A machine learning-based approach for smart agriculture via stacking-based ensemble learning and feature selection methods
Emna Ben Abdallah, Rima Grati, Khouloud Boukadi
Abstract
Smart irrigation has many advantages in optimizing resource usage (e.g., saving water, reducing energy consumption) and improving crop productivity. In this paper, we contribute to this field by proposing a robust and accurate machine learning-based approach that combines the power of feature selection methods and stacking ensemble method to effectively determine the optimal quantity of water needed for a plant. Random Forest, Recursive Feature Elimination (RFE), and SelectKBest are used to assess the importance of the features. Then, based on the best subset of features, a stacking ensemble model is proposed that combines CART, Gradient Boost Regression (GBR), Random Forest (RF) and XGBoost regressors. The different models involved in this approach are trained and tested using a collected dataset about various crops such as tomatoes, grapes, and lemon and encompasses different features such as meteorological data, soil data, irrigation data, and crop data. The experiments demonstrated the performance of RF in analyzing the feature importance. The findings of feature selection highlight the importance level of the evapotranspiration, the depletion, and the deficit to maximize the model’s accuracy. The results also showed that the proposed stacking model (Stacking_GBR+CART+RF+XGB) with the 10 most essential features outperforms individual models and other stacking models by achieving low error rates (i.e., MSE=0.0026, MAE=0.0279, RMSE=0.0509) and high R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score (i.e., 0.9927).