Data-driven multisource estimation of aboveground biomass density: a baseline for rangeland monitoring

Monitoring the condition and degradation of rangeland ecosystems is essential for sustainable rangeland management. Aboveground biomass density (AGBD) is a key indicator of rangeland health, providing insights into forage availability for pastoralists and climate change mitigation strategies. This work applied open access earth observation data to generate spatially continuous, high-resolution AGBD maps for an important pastoralist area in Ethiopia, where season-specific AGBD assessments remain limited. This study employed recursive feature elimination with cross-validation using random forest (RFECV_RF) with fivefold cross-validated grid search for variable selection and hyperparameter tuning. Convolutional neural network (CNN) and RF regression models were applied to estimate AGBD at 50-m resolution across seasonal (short and main rainy seasons) and combined annual datasets, using spaceborne Global Ecosystem Dynamics Investigation LiDAR measurements (GEDI L4A) data for training and validation. Predictor variables included Sentinel-1/2 backscatter and spectral bands, vegetation indices, topographic factors, and precipitation data. Both models performed well, with CNN consistently outperforming RF. For the combined annual analysis, CNN achieved a root mean square error (RMSE) of 4.77 t/ha, relative RMSE of 46.7%, a mean absolute error (MAE) of 2.55 t/ha, with a slight underestimation bias (mean error, ME -0.22 t/ha), and the coefficient of determination (R²) of 0.93. Errors were lower in the short rainy season (RMSE 4.88, MAE 1.90 t/ha) than in the main rainy season (RMSE 6.99, MAE 3.69 t/ha). This work establishes a novel, scalable framework and provides a critical high-resolution AGBD baseline to detect degradation hotspots and support season-specific strategies for sustainable grazing, restoration, and pastoral resilience.