Assessment of Machine Learning Methods for Modeling Alpine Grassland Biomass in Southern Qinghai Province, China

Tong Liang, Qisheng Feng, Jing Ge, Hongjie Xie, Tiangang Liang

Abstract: Accurate and effective modeling of grassland aboveground biomass (AGB) is an important basic task in monitoring and management of grassland and livestock interaction in pastoral areas. In this study, three machine learning methods, multi-layer perceptron network (MLP), support vector regression, and random forest regression (RF) are assessed for modeling the grassland AGB in the southern region, Qinghai Province, China. The results show that: 1) among the three methods, the MLP model performs the worst, with R2 and RMSE of 0.38 and 768.74 kg DW/ha, respectively for the test data, and of 0.34 and 745.06 kg DW/ha for the training dataset; 2) the RF model performs the best, with R2 and RMSE of 0.76 and 473.20 kg DW/ha, respectively for the test data, and of 0.95 and 208.88 kg DW/ha for the training dataset. This performance is similar to the best back propagation ANN model previously reported (0.66 and 556.57 kg DW/ha for test data, 0.85 and 355.04 kg DW/ha for training data, and 0.68 and 537.09 kg DW/ha for validation). The RF model is easy to apply in practice and is a robust tool for modeling grassland biomass based on DEM and remote sensing data.