| Forest fire is one of the most harmful natural disasters and an important factor that causes global forest loss.In recent years,with the increase of extreme weather in the world,forest fires have also shown an increasing trend,and the occurrence of major forest fire events has increased significantly.Therefore,how to effectively reduce the losses caused by forest fires is one of the urgent problems to be solved at present,and the construction of high-precision risk prediction model is a research hotspot at this stage.The occurrence conditions of forest fire behavior are complex,with distinct regional differences,and the main factors affecting forest fire occurrence will change in different regions,so there is a lack of unified prediction model at present.For typical areas,it is difficult to build a high-precision dynamic forest fire risk prediction model with regional characteristics.With the development of machine learning algorithms,their advantages in data mining have gradually become prominent.Especially when forest fire risk involves multi-factor coupling,machine learning methods have greater advantages and applicability than traditional methods.In this study,Liangshan Prefecture,Sichuan Province,China,was selected as an example.Based on multi-source remote sensing data,meteorological factors,social and economic data and human activity data,various machine learning training models were used to build forest fire risk prediction models,and various models were compared and analyzed to select the optimal regional model.The main research contents and achievements are as follows:(1)Based on the real fire point from 2005 to 2018,the historical fire situation was studied by mathematical statistics.The results show that from 2005 to 2018,the number of fire points varies greatly from year to year,showing great fluctuation.From 2005 to 2009,the number of fire points increased regularly.In 2010,the number of fire points reached the first peak of 186,accounting for 21.26%of the total.2014 is the second peak,the number of fire points reached a maximum of 240,accounting for 27.43%of the total fire points.The lowest number of fires occurred in 2005 and 2016.From the inter-monthly distribution law,the fire point is mainly distributed between November and April of the following year,which corresponds to the bright dry and wet seasons in Liangshan Prefecture.The number of fire points was the highest in April,then showed a steady downward trend,fell to the minimum value in December,and then increased,and the growth rate accelerated after January,reaching a small peak in February.From the spatial distribution of fire points,fire points are mainly concentrated in the central and western and southern parts,and the distribution of fire points decreases from west to east.The number of fire points distributed in Muli County is the largest,followed by Xichang City,Jinyang County has no fire point distribution,and other areas with fewer fire points are mainly distributed in the east.The spatial distribution of fire points is closely related to vegetation type,elevation and human activities.The analysis of 119 forest fires that occurred from 2017 to 2022 found that the fire time was relatively concentrated,with general fires and larger fires being the main ones,and the number of general fires was the highest.The cause analysis of 98 fires found that the proportion of man-made fires was 88.78%,and non-production fires were the main ones.The analysis of the area of the fire site found that the area of the fire site varied greatly from year to year.The maximum area of the fire site was 3,284.6 hectares in March 2020,and the minimum area of the fire site was 0.85 hectares in 2022.(2)A statistical study was carried out on fire points and 14 forest fire factors,and the results showed that 82.97%of fire points were distributed in high vegetation coverage areas;The number of fire points in evergreen coniferous forest area is more.In the range of fuel moisture content[0.4,0.7],85.14%of the ignition points are distributed;With the increase of TVDI and wind speed,the fire point showed an increasing trend.There are obvious distribution rules between fire point and terrain factor.About 50.51%of the fires were located in the areas with the least rainfall.83.77%of the ignition points were between 17-29℃.Nearly 90%of the fires are located in areas where the relative minimum humidity is less than 30%.Variance inflation factor and Pearson correlation coefficient were used to test the multicollinearity of the selected factors.The results showed that the 14 factors had low correlation and could be used for subsequent model training.(3)A forest fire risk prediction model based on logistic regression,decision tree,random forest,K_nearest neighbor and XGBoost algorithms was constructed,and the accuracy of the model was evaluated by confusion matrix and ROC curve.The results showed that the XGBoost fire risk prediction model had high sensitivity to fire point and the highest accuracy was 74%.The importance of forest fire driving factors was analyzed.The results showed that in the logistic regression model,the relative minimum humidity,elevation and settlement density were significantly negatively correlated with forest fire risk probability,and elevation was the most negatively correlated factor,while the other 11 factors were positively correlated with forest fire risk probability.The forest fire risk probability of decision tree model is greatly affected by wind speed and human factors.The characteristic importance of each factor in random forest and XGBoost prediction model is relatively balanced.In contrast,elevation,minimum relative humidity and settlement density are the most important factors.(4)The probability of forest fire risk in 2018 was predicted based on XGBoost model,and its feasibility was analyzed.The results showed that the real fire points were distributed in the high or extremely high risk areas predicted by the model;The forest fire risk areas of Liangshan Prefecture are mainly distributed in Xichang City,Yanyuan County,Ningnan County,Huili County and Muli County,while the forest fire risk level of eastern Liangshan Prefecture is low.Compared with the forecast results of the national forest fire weather grade,four fire points are distributed in the high risk area predicted by the model,but only one fire point is distributed in the risk area of the fire weather forecast.The paper summarizes the feasible forest fire prevention countermeasures suitable for Liangshan Prefecture,and provides theoretical support for scientific fire prevention work in the study area. |
PDF Full Text Request