Evaluation of Machine Learning Models for Estimating Grassland Pasture Yield Using Landsat-8 Imagery

Abstract

Accurate estimation of pasture yield in grasslands is crucial for the sustainable utilization of pasture resources and the optimization of grassland management. This study leveraged the capabilities of machine learning techniques, supported by Google Earth Engine (GEE), to assess pasture yield in the temperate grasslands of northern China. Utilizing Landsat-8 data, band reflectances, vegetation indexes (VIs), and soil water index (SWI) were extracted from 1000 field samples across Xilingol. These data, combined with field-measured pasture yields, were employed to construct models using four machine learning algorithms: elastic net regression (Enet), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM). Among the models, XGBoost demonstrated the best performance for pasture yield estimation, with a coefficient of determination (R2) of 0.94 and a precision of 76.3%. Additionally, models that incorporated multiple VIs demonstrated superior prediction accuracy compared to those using individual VI, and including soil moisture data further enhanced predictive precision. The XGBoost model was subsequently applied to map the spatial patterns of pasture yield in the Xilingol grassland for the years 2014 and 2019. The estimated average annual pasture yield in the Xilingol grassland was 1042.38 and 1013.49 kg/ha in 2014 and 2019, respectively, showing a general decreasing trend from the northeast to the southwest. This study explored the effectiveness of common machine learning algorithms in predicting pasture yield of temperate grasslands utilizing Landsat-8 data and ground sample data and provided the valuable support for long-term historical monitoring of pasture resources. The findings also highlighted the importance of predictor selection in optimizing model performance, except for the reflectance and vegetation indices characterizing vegetation canopy information, the inclusion of soil moisture information could appropriately improve the accuracy of model predictions, especially for grasslands with relatively low vegetation cover.