Wildfire detection with deep learning—A case study for the CICLOPE project

• Autor(a) principal: Gonçalves, A. M.
• Data de Publicação: 2024
• Outros Autores: Brandão, T., Ferreira, J. C.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10071/31906
• topic: Computer vision; Convolutional neural networks; Deep learning; Smoke detection; Wildfire detection

In recent years, Portugal has seen wide variability in wildfire damage associated to high unpredictability of climatic events such as severe heatwaves and drier summers. Therefore, timely and accurate detection of forest and rural wildfires is of great importance for successful fire containment and suppression efforts, as wildfires exponentially increase their spread rate from the moment of ignition. In the field of early smoke detection, the CICLOPE project currently trailblazes in the employment of a network of Remote Acquisition Towers for wildfire prevention and observation, along with a rule-based automatic smoke detection system, covering over 2, 700, 000 hectares of wildland and rural area in continental Portugal. However, the inherent challenges of automatic smoke detection raise issues of high false alarm rates that affect the system’s prediction quality and overwhelm the Management and Control Centers with numerous false alarms. The research work presented in this paper evaluates the potential improvement in wildfire smoke detection accuracy and specificity using deep learning-based architectures. It proposes a solution based on a Dual-Channel CNN that can be deployed as a secondary prediction confirmation layer to further refine the CICLOPE automatic smoke detection system. The proposed solution takes advantage of the high true alarm coverage of the current detection system by taking only the predicted alarm images and respective bounding box coordinates as inputs. The Dual-Channel network combines the widely used DenseNet architecture with a novel detail selective network with spatial and channel attention modules trained separately with image data obtained from CICLOPE, fusing the extracted features from both networks in a concatenation layer. The results demonstrate that the proposed Dual-Channel CNN outperforms both single-channel networks, achieving an accuracy of 99.7% and a low false alarm rate of 0.20% when re-examining the alarms produced by the CICLOPE surveillance system.