EVALUATION OF LIGHTWEIGHT MODELS FOR WILDFIRE LOCALIZATION

Deep Neural Networks (DNNs) have become the de facto standard for computer vision applications, including its use in smoke and fire detection. Nevertheless, it is challenging to apply these network models on drones due to the high computational complexity, memory intensive, and the massive number of parameters. Therefore, we employ various model compression techniques to help with resource and latency constraints of devices like drones and to enable early fire and smoke detection. The potential to detect fire and smoke disasters using lightweight models mounted on drones enables early detection, rapid firefighter response, and prompt alerting of associated hazards to nearby communities. In order to reduce the computational burden, the methods use lightweight fully connected layers to accelerate reasoning, pruning was done to remove redundant parameters and to reduce multiply-accumulate operations, and then quantization to reduce the size of the model even further. Network Pruning and Quantization are effective in reducing the network complexity and addressing the problem of overfitting. Pruning can bring regularization to neural networks and Quantization compresses the network by reducing the number of bits required to represent each weight. The results show that the model can be compressed to 12MB with a fire MIoU and smoke MIoU of 0.6705 and 0.7197 respectively.