Estimation of Grassland Height Based on the Random Forest Algorithm and Remote Sensing in the Tibetan Plateau

Abstract

Grassland height is one of the main factors used to evaluate grassland conditions. However, the retrieval of natural grassland height at the regional scale by remote sensing data and conventional statistical models will result in large errors, especially in the heterogeneous alpine grassland of the Tibetan Plateau (TP). In this article, we aimed to construct a model based on multiple variables (biogeographical, meteorological, and Moderate Resolution Imaging Spectroradiometer (MODIS) product) using a random forest (RF) algorithm to predict the spatial distribution of grassland height in the TP from 2003 to 2017. The results show the following conditions. 1) Seven variables (elevation, slope, aspect, enhanced vegetation index, reflectance in band seven of MODIS (B7), annual accumulated temperature (≥0 °C), and annual precipitation) that were selected by recursive feature elimination from 11 variables have high importance in the RF model. The final model exhibits good performance, with mean R2 and root mean squared error values of 0.51 and 6.15 cm, respectively, which were determined via 10-fold crossvalidation. 2) The mean grassland height (2003-2017) predicted by the RF model ranges from 5 to 10 cm in most areas of the TP, and the mean height is 10 cm. The grassland height in the east and southeast of the TP is significantly higher than that in other areas. 3) This article achieves a relatively accurate estimation of grassland height over a large spatial scale at 500-m spatial resolution, which plays an important role in accurately estimating aboveground biomass and evapotranspiration over grassland.

Keywords

• Grassland height
• random forest (RF)
• Tibetan Plateau (TP)