Modeling Growing Stock Volume Using SAR Data and OBIA: Effects of Scale Parameter and Textural and Geometrical Features

Abstract

This study applies an object-based image classification approach to the modeling of growing stock volume (GSV) for three test sites in boreal forests of Central Siberia. Assessing GSV is of great importance in the context of climate change and modeling of the global carbon cycle. In this study, dual-polarized (HH and HV) L-band radar data are used. The main objective of this study is to improve the model accuracy of object-based GSV estimation. Thus, the applied methodology uses backscatter intensities as well as geometrical and textural features computed using Trimble eCognition Developer. Furthermore, the impacts of these feature groups and of different scale parameters on the model accuracy are analyzed. The scale parameter is of great importance in image segmentation, defining the size of the resulting objects. For modeling GSV, the random forest algorithm is used, and is trained using forest inventory data. The application of this method yields a coefficient of correlation (R²) between 0.42 and 0.51, and a relative root mean square error (RMSE) between 27% and 37%. These results reveal that the combined use of spectral, textural, and geometrical features and a smaller scale parameter enhance the model accuracy. These findings are encouraging and indicate that the model performance of object-based GSV estimation models can still be improved.