildfire monitoring. However, the small-scale nature of early wildfire targets and the complexity of the forest environment pose significant challenges to accurately and promptly identify fires. To address challenges such as high false-positive rates and inefficiency in existing methods, we propose a Forest Wildfire and Smoke Recognition Network termed FWSRNet. Firstly, we adopt Vision Transformer, which has shown superior performance in recent traditional classification tasks, as the backbone network. Secondly, to enhance the extraction of subtle differential features, we introduce a self-attention mechanism to guide the network in selecting discriminative image patches and calculating their relationships. Next, we employ a contrastive feature learning strategy to eliminate redundant information, making the model more discriminative. Finally, we construct a target loss function for model prediction. Under various proportions of training and testing dataset allocations, the model exhibits recognition accuracies of 94.82, 95.05, 94.90, and 94.80% for forest fires. The average accuracy of 94.89% surpasses five comparative models, demonstrating the potential of this method in IoT-enhanced aerial forest fire recognition.