Wildfire and smoke early detection for drone applications: a light-weight deep learning approach

Abstract

Drones have become a crucial element in current wildfire and smoke detection applications. Several deep learning architectures have been developed to detect fire and smoke using either colour-based methodologies or semantic segmentation techniques with impressive results. However, the computational demands of these models reduce their usability on memory-restricted devices such as drones. To overcome this memory constraint whilst maintaining the high detection capabilities of deep learning models, this paper proposes two lightweight architectures for fire and smoke detection in forest environments. The approaches use the Deeplabv3+ architecture for image segmentation as baseline. The novelty lies in the incorporation of vision transformers and a lightweight convolutional neural network architecture that heavily reduces the model complexity, whilst maintaining state-of-the-art performance. Two datasets for fire and smoke segmentation, based on the Corsican, FLAME, SMOKE5K, and AI-For-Mankind datasets, are created to cover different real-world scenarios of wildfire to produce models with better detection capabilities. Experiments are conducted to show the benefits of the proposed approach and its relevance in current drone-based wildfire detection applications.