Regionalization Analysis of Environmental Drivers of CONUS Grazing Land Biomass

Grazing lands are essential ecosystems that contribute significantly to global food production and ecological balance. Effectively managing these lands requires precise biomass assessment, which is challenging due to diverse environmental conditions. This study focuses on regionalizing the analysis of grazing lands in the Conterminous United States by combining Earth observation data on Google Earth Engine, with machine learning techniques. Key environmental factors, including precipitation, elevation, land surface temperature, vegetation cover, and soil texture, were used for unsupervised clustering. We examined the correlation of each cluster's environmental factors with annual herbaceous above-ground biomass estimates from the Rangeland Analysis Platform. The study identified 18 distinct environmental clusters, highlighting the complex variability within grazing lands and underscoring the importance of considering multiple environmental factors for reliable biomass estimation. Investigating the performance of several machine learning approaches in reproducing RAP biomass, the Random Forest model performed best, with a mean absolute error (MAE) of 373 lb/acre and a coefficient of determination (R²) of 0.66. However, the incorporation of adaptive modeling through clustering did not yield a significant improvement in the accuracy of the Random Forest model likely due to its inherent capability to capture regional variations through complex decision tree structures. In contrast, the multiple linear regression model demonstrated a substantial improvement of around 16%, with the R² value increasing from 0.388 to 0.556. These findings underscore the effectiveness of regionalization in enhancing model interpretability and applicability while improving accuracy, providing a more efficient framework for large-scale biomass estimation and monitoring.