CNN-based, contextualized, real-time fire detection in computational resource-constrained environments

Abstract

The increasing occurrence of wildfires, amplified by the changing climate conditions and drought, poses threats to human lives, the environment and the geographically dispersed infrastructures. Such impact necessitates the prompt identification of wildfires so that appropriate countermeasures are taken. The availability of electronic equipment, such as Unmanned Aerial Vehicles, allows for images from dynamically changing, geographical areas, which must be directly processed for wildfire identification and contextualization. In this work, we identify the requirements and the constraints in terms of computational resources of this workflow, and investigate lightweight CNNs to be used. SqueezeNet, ShuffleNet, MobileNetv2 as well as ResNet50 are used for fire identification. To simulate the realistic conditions, we have investigated multiple datasets, selecting Forest-Fire and Fire-Flame datasets and images from 3rd party sources and performed cross-dataset identification evaluation. To rationalize the required computational resources and the operation cost, lightweight networks have been selected and compared with ResNet-50, which is more complex. The contextualization, i.e. the detection of elements related to energy infrastructures, has been based on image semantic segmentation, performed through ResNet-18. The identification results, expressed as classification accuracy has reached 96%, with satisfactory results in the cross dataset scenarios, while we have identified five classes from the CamVid dataset which can be used for the contextualization needs.