Estimation of tree biomass with GF-2

Biomass data from a total of 192 plots (112 pure forests of Chinese fir, 80 pure forests of Pinus massoniana) in Jiangle State Forest Farm in Sanming City, Fujian Province were obtained from field measured sample plot data and second-class survey data of Fujian Province．Two scene GF-2 images from the study area were preprocessed, spectral information, vegetation index, texture features and topographic factors were extracted, factors highly-correlated with the biomass as independent variables were screened out．Biomass models of fir and Pinus massoniana were established from support vector machine, random forest and multiple stepwise regressions．Fitting of the two machine learning models was found to be better than the multiple stepwise regression model．The random forest model showed the highest determination coefficient R² (0.65 and 0.72 for the 2 plots), and the highest estimation accuracy (65.28% and 76.82%)．The mean root square errors in the 3 models for the Chinese fir plot were 64.27, 48.16 and 77.03．The mean root square errors in the three models for the Pinus massoniana plot were 54.79,48.16 and 65.63, with the random forest model showing the lowest value．It is concluded that the random forest model is the most optimal among all three models．