Estimation Of Forest Aboveground Biomass And Determination Of Its Saturation Values Based On Passive And Active Data

The Asian temperate mixed forest in northeastern China is one of the three major temperate mixed forests in the world(i.e.,northeastern North America,Europe,and East Asia),which is of great strategic importance to the carbon trading of China.Currently,forest ecosystems in northeast China are gradually recovering from the over-harvesting of the 20th century,and the quality of forest ecosystems is gradually improving.Natural secondary forests(NSFs)are important not only for China's timber supply but also as an important reservoir of biodiversity,potential carbon sequestration,ecotourism destination,and vast ecological barrier in northeastern China.Forest biomass is the accumulation of dry matter produced by the forest plant community during its life and is expressed as the mass of dry matter accumulated per unit area or per unit time,usually including the total weight of stem,branches,leaves,and roots.Aboveground biomass(AGB)is only the aboveground biomass of forest trees,which usually accounts for 70?80%of the forest biomass.Thus,the accurate estimation of AGB has a critical effect on the understanding of forest quality and recovery in the NSFs of northeastern China.In this study,based on Landsat 8 OLI imagery,Airborne Laser Scanning(ALS)data,and 195 fixed plot data,three feature experiments were designed to explore the effects of different synergistic methods of Landsat 8 imagery and ALS data on forest AGB estimation,effects of seven classic machine learning algorithms on AGB estimation using different feature combinations,including BP(back propagation neural network),extreme learning machine(ELM),regression tree(RegT),random forest(RF),and support vector regression(SVR),KNearest Neighbor(KNN),and convolutional neural network(CNN).Four classical machine learning algorithms(RF,SVR,KNN,CNN)with good performances were used as the base models and meta models to establish ensemble learning algorithms based on the stacked generalization(SG).The performances of the ensemble learning algorithms,including SG(RF),SG(SVR),SG(KNN),and SG(CNN),were investigated for AGB estimation using different feature combinations.Based on the well-performed feature combinations,the uncertainties of machine learning and ensemble learning algorithms were investigated using Monte Carlo simulation.The saturation values of AGB obtained by different feature combinations and algorithms were predicted using Quantile Generalized Additive Models(QGAM)and spherical model.The specific contributions and main conclusions of this study were as follows:(1)Based on three feature experiments,the performance of multiple linear regression,seven machine learning algorithms,and four ensemble learning algorithms were compared for AGB estimation.The results showed that:compared with applying Landsat 8 or ALS data alone,whether classical machine learning or ensemble learning algorithms could improve the accuracy of AGB estimation if the active and passive remotely sensed data were combined;among the novel extracted features,COLI2 features are more suitable than COLI1 features for estimating AGB of natural secondary forests of the Maoershan area;regardless of the feature combinations,the CNN model is significantly better than multiple linear regression and other classic machine learning algorithms.The CNN model had the highest accuracy(R2=0.99,RMSE=6.85,rRMSE=0.04,MAE=2.95,MAPE=1.02,PM=0.03)when the optimal ALS and Landsat 8 features and all COLI2 features were used.(2)Compared to the corresponding base models(RF,SVR.KNN,and CNN),the ensemble learning algorithms could substantially improve the accuracy of AGB estimation,especially for the Landsat 8 features and the combination of Landsat 8 and ALS features.Since the ensemble learning algorithms were adept in training weak learners into strong learners.the SG algorithms for relatively weak learners(e.g.,RF.SVR,KNN)had more room for improvement than CNN.SG(CNN)had the high accuracy based on the combination of ALS.Landsat 8 and all COLI2 features(R~2=0.99,RMSE=2.02,rRMSE=0.01,MAE=0.87.MAPE=0.73,PM=0.02).(3)Based on Monte Carlo simulation,the uncertainty of AGB estimation using six feature combinations and eight machine learning algorithms was investigated.Results showed that the uncertainties of RF,SVR,and KNN,except for CNN,were small and had little impact on AGB estimation.The uncertainties generated by the ensemble learning algorithms were larger than those generated by the corresponding base models.Among them,SG(SVR)had the smallest uncertainty,while SG(CNN)had the largest uncertainty,which was three to four times larger than that of the SG(SVR)algorithm.(4)In this study,the saturation values of AGB based on six feature combinations,and four machine learning methods(RF,SVR,KNN,and CNN)and four ensemble algorithms(SG(RF).SG(SVR).SG(KNN),SG(CNN))were predicted by QGAM and spherical models.Results showed that the application of the QGAM model for predicting the saturation values based on different feature combinations and algorithms could effectively avoid the information loss caused by using a single-variable model.Meanwhile,the saturation range could be determined by QGAM based on different data,which was more accurate and reasonable than the unique saturation value predicted by the spherical model.The light saturation values of AGB of the northern NSFs predicted by the QGAM(0.25?q?5)model ranged from 148.068 to 231.387 t/ha,suggesting that the light saturation values of AGB varied with the variation of AGB estimation models even if only optical imagery were applied.(5)The saturation values of AGB were closely related to both features and the estimation algorithms.Among the traditional machine learning algorithms.SVR yielded low AGB saturation values,all less than 160 t/ha;while RF yielded high AGB saturation values,roughly between 188 and 244 t/ha;and both KNN and CNN yielded wide ranges of AGB saturation values.No matter of algorithms,the saturation values obtained by the COLI2 generated by Landsat 8 and ALS were much higher than those obtained by other feature combinations.The maximum saturation value(253.361 t/ha)was yielded by applying COLI2 features and CNN.(6)The saturation values of AGB obtained by the ensemble learning algorithms were significantly higher than those obtained by the classic machine learning algorithms,with most of them above 200 t/ha,with a maximum of 284.788 t/ha and a minimum value of 167.273 t/ha.Except for COLI2 features,SG(SVR)and SG(KNN)estimated the largest and smallest saturation values,respectively.For most feature combinations,the QGAM models with quartiles from 0.5 to 0.75 accurately covered the AGB saturation values of NSFs estimated by the ensemble learning algorithms.This study investigated the effects of applying different synergistic methods of active and passive remotely sensed data,classic machine learning algorithms,and ensemble learning algorithms on the AGB estimation,and also explored how to determine the AGB saturation values of NSFs in northeastern China.This study would provide technical support for accurate large-scale AGB estimation based on an area-based approach,and lay a solid foundation for the research related to precise forest survey,and forest carbon sink,and so on.