基于地理加权随机森林的黑龙江省森林碳储量遥感估测

Geographically weighted random forest approach to predict forest carbon storage by remote sensing in Heilongjiang

【目的】构建地理加权随机森林(Geographically weighted random forest,GWRF)模型估算森林碳储量以解决区域尺度范围内森林碳储量估算精度不高的问题,对科学经营管理森林、推动碳循环和碳汇相关研究、实现我国“双碳”目标有重要指导意义。【方法】以黑龙江省小兴安岭、长白山地区森林植被碳储量为研究对象,基于2015年森林资源连续清查数据和Landsat8-OLI影像,采用普通最小二乘(Ordinary least squares,OLS)、随机森林(Random forest,RF)模型、地理加权回归(Geographically weighted regression,GWR)模型以及地理加权随机森林模型分别构建不同林型及总体(不分林型)的森林碳储量估测模型,比较是否区分林分类型时,不同模型预测精度之间的差异,实现对研究区森林碳储量的精准反演。【结果】1)各个模型在区分林型时的预测精度均高于总体(不分林型)情况,以GWRF模型精度最优,其中针叶林精度最高(R^(2)=0.58,RMSE=15.97 t/hm^(2));阔叶林次之(R^(2)=0.46,RMSE=17.66 t/hm^(2));针阔混交林随后(R^(2)=0.45,RMSE=19.51 t/hm^(2));总体(不分林型)最低(R^(2)=0.40,RMSE=20.22 t/hm^(2))。2)4种模型的检验精度GWRF>RF>GWR>OLS。与OLS相比,GWRF在针叶林、阔叶林、针阔混交林和总体(不分林型)中提升的ΔR^(2)分别为0.15、0.09、0.16和0.04;降低的ΔRMSE分别为2.09、1.35、3.47和0.89 t/hm^(2);与RF相比,GWRF提升的ΔR^(2)分别为针叶林0.14、阔叶林0.06、针阔混交林0.04、总体(不分林型)0.02;降低的ΔRMSE分别为针叶林1.95 t/hm^(2)、阔叶林0.86 t/hm^(2)、针阔混交林0.67 t/hm^(2)、总体(不分林型)0.29 t/hm^(2)。3)研究区森林碳储量密度最高预测值为77.08 t/hm^(2),最低值为5.24 t/hm^(2),平均值为41.07 t/hm^(2),总量为552.04 Tg;从空间上看,森林碳储量高值分布在小兴安岭东南部、张广财岭等地区,呈现斑状不均匀性分布。【结论】相比于其他3种模型,GWRF作为局部模型,考虑到空间异质性,在区域尺度范围内估测森林碳储量有较好的应用前景。区分林分类型能提高预测精度,在今后对森林生物量或碳储量的研究中,应考虑区分林分类型建模。本研究的模型和方法有一定适应性,可为森林资源的快速和精准监测提供方法借鉴。

【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 (R^(2)=0.58,RMSE=15.97 t/hm^(2));followed by broadleaf forest (R^(2)=0.46,RMSE=17.66 t/hm^(2));mixed forest (R^(2)=0.45,RMSE=19.51 t/hm^(2));and the lowest accuracy for the total (no forest type) (R^(2)=0.40,RMSE=20.22 t/hm^(2)).2) The test accuracy of the four models was GWRF>RF>GWR>OLS.Compared with OLS,GWRF increased ΔR^(2) 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/hm^(2),respectively.Compared with RF,the ΔR^(2) 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 is1.95 t/hm^(2) in coniferous forest,0.86 t/hm^(2) in broadleaf forest,0.67 t/hm^(2) in mixed forest,and 0.29 t/hm^(2) in total (no forest type).3) The highest predicted forest carbon storage density in the study area was 77.08 t/hm^(2),the lowest is 5.24 t/hm^(2),the average was 41.07 t/hm^(2),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.