Prediction of aboveground grassland biomass on the Loess Plateau, China, using a random forest algorithm

Yinyin Wang,Kaibo Wang,Gaolin Wu,Zhuangsheng Tang,Wenyi Sun,Lei Deng,Zhouping Shangguan

doi:10.1038/s41598-017-07197-6

Abstract

Grasslands are an important component of terrestrial ecosystems that play a crucial role in the carbon cycle and climate change. In this study, we collected aboveground biomass (AGB) data from 223 grassland quadrats distributed across the Loess Plateau from 2011 to 2013 and predicted the spatial distribution of the grassland AGB at a 100-m resolution from both meteorological station and remote sensing data (TM and MODIS) using a Random Forest (RF) algorithm. The results showed that the predicted grassland AGB on the Loess Plateau decreased from east to west. Vegetation indexes were positively correlated with grassland AGB, and the normalized difference vegetation index (NDVI) acquired from TM data was the most important predictive factor. Tussock and shrub tussock had the highest AGB, and desert steppe had the lowest. Rainfall higher than 400 m might have benefitted the grassland AGB. Compared with those obtained for the bagging, mboost and the support vector machine (SVM) models, higher values for the mean Pearson coefficient (R) and the symmetric index of agreement (λ) were obtained for the RF model, indicating that this RF model could reasonably estimate the grassland AGB (65.01%) on the Loess Plateau.

Highlights

Grasslands are indispensable terrestrial ecosystems[1,2,3,4] for maintaining the ecological balance of arid and semi-arid regions under global climate change[5,6,7]
(1) Can an random forest (RF) model be used to predict the grassland aboveground biomass (AGB) on the Loess Plateau using meteorological and Remote sensing (RS) data? (2) What is the spatial distribution of the grassland AGB on the Loess Plateau? (3) How does the grassland AGB vary along the rainfall gradient? (4) How well does an RF model perform based on an accuracy assessment?
The tussock and shrub tussock vegetation types had the highest AGB, followed by forest steppe; in general, the AGB in forest steppe is higher than the AGB in typical steppe, which is higher than the AGB in desert steppe

Summary

OPEN Prediction of aboveground grassland biomass on the Loess

Received: 24 November 2016 Accepted: 27 June 2017 Published online: 31 July 2017 forest algorithm. Barrachina et al.[20] employed Landsat TM-5 data to estimate the AGB in mountain meadows and pastures, and Li et al.[21] developed a pure vegetation index model to predict the grassland AGB in the Inner Mongolian region of China. These studies indicated that AGB assessment using RS data is feasible, but the study areas were so different from the Loess Plateau that the fit of these models in that context cannot be validated. We attempted to predict the grassland AGB across the Loess Plateau, to understand the large-scale spatial characteristics of grasslands in this region by addressing the following questions. (1) Can an RF model be used to predict the grassland AGB on the Loess Plateau using meteorological and RS data? (2) What is the spatial distribution of the grassland AGB on the Loess Plateau? (3) How does the grassland AGB vary along the rainfall gradient? (4) How well does an RF model perform based on an accuracy assessment?

Results

Methods

Horizontal Date of image and vertical acquisition tile number

Longitude Latitude

Additional Information