Based on the multi-scale geographically weighted regression (MGWR) model, combined with remote sensing data, the above ground carbon (AGC) of forests in Liaoning Province was estimated, and the impacts of spatial heterogeneity on the estimation of AGC were assessed through the selection of different characteristic variables.
Using 1 068 forest inventory sample plots, digital elevation model (DEM) data and Landsat 8 OLI images in Liaoning Province in 2015 as data sources, 63 feature variables (2 topographic features, 6 band features, 7 vegetation indices, 48 texture features) were extracted and selected the feature variables using stepwise regression and random forest (RF) algorithms, and construct ordinary least squares (OLS), geographically weighted regression (GWR), MGWR and random forest regression (RFR), MGWR model and Random Forest Regression (RFR) model to estimate AGC, compare the differences between the prediction accuracy of different models, and realize the accurate inversion of AGC in the study area.
1) The prediction accuracies of each model under the stepwise regression method of selection feature variables are higher than those of the RF method of selecting feature variables, and the MGWR model has the highest accuracy (R2=0.74, RMSE=11.66 t/hm2); the GWR model is the next highest (R2=0.72, RMSE=12.30 t/hm2); and the RFR model is the next highest (R2=0.37, RMSE=16.32 t/hm2). The OLS model is the lowest (R2=0.26, RMSE=16.41 t/hm2); 2) In the stepwise regression of the feature selection method, the MGWR model improves the R2 by 0.48, 0.02, and 0.37, respectively; and decreases the RMSE by 4.75 t/hm2, 0.64 t/hm2, and 4.66 t/hm2, respectively. The R2 of the MGWR model is improved by 0.46, 0.05, and 0.33 compared with those of OLS, GWR, and RFR, respectively, while the RMSE is reduced by 6.95 t/hm2, 1.23 t/hm2, and 5.36 t/hm2, respectively, in the selection of feature variables by the RF method; 3) The prediction of aboveground carbon stock in forests of the study area ranges from 37.25 t/hm2 to 190.22 t/hm2, totaling 135.165 Tg. Spatially, the high value of forest aboveground carbon stock is mainly distributed in the west and a small area in the east, while most of the eastern area decreases, and the central plains area is the lowest.
Stepwise regression selection of feature variables to construct the MGWR model effectively solves the problem of spatial heterogeneity through flexible scale (bandwidth) settings, reveals the multi-scale spatial distribution characteristics of AGC in Liaoning Province, which is applicable to the accurate estimation of large-scale AGC, and provides a scientific basis for the enhancement of regional carbon management and ecosystem services.
京公网安备11010802044758号