An Integrated Machine Learning Framework for Citrus Grading and Yield Prediction Using Classifier-Informed Regression

Agriculture remains a cornerstone of human sustenance, yet accurate citrus fruit grading and yield predictions remain significant challenges due to visual similarities between varieties such as oranges and grapefruits. This includes the inefficiencies of manual grading. This study presents a new dual-function machine learning system that integrates classification and regression to simultaneously grade citrus fruits and forecast crop yields. This study utilizes a citrus classification dataset comprising 10,000 samples (5,000 oranges and 5,000 grapefruits) characterized by 5 features (diameter, weight, and RGB color values), and a crop yield dataset spanning 1990-2022 from 39 countries across Africa (17 countries, 1,122 records) and the Americas (22 countries, 1,452 records) totaling 2,574 records with environmental and nutrient features. An Artificial Neural Network (ANN) classifier first distinguishes visually similar citrus types, which achieves a grading accuracy of 98.5% with minimal variance. This output then informs a Random Forest (RF) regressor, which predicts yields with a high degree of precision (R² = 0.905). Compared to existing methods, such as CNN-SVM and XGBoost-based approaches that achieve lower accuracy and R² scores, such as 90.6% and 0.853, respectively, the proposed system demonstrates superior performance across both tasks. The integrated ANN-RF pipeline architecture uses classifier predictions as informative features for the regression model. This integrated design aligns with real-world agricultural practices that demand concurrent quality control and forecasting with improved predictive reliability. The system’s modular architecture allows adaptation across different geographies, reinforcing its relevance to precision agriculture and food security.