Model Transferability Based On Remotely Sensed Data For Moso Bamboo Forest Aboveground Carbon Storage Estimation

Quantitative estimation of forest carbon stocks using remote sensing is widely recognized, which isimportant progress of quantitative study on terrestrial vegetation carbon stocks. Remote sensing is oneof the important means of future forest carbon estimation and its dynamic change detection. Researcheson a variety of remote sensing estimation models for biomass carbon stocks have been reported. Thecharacteristics of those models are simple, but structurally diverse. The defects of those models arevulnerable to vegetation type, light conditions, observation position, and canopy structure, while arealso more sensitive to the non-vegetation factors, such as soil background. Therefore, the statisticalmodel with poor transferability is also a major problem in estimation of biomass carbon stocks usingremote sensing. Aimed to this issue and taken Lin'an City, Zhejiang Province, Anji County, LongquanCity as study areas, the transferability of bamboo forest aboveground biomass estimation model basedon remote sensing will be discussed in this study.This research mainly includes the following three aspects.1,Remote sensing data preprocessing for three study areas, including spatial registration, terraincorrection, atmospheric correction, and bamboo forest remote sensing information extraction.2,Linear and nonlinear models, stepwise regression model, multivariate linear model, and Erf-BPneural network model were built to estimate bamboo forest carbon stocks using remote sensing data.3,The transferability of bamboo forest aboveground biomass estimation model was analyzed andevaluated. The models for one study area with better accuracy were transplanted to the other two areas,and results of model transferability were tested using the actual ground survey samples.There are three main conclusions:1,For the three study areas, Erf-BP neural network model had the highest accuracy, followed by thestepwise regression model and nonlinear model.2,The transferability of the Erf-BP neural network model was superior to the stepwise regressionmodel and nonlinear model, and Erf-BP model had strong transferability.3,Model type and independent variables of model had a significant impact on the transferability ofthe statistical model.