Estimation Of Forest Stock Volume Based On Machine Learning

In recent years,with the increasing number of satellite launches and the gradual maturity of computer technology,the estimation of forest stock by using computer technology combined with remote sensing has gradually become one of the research hotspots.At present,a large number of studies have proved that remote sensing data and a small part of field data can be used to estimate forest stock.How to further improve the estimation accuracy of forest stock is of great significance to meet the needs of human production.In this study,paiyashan State-owned Forest Farm in Jingzhou County,Hunan Province was selected as the research area,and Forest Resources Management Inventory data were selected as the basis,supplemented by field survey data as the ground data.Landsat8 OLI and Sentinel-2 data were selected as remote sensing data sources,and the variable fusion of the two satellite data was used as a new data source.Four single stock estimation models(multiple linear regression model,support vector machine model,random forest model,and Boruta variable selection method)were constructed by using four variable screening methods(PCA-P,Pearson correlation coefficient,principal component analysis method and Boruta variable selection method)respectively.K nearest neighbor model).Based on the results of single model construction,extreme gradient elevation model(XGBOOST)was used as the meta-model to construct the Stacking integrated learning model and draw the forest stock distribution map with the Stacking model results.The main conclusions of this results are as follows:(1)Among the single models constructed by using Pearson correlation coefficient,Boruta,PCA and the newly developed PCA-P as screening variables,the best model is the support vector machine model constructed by using PCA-P screening variables for fusion data variable sets,and its determination coefficient(R2)is 0.60.The root mean square error(RMSE)was 46.3m3.and the relative root mean square error(rRMSE)was 22.0%.Compared with the optimal model constructed by the other two kinds of data,the relative root mean square error of the fusion data is reduced by 0.1-0.2%.(2)Pearson correlation coefficient method combined with principal component analysis can effectively reduce the number of variables and improve the speed of modeling.By comparing the precision index table of each single model,it can be found that the optimal model of each variable set is constructed by using PCA-P screening variables,and the average accuracy of the model constructed by using PCA-P as screening variables is higher than that of the other three variable screening methods.Therefore,this experiment considers that PCA-P,as a variable screening method,can not only improve the timeliness,but also improve the model inversion accuracy when applied to the estimation of forest stock volume.(3)After significant difference test,support vector machine and kNN model were selected as the base model of Stacking model,and extreme gradient elevation model(XGBOOST)was selected as the meta-model to construct the Stacking integration model of three data sources.By comparison,it is found that the Stacking model has higher accuracy than the estimation results using a single algorithm.Among all the models,the Stacking model built with integrated data has the best Stacking performance,with R2 of 0.62,root mean square error of 43.7 m3.Hm-2 and relative root mean square error of 20.8%.Three Stacking models have achieved good results.