Geographically weighted random forest approach to predict forest carbon storage by remote sensing in Heilongjiang
[Objective]To construct a geographically weighted random forest(GWRF)model for characterizing forest carbon storage in order to address the problem of low accuracy in estimating forest carbon stocks at the regional scale.This has significant implications for the scientific management of forests,the advancement of research on the carbon cycle and carbon sequestration,and the achievement of our country's"double carbon"goal.[Method]Focusing on the carbon storage of forest vegetation in the Xiaoxing'an mountains and Changbai mountains of Heilongjiang province,this paper was based on the 2015 continuous forest resource inventory data and Landsat 8-OLI imagery.Different forest carbon storage estimation models were constructed for various forest types and the total(no forest type),using ordinary least squares(OLS),random forest(RF),geographically weighted regression(GWR),and geographically weighted random forest(GWRF).Additionally,this paper also compared the differences in prediction accuracy among different models whether distinguishing forest stand types and achieved accurate inversion of forest carbon storage in the study area.[Result]1)Each model exhibited higher predictive accuracy when distinguishing between forest types compared to the total(no forest type)situation.The GWRF model achieved the highest accuracy,with the highest precision for coniferous forest(R2=0.58,RMSE=15.97 t/hm2);followed by broadleaf forest(R2=0.46,RMSE=17.66 t/hm2);mixed forest(R2=0.45,RMSE=19.51 t/hm2);and the lowest accuracy for the total(no forest type)(R2=0.40,RMSE=20.22 t/hm2).2)The test accuracy of the four models was GWRF＞RF＞GWR＞OLS.Compared with OLS,GWRF increased ΔR2 by 0.15,0.09,0.16,and 0.04 in coniferous forest,broadleaf forest,mixed forest,and total(no forest type);and decreased ΔRMSE by 2.09,1.35,3.47 and 0.89 t/hm2,respectively.Compared with RF,the ΔR2 increased by GWRF is 0.14 for coniferous forest,0.06 for broadleaf forest,0.04 for mixed forest,and 0.02 for total(no forest type);the reduced ΔRMSE is 1.95 t/hm2 in coniferous forest,0.86 t/hm2 in broadleaf forest,0.67 t/hm2 in mixed forest,and 0.29 t/hm2 in total(no forest type).3)The highest predicted forest carbon storage density in the study area was 77.08 t/hm2,the lowest is 5.24 t/hm2,the average was 41.07 t/hm2,and the total was 552.04 Tg.From a spatial perspective,high values were concentrated in the southeastern regions of the Xiaoxing'an mountains and Zhangguangcai mountains,displaying a patchy and uneven distribution.[Conclusion]Comparing to the other three models,GWRF,as a local model that accounts for spatial heterogeneity,has promising applications for estimating forest carbon storage on a large scale.Differentiating forest stand types can improve the accuracy of prediction,we should take into account distinguishing stand type modeling in future research on forest biomass or carbon stocks.The models and methods studied in this paper have a certain level of adaptability and can provide methodological references for the rapid and precise monitoring of forest resources.
NETL
NSTL
万方数据
