FireFormer: an efficient Transformer to identify forest fire from surveillance cameras

Abstract

Background An effective identification model is crucial to realise the real-time monitoring and early warning of forest fires from surveillance cameras. However, existing models are prone to generate numerous false alarms under the interference of artificial smoke such as industrial smoke and villager cooking smoke, therefore a superior identification model is urgently needed. Aims In this study, we tested the Transformer-based model FireFormer to predict the risk probability of forest fire from the surveillance images. Methods FireFormer uses a shifted window self-attention module to extract similarities of divided patches in the image. The similarity in characteristics indicated the probability of forest fires. The GradCAM algorithm was then applied to analyse the interest area of FireFormer model and visualise the contribution of different image patches by calculating gradient reversely. To verify our model, the monitoring data from the high-point camera in Nandan Mountain, Foshan City, was collected and further constructed as a forest fire alarm dataset. Key results Our results showed that FireFormer achieved a competitive performance (OA: 82.21%, Recall: 86.635% and F1-score: 74.68%). Conclusions FireFormer proves to be superior to traditional methods. Implications FireFormer provides an efficient way to reduce false alarms and avoid heavy manual re-checking work.

Sensor-Based Indoor Fire Forecasting Using Transformer Encoder

“…There are a few studies that have applied the self-attention mechanism of Transformer to fire prediction, but they have focused only on the image or video data. For example, FireFormer is a Transformer-based architecture dedicated to forest fire detection using surveillance cameras [17]. In Ref.…”