Forecasting Crop Residue Fires in Northeastern China Using Machine Learning

Abstract

With repeated changes to local crop residue disposal policies in recent years, the distribution and density of crop residue fire events have been irregular in both space and time. A nonlinear and complex relationship between natural and anthropogenic factors often affects the occurrence of crop residue field fires. To overcome this difficulty, we used the Himawari-8 wildfire data for 2018–2021 to verify the likelihood of crop residue fires against the results of three machine learning methods: logistic regression, backpropagation neural network (BPNN), and decision tree (DT). The results showed the verified accuracies of BPNN and DT methods were 68.59 and 79.59%. Meantime, the sensitivity and specificity of DT performed the best, with the value of area under the curve (AUC) 0.82. Furthermore, among all the influencing factors, open burning prohibition constraints, relative humidity and air pressure showed significant correlations with open burning events. As such, BPNN and DT could accurately forecast the occurrence of agricultural fires. The results presented here may improve the ability to forecast agricultural field fires and provide important advances in understanding fire formation in Northeastern China. They would also provide scientific and technical support for crop fire control and air quality forecasting.