Detecting Forest Fires in Southwest China From Remote Sensing Nighttime Lights Using the Random Forest Classification Model

Abstract

Forest fires are one of the most common natural disasters and public crisis events, presenting a serious threat to the ecosystem and human security. The traditional forest fire monitoring is time-consuming and inaccurate, but nighttime light remote sensing is enhancing the efficiency and speed of forest fire detection. Thus, this study focused on forest fire detection in Southwest China that was affected by fires in 2021 and 2022 using daily remote sensing nighttime light data from the black marble product VNP46A2 of the Suomi National Polar-Orbiting Partnership-Visible Infrared Imaging Radiometer Suite. Multiple features were developed based on the temporal features of day-level nighttime light radiances, which were divided into black pixels, stable light pixels, and forest fire pixels using the random forest classification model. To verify the accuracy, a confusion matrix was constructed to calculate the accuracy of the forest fire identification results by combining multisource data, and the results were compared with the results of different machine learning models and fire products. Moreover, the spatiotemporal distribution of forest fires was analyzed using trend analysis and nearest-neighbor analysis. Results show that the overall accuracy of pixel classification was over 92% in both years, with forest fire pixels classified with user's accuracy exceeding 99%. The forest fires were mainly dispersed and concentrated around the stable light pixels 400–700 m, and the distribution in different regions showed obvious spatial heterogeneity. Compared with 2021, the distribution of forest fires decreased significantly in 2022. This study can support the future management and protection of forest resources and the prediction of fire disasters. The dataset (2021 and 2022) of forest fires in Southwest China is available free of charge.1

Keywords

• Forest fires
• nighttime lights (NTLs)
• random forest classification
• Southwest China
• spatiotemporal analysis