The Study Of Forest Biomass Estimation Model Based On Neural Network

Based on the analysis of correlation between each organ biomass (Aboveground biomass, Stem biomass, Crown biomass, Branch biomass and Leaf-flower-fruit biomass) and the stand factors(DBH, Height, Age, etc.) for Pinus massoniana, we have built linear, nonlinear and polynomial biomass models, and built the contrast analysis to compare multivariate models and single-variable models. The optimal BP-artificial-neural-network model has been built after the topological structure was determined. The topological structure of the optimal model was determined through screening of12algorithms and choosing of the number of inputs, outputs and hidden nodes. To explain the impact of the number of input variable on the accuracy, double-inputs BP model was compared with single-input BP model. To explain the impact of the number of output variable on the accuracy, multiple-outputs BP model was compared with single-output BP model. The optimal BP model was compared with allometric equations to verify the feasibility. The forest biomass BP-artificial-neural-network model has been built, using50plots data for training samples. The purpose is to estimate the forest biomass in Jiangle County, and to analyze the variation of biomass with altitude, slope, and aspect. The results show that:(1) Estimation results of the nonlinear model and polynomial model is better than the linear model. For linear and nonlinear models, multivariate models are superior to single-variable models. For polynomial models, high-order polynomial models are superior to low-order polynomial models. Models for aboveground biomass and stem biomass are superior to models for crown biomass, branch biomass and leaf-flower-fruit biomass.(2) The algorithm of optimal model is Levenberg-Marquardt algorithm, with the DBH and height inputs variable. The outputs are total weight, weight of above ground and weight of root. The number of hidden nodes is eight. Adding input variable and output variable would not decrease the accuracy of the BP neural network model. The optimal BP model had a good performance in estimating Pinus massoniana biomass. Its accuracy was higher than the relative growth model.(3) The fitting effect of the trilaminar BP model for forest biomass is good. The distribution map of forest biomass in Jiangle County has been generated by using this BP model. The result shows that there are obvious correlation between forest biomass and topographic factors.The purpose is to explore and verify the applicability of the BP neural network model on the biomass modeling and estimation, which makes the workload of biomass modeling and estimation more simply, and provides practical reference for the modeling work of the national forest biomass.