Comparing the performance of machine learning algorithms for estimating aboveground biomass in typical steppe of northern China using Sentinel imageries

Abstract

Monitoring aboveground biomass (AGB) is crucial for assessing, managing, and utilizing grassland ecosystems. While the technical form of combining remote sensing and machine learning algorithms is widely used to estimate AGB at a regional scale, few studies have assessed and compared the performance of popular algorithms on the typical steppe in northern China. In this study, the northern Xilinhot, a representative area of typical steppe in China, was selected as the study area to compare the performance of six widely used machine learning algorithms for AGB estimation, namely stepwise linear regression (SLR), partial least square regression (PLS), principal component regression (PCR), random forest (RF), support vector machines (SVM), and k-nearest neighbors (KNN). Additionally, the study explored the modeling capability of multisource variables from Sentinel imagery and auxiliary data. The results showed that (1) considering the aspects of prediction accuracy, noise resistance, ease of operation, and transferability, the SLR algorithm is more suitable for estimating typical steppe AGB in northern China at the Sentinel scale. (2) Vegetation Indices (VI) play a significant role in the development of selected models, with significant contributions from both traditional and soil-adjusted indices. (3) Sentinel C-band synthetic aperture radar (SAR) is unsuitable for modeling typical steppe AGB. (4) Among the selected environmental factors, only clay content and soil pH are significantly linearly correlated with AGB, while elevation, precipitation, temperature, soil pH, and sand content are advantageous for RF prediction. This study can provide important technical references for the research on AGB in typical steppe in northern China.

Keywords

• Remote sensing
• Grassland ecosystem
• Nonparametric algorithm
• Vegetation modelling