A Self-supervised Wildfire Risk Prediction Model Incorporating Spatio-temporal Features

Wildfires affect the global carbon cycle and pose great hazards to the ecological environment and human activities.The existing studies summarize the regularity of wildfire occurrence mainly focus on factors such as excavation climate conditions,combustible material status,and human activities.Temperature,humidity,wind speed,and humus thickness do have important effects on the occurrence of wildfires.Besides,some temporal factors,such as season and day and night,lead to changes in meteorological conditions and burning status;some spatial factors,such as the sunny and shady sides of slopes and steepness of hills,change combustible characteristics by affecting the distribution of vegetation.These spatial and temporal constraints make it challenging to predict fire activity.Several scholars have begun to study the relationship between forest fires and time series and spatial dispersion on.In this paper,we further explore the influence of spatial and temporal distributions on forest fire risk based on traditional influences and previous studies using a monthly wildfire sample collection(January 2000 to December 2003)from the historical fire dataset of Montesinho Natural Park.The specific study and results are as follows:(1)The correlation patterns between wildfires and time series,spatial dispersion,and forest fire weather index were explored.Using correlation coefficients and statistical indicators,it was concluded that the occurrence of wildfire showed a certain aggregation in space,a certain regularity in time,especially in monthly time series,and a close relationship with forest fire weather index,and most of them were positively correlated.Among them,the highest correlation coefficients were found between the water content of surface combustible material in the upper layer of forest humus and the total number of wildfires and the total burned area,indicating that this index has the most significant influence on the occurrence of forest fires and the burned area;there were strong positive correlations between temperature and precipitation and the number of wildfires and the burned area,which indicated that the increase in temperature and decrease in precipitation was one of the important factors leading to the occurrence of wildfires.There is a negative correlation between relative humidity and wind speed and the number of wildfires and the area burned,which means that the increase of relative humidity and wind speed can effectively reduce the occurrence of wildfires and the area burned.(2)Automatic sample labeling using self-supervised learning.Using the structure and attributes of the dataset itself,the wildfire samples were unfolded in a monthly sequence,and the patterns of wildfire distribution were analyzed using the K-means++clustering algorithm for each subset,and the monthly sequence samples were automatically labeled as flammable or non-flammable based on Euclidean distances and blocks(X,Y coordinates).(3)A wildfire prediction model based on spatio-temporal characteristics was trained.Four machine learning models were trained using a hierarchical five-fold cross-validation approach: Local Cascade Ensemble(LCE),Support Vector Machine(SVM),Random Forests(RF),Extreme Gradient Enhanced Tree(XGBoost).The XGBoost model performed best under the evaluation of five metrics: Accuracy(ACC),Precision(P),Recall(R),F1＿Score(F1),and Area under the Receiver Operating Characteristic Curve(AUC)(ACC=0.888,P=0.889,R=0.882.F1=0.885,AUC=0.888).Compared with the method of classifying wildfires based on overfire area size(ACC=0.7230),the performance of the model improved significantly,demonstrating that spatio-temporal heterogeneity has a broad impact on wildfire occurrence.Exploring the spatio-temporal distribution relationship of wildfire occurrence can help provide management strategies for forest and grassland fire prevention and disaster prevention,and reduce the risk of large fires.