Transferability Of Renmote Sensing Based Models For Estimating Aboveground Forest Biomass

Remote sensing is the most important means of forest biomass estimation. Researchers have focused on the modeling methods for forest biomass. There are many research on forest biomass estimation models using remote sensing. The model was built in one research area, the prediction of its forest biomass is higher, but when the model was applied to other areas, the forest biomass estimation precision was reduced, so the transferability of model is poor.To solve this problem, this paper using remote sensing data of Landsat5 TM and the data of forest management inventory, multiple linear regression model, partial least square regression model, random forest regression model, and Erf-BP neural network model were built for Jinyun, Jiangshan, Dongyang and Tiantai four counties.Then we compared the four kinds of model accuracy for one place.The transferability of aboveground forest biomass estimation model was analyzed and evaluated. The models for one study area with better accuracy were transplanted to the other three areas, and results of model transferability were tested using the actual ground survey samples. Last,the model for one area with better accuracy were applied to the other three areas, the model transferability were used sample data for analaysis and evaluation.The main conclusions of the study are as follows:(1) Multiple linear regression model, partial least square regression model, random forest regression model, and Erf-BP neural network model were built for Jinyun, Jiangshan, Dongyang and Tiantai, had come to the same conclusion: random forest regression have the highest accuracy, followed by the Erf-BP neural network, partial least square regression model and multiple linear regression model.(2) Choose the three better precision models for one area to the other three areas: partial least square regression model, random forest regression model, and Erf-BP neural network model.The final results: the random forest regression and Erf-BP neural network model were superior to the partial least squares regression. The partial least square regression model built in one area tranplanted to other three areas, the prediction of forest biomass was higher or lower than the actually forest biomass,. The random forest regression and Erf-BP neural network model applied to other areas, the predivtion of forest biomass close to the actually forest biomass, they have strong tansferabilty.(3) The sample quantity and quality, the selection of independent variable and the model type have impact on the transferability of those model.Random forest regression and Erf-BP neural network model, which is better transferability need to consider other factors. And also, more deeply research in this aspect need to do more work to explore and discuss in future.