Abstract: Estimation models of aboveground biomass change of Pinus densata in Shangri-La were established based on remote sensing. The estimation accuracy of the models with or without terrain factors being considered was compared. Based on the continuous inventory plots and the Landsat TM/OLI images during 1987 to 2017, the remote sensing factors were extracted, and the change rates of those remote sensing factors and aboveground biomass of the corresponding plots were calculated. Multiple linear regression (MLR), random forest (RF) and gradient boosting regression tree (GBRT) were then used in the modeling. Finally, the factors of elevation, slope and aspect were added to these three models and their modeling effects were compared. The change rate models constructed with the three methods varied a lot, with the fitted results of R² being 0.457, 0.946 and 0.887, the RMSE being 2.431, 0.692 and 1.027 t·hm^-2·a^-1, the rRMSE of predicted effect being 56.66%, 33.17% and 35.30%, and the precision being 44.31%, 78.77% and 70.95%, respectively. The accuracy of the three models was improved, and the confidence intervals of RF and GBRT were narrowed after adding terrain factors. In all models, the result of RF model with slope aspect factor was best, the R² was 0.976 (with an enhancement of 3.1%), the RMSE was 0.502 t·hm^-2·a^-1 (a reduction of 27.46%), the rRMSE was 31.50% (a reduction of 5.05%), the precision was 82.07%, (an increase of 4.18%). Adding terrain factor into the change rate model of RF can improve the accuracy of the model, and the aspect factor has the best performance on improving the model accuracy. Our results can provide a useful basis for accurate estimation of aboveground biomass of Pinus densata in Shangri-La.

Key words: Landsat, random forest, gradient boosting regression tree, change rate, topographic factor, aboveground biomass.