Precision Agriculture Using AI-Based Crop Yield Prediction Models

Abstract

Accurate crop yield prediction is central to the adoption of precision agriculture and the mitigation of food-security risks
under intensifying climate variability. This study evaluates five machine-learning and deep-learning algorithms--Random
Forest (RF), Gradient Boosting Machine (GBM), Long Short-Term Memory networks (LSTM), Convolutional Neural
Networks (CNN), and Support Vector Regression (SVR)--for season-long yield forecasting of wheat, maize, and rice
across three agro-climatic zones in Italy. Multi-source input features comprising remote-sensing vegetation indices
(NDVI, EVI), soil physicochemical properties, cumulative growing-degree days (GDD), and meteorological parameters
were integrated over a six-year observational period (2018-2024). Model performance was assessed using Root Mean
Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R2). LSTM
achieved the highest predictive accuracy (R2 = 0.947, RMSE = 0.31 t/ha) followed by GBM (R2 = 0.921), demonstrating
the superiority of sequential deep-learning architectures in capturing temporal phenological dynamics. RF yielded the
highest feature-importance consistency, identifying GDD and NDVI at heading stage as the dominant yield determinants.
The findings advocate a hybrid modelling framework combining LSTM temporal encoders with RF-derived feature
selection for operational deployment in precision-farming decision-support systems.