Abstract
Digital tools and non-destructive monitoring techniques are crucial for real-time evaluations of crop output and health in sustainable agriculture, particularly for precise above-ground biomass (AGB) computation in pearl millet (Pennisetum glaucum). This study employed a transfer learning approach using pre-trained convolutional neural networks (CNNs) alongside shallow machine learning algorithms (Support Vector Regression, XGBoost, Random Forest Regression) to estimate AGB. Smartphone-based RGB imaging was used for data collection, and Shapley additive explanations (SHAP) methodology evaluated predictor importance. The SHAP analysis identified Normalized Green-Red Difference Index (NGRDI) and plant height as the most influential features for AGB estimation. XGBoost achieved the highest accuracy (R(2) = 0.98, RMSE = 0.26) with a comprehensive feature set, while CNN-based models also showed strong predictive ability. Random Forest Regression performed best with the two most important features, whereas Support Vector Regression was the least effective. These findings demonstrate the effectiveness of CNNs and shallow machine learning for non-invasive AGB estimation using cost-effective RGB imagery, supporting automated biomass prediction and real-time plant growth monitoring. This approach can aid small-scale carbon inventories in smallholder agricultural systems, contributing to climate-resilient strategies.