Modeling Biomass for Natural Subtropical Secondary Forest Using Multi-Source Data and Different Regression Models in Huangfu Mountain, China

Forest biomass estimation is an important parameter for calculating forest carbon storage, which is of great significance for formulating carbon-neutral strategies and forest resource management measures. We aimed at solving the problems of low estimation accuracy of forest biomass with complex canopy structure and high canopy density, and large differences in the estimation results of the same estimation model under complex forest conditions. The Huangfu Mountain Forest Farm in Chuzhou City was used as the research area. As predictors, we used Gaofen-1(GF-1) and Gaofen(GF-6) satellite high-resolution imaging satellite data, combined with digital elevation model (DEM) and forest resource data. Multiple stepwise regression, BP neural network and random forest estimation models were used to construct a natural subtropical secondary forest biomass estimation model with complex canopy structure and high canopy closure. We extracted image information as modeling factors, established multiple stepwise regression models of different tree types with a single data source and a comprehensive data source and determined the optimal modeling factors. On this basis, the BP neural network and random forest biomass estimation model were established for Pinus massoniana, Pinus elliottii, Quercus acutissima and mixed forests, with the coefficient of determination (R2) and root mean square error (RMSE) as the judgment indices. The results show that the random forest model had the best biomass estimation effect among different forest types. The R2 of Quercus acutissima was the highest, reaching 0.926, but the RMSE was 11.658 t/hm2. The R2 values of Pinus massoniana and mixed forest were 0.912 and 0.904, respectively. The RMSE reached 10.521 t/hm2 and 6.765 t/hm2, respectively; the worst result was the estimation result of Pinus elliottii, with an R2 of 0.879 and an RMSE of 14.721 t/hm2. The estimation result of the BP neural network was second only to that of the random forest model in the four forest types. From high precision to low precision, the order was Quercus acutissima, Pinus massoniana, mixed forest and Pinus elliottii, with R2s of 0.897, 0.877, 0.825 and 0.753 and RMSEs of 17.899 t/hm2, 10.168 t/hm2, 18.641 t/hm2 and 20.419 t/hm2, respectively. In this experiment, the worst biomass estimation performance was seen for multiple stepwise regression, which ranked the species in the order of Quercus acutissima, Pinus massoniana, mixed forest and Pinus elliottii, with R2s of 0.658, 0.622, 0.528 and 0.379 and RMSEs of 29.807 t/hm2, 16.291 t/hm2, 28.011 t/hm2 and 23.101 t/hm2, respectively. In conclusion, GF-1 and GF-6 combined with data and a random forest algorithm can obtain the most accurate results in estimating the forest biomass of complex tree species. The random forest estimation model had a good performance in biomass estimation of primary secondary forest. High-resolution satellite data have great application potential in the field of forest parameter inversion.