STUDY ON EARLY WARNING MODEL FOR REGIONAL LANDSLIDES BASED ON RANDOM FOREST IN FUJIAN PROVINCE

Landslides occur in a large number of places in Fujian Province,China,so it is an important measure to carry out regional landslides warning,and the scientificity and effectiveness of the warning model is the key. Due to the complex of landslide mechanism,limited data accumulation,and insufficient big data method,etc. traditional regional landslides warning models have many problems,such as limited warning accuracy and lack of refinement. Based on the geological and meteorological big data in the past decade,using the random forest algorithm,the regional landslides warning model in Fujian Province was studied in this paper. Then,the case verification was carried out. (1)The building method of regional landslide training sample set was optimized,and the regional landslides training set of Fujian Province was constructed,which included 26 input characteristics and 1 output characteristic,covered nearly nine years in Fujian Province(2010~2018)total sample size of 15589(3562 positive samples and 12027 negative samples). (2)Based on the random forest algorithm,the training sample set was trained,optimized and stored. The model was trained with the five-fold cross-validation method,and model parameters were optimized with the Bayesian optimization algorithm. Then,the accuracy value,ROC curve and AUC value were used to verify the model accuracy and generalization ability. Finally,the optimized model had good accuracy and generalization ability(the accuracy value was 94.3%,and AUC was 0.954). (3)The actual landslides on June 22 and 28,2021 were selected to verify the new warning model. Then,verified result was that the hit ratio of new warning model were 100% in both June 22 and 28. Compared with the warning results of the original explicit statistical model,the hit ratio of the new model is 6 times(on June 22)or equivalent(on June 28)to that of the original model,and the landslide density in the warning area of the new model is 1.6 to 1.7 times to that of the original model. Therefore,the new model based on random forest has obvious advantages,which is the new model warning result has higher hit rate and smaller warning area. In other words,the new model can achieve a more accurate warning. In the follow-up,new landslides in the study area will be tracked to verify and modify the model.