An Improved Stacking Model for Forest Fire Susceptibility Prediction in Chongqing City, China

Forest fires are a type of serious natural disaster causing great damage to human life and property. Developing a scientific and accurate prediction model is an effective means to reduce forest fire risk. However, previous studies have primarily relied on single machine-learning models with the disadvantage of poor prediction accuracy. To address this issue, we proposed an improved stacking model that utilized accuracy weighting, random sampling methods, and stacking ensemble technology. Taking Chongqing City, China as the study region, this study used remotely sensed anomalous hotspot data as the forest fire sample. Sixteen factors related to meteorology, topography, vegetation, and human activities were used as environmental variables affecting forest fire occurrence. The prediction performance of the proposed model was assessed using the receiver operating characteristic (ROC) curve and several statistical measures. The results confirmed that the area under the ROC curve (AUC) of the improved stacking model was 0.9447, which was 0.0721-0.0242 higher than the traditional machine learning methods. In general, this proposed improved stacking model can effectively generate high-quality forest fire risk susceptibility maps and can be applied to other disaster prediction fields.