Real-Time Integration of Segmentation Techniques for Reduction of False Positive Rates in Fire Plume Detection Systems during Forest Fires

Abstract

Governmental offices are still highly concerned with controlling the escalation of forest fires due to their social, environmental and economic consequences. This paper presents new developments to a previously implemented system for the classification of smoke columns with object detection and a deep learning-based approach. The study focuses on identifying and correcting several False Positive cases while only obtaining a small reduction of the True Positives. Our approach was based on using an instance segmentation algorithm to obtain the shape, color and spectral features of the object. An ensemble of Machine Learning (ML) algorithms was then used to further identify smoke objects, obtaining a removal of around 95% of the False Positives, with a reduction to 88.7% (from 93.0%) of the detection rate on 29 newly acquired daily sequences. This model was also compared with 32 smoke sequences of the public HPWREN dataset and a dataset of 75 sequences attaining 9.6 and 6.5 min, respectively, for the average time elapsed from the fire ignition and the first smoke detection.