Predictive modelling employing machine learning, convolutional neural networks (CNNs), and smartphone RGB images for non-destructive biomass estimation of pearl millet (Pennisetum glaucum)

Faten Dhawi^1*Abdul Ghafoor^2*Norah Almousa³Sakinah Ali³Sara Alqanbar³

• ¹Department of Agricultural Biotechnology, College of Agricultural and Food Sciences, King Faisal University, Al-Ahsa, Eastern Province, Saudi Arabia
• ²Center for water and Environmental Studies, king Faisal University, Al Ahsa, Saudi Arabia
• ³Fab Lab, Abdulmonem Al Rashed Humanitarian Foundation, Al-Ahsa, Saudi Arabia, Al Ahsa, Saudi Arabia

Digital tools and non-destructive monitoring techniques are crucial for real-time evaluations of crop output and health in sustainable agriculture. Sustainable agriculture and environmental stewardship require precise AGB computation since it provides valuable insights into efficient crop management and carbon balance in ecosystems.This study uses cutting-edge computer vision and machine learning techniques to propose a novel, non-invasive approach for quantifying above-ground biomass (AGB) of pearl millet. We employed a transfer learning approach, leveraging pre-trained CNN models alongside shallow machine learning algorithms-such as Support Vector Regression (SVR), XGBoost, and Random Forest Regression (RFR). Smartphone-based RGB imaging was utilized for data collection. Convolutional neural networks (CNNs) were trained to predict the above-ground biomass of pearl millet using transfer learning, which incorporated the VegAnn model and its parameters into the convolutional layers of our model. Shapley additive explanations (SHAP) methodology was employed to evaluate predictor importance systematically. The SHAP analysis confirmed that the most influential features (such as Normalized Green-Red Difference Index, NGRDI, and plant height) substantially contributed to AGB estimation accuracy, whereas features with low SHAP values might be excluded from the final model without compromising predictive performance. A comparison of four machine learning models was performed using several feature sets, including all features, the five most significant, and the two most salient. The study found that XGBoost has a comprehensive feature set (R² = 0.98, RMSE = 0.26), while CNN-based models also showed high predictive ability. Notably, RFR performs best with the two most important features, whereas SVR is the least effective model throughout the analysis. These findings demonstrate the effectiveness of CNNs and shallow machine learning in estimating AGB non-invasively using cost-effective RGB imagery. The results show that our technique may be utilized to support automated biomass prediction and real-time plant growth monitoring. These tools could form the basis for small-scale carbon inventories to measure the carbon sequestered within vegetation biomass in smallholder agricultural systems. These inventories are crucial for understanding how agricultural practices contribute to carbon sequestration at the local level and for informing climateresilient strategies