Integrating Proximal and Remote Sensing with Machine Learning for Pasture Biomass Estimation

This study tackles the challenge of accurately estimating pasture biomass by integrating proximal sensing, remote sensing, and machine learning techniques. Field measurements of vegetation height collected using the PaddockTrac ultrasonic sensor were combined with vegetation indices (e.g., NDVI, MSAVI2) derived from Landsat 7 and Sentinel-2 satellite data. We applied the Boruta algorithm for feature selection to identify influential biophysical predictors and evaluated four machine learning models-Linear Regression, Decision Tree, Random Forest, and XGBoost-for biomass prediction. XGBoost consistently performed the best, achieving an R² of 0.86, an MAE of 414 kg ha⁻¹, and an RMSE of 538 kg ha⁻¹ using Landsat 7 data across multiple years. Sentinel-2's red-edge indices did not substantially improve predictions, suggesting a limited benefit from finer spectral resolutions in this homogenous pasture context. Nonetheless, these indices may offer value in more complex vegetation scenarios. The findings emphasize the effectiveness of combining detailed ground-based measurements with advanced machine learning and remote sensing data, providing a scalable and accurate approach to biomass estimation. This integrated framework provides practical insights for precision agriculture and optimized pasture management, significantly advancing efficient and sustainable rangeland monitoring.