Modeling Alpine Grassland Above Ground Biomass Based on Remote Sensing Data and Machine Learning Algorithm: A Case Study in East of the Tibetan Plateau, China

Abstract

Effective and accurate assessment of grassland above-ground biomass (AGB) especially via remote sensing (RS), is crucial for forage-livestock balance and ecological environment protection of alpine grasslands. Because of complexity and extensive spatial distribution of natural grassland resources, the RS estimation models based on moderate resolution imaging spectroradiometer (MODIS) data exhibited low accuracy and poor stability. In this study, various methods for estimating the AGB of alpine grassland vegetation using MODIS vegetation indices were evaluated by combining with meteorology, soil, topography geography and in situ measured AGB data (during grassland growing season from 2011 to 2016) in Gannan region. Results show that 1) five out of ten factors (elevation, slope, aspect, topographic position, temperature, precipitation and the concentration of clay and sand in the soil) exert significant effects on grassland AGB, with R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> 0.04-0.39, and RMSE 859.68-1075.09 kg/ha, respectively; 2) the accuracy and stability of AGB estimation model can be improved by constructing multivariate models, especially using multivariate nonparameter models; 3) the optimum estimation model is constructed on the basis of random forest algorithm (RF). Compared with univariate/multivariate parameter models, RMSE of RF model decreased 26.45%-44.27%. Meanwhile, RF models can explain 89.41% variation in AGB during grass growing season. This study presented a more suitable RS inversion model integrated MODIS vegetation indices and other effect factors. Besides, the accuracy based on MODIS data was greatly improved. Thus, our study provides a scientific basis for effective and accurate estimating alpine grassland AGB.