基于多源数据并顾及森林类型的森林冠层高度反演

[Objective]Based on the Random Forest method,forest canopy height estimation models corresponding to coniferous forest,broadleaf forest and mixed forest were established by synergizing ICESat-2 lidar data,Sentinel-1 SAR data,Sentinel-2 optical images and topographic data,and then the accuracy of the established models for different forest types were validated and compared.Meanwhile,the optimal variable set corresponding to each forest type and importance score of each variable in the set was obtained with the help of Random Forest,and then the contribution of each optimal variable in the model establishment process was compared and analyzed quantitatively.[Method]Firstly,the accuracy of the RH95 height metric provided by ATL08 product of ICESat-2 satellite is validated by comparing with the canopy height derived from airborne lidar data at different spatial resolutions.Subsequently,elevation and slope data provided by SRTM product,VV and VH provided by Sentinel-1 SAR data,texture features,vegetation index and biophysical features provided by Sentinel-2 optical images were obtained.And then the cover information of coniferous/broadleaf/mixed forest within the study area was obtained through the NGCM(National Geographical Condition Monitoring Project)data.Based on the Random Forest method and the above derived parameters,the canopy height estimation models for whole forest(Not distinguishing forest types)and coniferous/broadleaf/mixed forest were established,and then the accuracy of each model was validated and compared.Finally,the optimal variable set corresponding to different forest types was obtained though Random Forest method and the contribution of each optimal variable in establishing the estimation model was compared and analyzed quantitatively.[Results]The RH95 height metric provided by ATL08 product has the best consistency with the canopy height derived from airborne lidar data at a spatial resolution of 250 mwith the R and RMSE of 0.80 and 1.98 m,respectively.The accuracy of canopy height estimation model established without distinguishing forest types is lower than that of estimation models established for coniferous/broadleaf/mixed forest.The two accuracy indicators R and RMSE corresponding to whole/coniferous/broadleaf/mixed forest are 0.59,0.72,0.59,0.62 and 3.68 m,3.15 m,3.37 m,3.26 m,respectively.In the process of model establishment,the optimal variable set corresponding to different models contains different types of variables and the number of variables within each set varies obviously,and there are also significant differences in the importance scores of the same variable in different sets.[Conclusion]Spatially continuous canopy height for a large area with high precision can be obtained by the estimation models established in this study by synergizing multi-source data.And the accuracy of estimation models considering forest types is significantly higher than that of estimation models established for whole forest,proving the necessity of establishing estimation models for different forest types.There are obvious differences in the optimal variables corresponding to different forest types,but the sum of importance scores of vegetation indices extracted from optical data in each set is higher than that of other types of variables,demonstrating the importance of vegetation indices in predicting forest canopy height through multi-source data.