Utilisation des modèles de machine learning pour l'estimation des baisses du rendements des cultures sous contraintes climatiques

Abstract

Climate change has emerged as one of the most critical challenges facing modern agriculture, with significant implications for global food security. To assess its impact on wheat yield reduction, we conducted a study estimating winter wheat yield declines using machine learning models from 1981 to 2022, focusing on the Sétif region of Algeria. First, we used the CROPWAT software to calculate yield reduction based on climatic variables (minimum/maximum temperature, rainfall, humidity, sunshine, wind speed, and reference evapotranspiration). Next, yield reductions were estimated using eight machine learning models: Support Vector Machines (SVM), K-Nearest Neighbors (K-NN), Adaptive Boosting (AdaBoost), Random Forest (RF), Gradient Boosting, Linear Regression, Neural Networks, and Decision Trees. The most influential climatic variables were identified using the RANK tool, supported by univariate regression analysis. Results revealed that relative humidity ranked highest (score:17.808), followed by reference evapotranspiration (score:12.812) and rainfall (score:12.418). Among the ML models, SVM demonstrated the best performance (R²= 0.801, , RMSE: 3.796%), closely followed by Random forest (R²: 0.758, RMSE: 4.184%). K-NN also showed promising results (R²: 0.746, RMSE: 4.287%)