Comparison of Forest Stock Volume Inversion Methods Coupled With Multiple Features - A Case Study of Forest in Yarlung Zangbo River Basin

Remote sensing monitoring of forest resources is one of the important application directions of remote sensing. Traditional measurement methods cost a lot of workforce and material resources. Scientific forest resource prediction can improve work efficiency and reduce measurement costs. Forest stock volume is an important index to evaluate the quality of forest ecosystems. The forest stock volume inversion model is a mathematical model used to estimate the forest stock volume, which has the functions of learning and prediction. The same ground features are quite different in different light or shadow areas. The band ratio can be used to reduce the error of the results in modeling light and shadow areas to a certain extent. The forest stock volume prediction model usually selects spectral information and texture features as the main modeling factors. It does not fully consider the impact of different models on the prediction accuracy when selecting multi-characteristic variables such as band ratio, vegetation index, and topographic factors. In order to compare the accuracy of different models, this article takes Milin County in Tibet Autonomous Region as the research area, and uses Landsat OLI images, DEM data and forest resource survey data as data sources to extract analyze spectral information, texture features and topographic factors. Three forest volume inversion models based on multi-features are established. The three models are multiple stepwise regression models, BP neural network models and random forest models. The effects of different methods on the inversion of forest stock are studied. The coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) are used to evaluate the fit and accuracy of the model. The results showed that the fit and accuracy of the random forest model are the best (R2=0.739, MAE=55.352 m3·ha-1, RMSE=63.195 m3·ha-1). The result is higher than the multiple stepwise regression model (R2=0.541, MAE=58.317 m3·ha-1, RMSE=71.562 m3·ha-1) and BP neural network model (R2=0.477, MAE=67.503 m3·ha-1, RMSE=73.226 m3·ha-1). The predicted value range of the model is 121.3~372.8 m3·ha-1 and it is relatively close to the actual value. The results showed that the inversion of forest stock volume based on multi-features is effective in practical applications, and different models have different effects on the inversion accuracy of forest stock volume. The random forest regression model has the highest accuracy in this inversion study of forest stock volume, and it can be better applied to remote sensing monitoring of forest resources. This study can provide a reference for selecting forest stock volume inversion methods and help continuously improve the forest resource remote sensing monitoring system. © 2022 Science Press. All rights reserved.