An Integrated Approach for Wildfire Photography Telemetry using WRF Numerical Forecast Products

Abstract

Forest fire detection using machine vision has recently emerged as a hot research topic. However, the complexity of background information in smoke images often results in deep learning models losing crucial details while capturing smoke image features. To address this, we present a detection algorithm called Multichannel Smoke YOLOv5s (MCSYOLOv5s). This algorithm comprises a smoke flame detection module, multichannel YOLOv5s (MC‐YOLOv5s), and a smoke cloud classification module, Smoke Classification Network (SCN). MC‐YOLOv5s uses a generative confrontation structure to design a dual‐channel feature extraction network and adopts a new feature cross-fusion mechanism to enhance the smoke feature extraction ability of classic YOLOv5s. The SCN module combines Weather Research and Forecasting numerical forecast results to classify smoke and clouds to reduce false positives caused by clouds. Experimental results demonstrate that our proposed forest fire monitoring method, MCS‐YOLOv5s, achieves higher detection accuracy of 95.17%, surpassing all comparative algorithms. Moreover, it effectively reduces false alarms caused by clouds.