Remote sensing inversion of forest biomass based on hyperparametric optimized random forests algorithm
[Objective]Accurately estimating forest aboveground biomass(AGB)is crucial for large-scale forest resource investigation and management.Machine learning algorithms can achieve high-precision estimation of forest AGB,but the setting of hyperparameters can directly affect model performance.To improve the efficiency and prediction accuracy of the model,machine learning algorithms optimized by hyperparameters were constructed for forest AGB estimation,and the model error changes under different hyperparameters were compared.[Method]The study takes natural forests in Jiangda county,Tibet Autonomous Region as the research object,and the measured forest AGB data was extracted by forest resource survey data.Sentinel-2 multispectral images were used to extract remote sensing variables.The remote sensing variables were screened using stepwise regression method and Boruta method respectively,and multiple linear regression model(MLR),support vector machine(SVM)model and random forests(RF)model were constructed for forest AGB inversion.In addition,hyperparameter optimization was performed for the SVM model and RF model to improve the model accuracy.[Result](1)The RF model achieved the best estimation accuracy among all inversion models.The RF achieved a coefficient of determination of 0.63,while achieving the lowest root mean square error(RMSE)and relative root mean square error(rRMSE)of 28.06 t/hm2 and 23.03%,respectively.The RMSE was reduced by 22.2%and 12.1%compared to the MLR model and SVM model,respectively.(2)Hyperparameter optimization can effectively improve the model estimation accuracy.By analyzing the error variation trend under different parameter combinations and determining the best parameter combination,the model estimation error was effectively reduced.(3)The higher forest AGB values were mainly distributed in the eastern,southern and southeastern regions,and the forest AGB values were smaller in the central region and some northern regions.The forest AGB inversion results of the hyperparametric optimized random forests model were in good agreement with the actual forest distribution in the study area,and the overall inversion effect is satisfactory.[Conclusion]The RF model with hyperparameter optimization combined with Sentinel-2 remote sensing images can achieve a better inversion of forest AGB,which can provide an effective reference for forest resource dynamics monitoring.
NETL
NSTL
万方数据
