A method to avoid spatial overfitting in estimation of grassland above-ground biomass on the Tibetan Plateau. (June 2021)

• Record Type: Journal Article
• Authors: Yu, Hui; Wu, Yufeng; Niu, Liting; Chai, Yafan; Feng, Qisheng; Wang, Wei; Liang, Tiangang
• Is Part Of: Ecological indicators. Volume 125(2021)
• Publication Date: 2021-06
• Subjects: Alpine grassland -- Above-ground biomass -- Random forest -- Cross-validation -- Overfitting; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2021.107450
• Languages: English
• ISSNs: 1470-160X

Abstract

Highlights: The FFS was able to improve the performance of RF model to predict the AGB. The LLO-CV could reduce the spatial overfitting for AGB estimation. The NDVI, AAT and AP were most important to accurately estimate AGB. The grassland AGB increased more than it decreased on Tibetan Plateau (2001–2018). Abstract: Accurate assessments of grassland above-ground biomass (AGB) are crucial for the sustainable utilization and protection of grassland resources and the eco-environment. In this study, a random forest (RF) model combined with the forward feature selection (FFS) and leave-location-out cross-validation (LLO-CV) methods was trained to predict the dry weight (DW) of grassland AGB based on multiple factors. The final model exhibited a performance of R 2 = 0.66, root mean square error (RMSE) of 503.86 kg DW/ha and mean absolute error (MAE) of 376.51 kg DW/ha. The spatial distribution of grassland AGB increased from northwest to southeast over the entire Tibetan Plateau (TP) from 2001 to 2018. Grassland AGB increased more than it decreased (70.6% vs 29.4%, respectively) during the study period. Using a combination of FFS and LLO-CV, spatial overfitting was reduced, and the predictive accuracy of the RF was improved, thus enhancing the ability to predict the AGB in unknown locations from training data. This study proposes a robust methodology with which to improve the transferability of machine learning algorithms to predict grassland AGB in unknown locations.