A state-of-the-art novel approach to predict potato crop coefficient (Kc) by integrating advanced machine learning tools

Abstract

The crop coefficient (K_c) is one of the important elements of the actual evapotranspiration estimation. The current study aims to develop a machine learning approach to estimate the crop coefficient of potatoes (Russet Burbank variety) in Prince Edward Island province, one of Canada's most important producers. The study was conducted at drainage-type lysimeters placed in the potato field with three types of soils (sandy loam, loamy sand, and loam). A machine learning approach using XGBoost, optimized with the Chaos Game algorithm (CGO-XGBoost), was employed to predict K_c. Three input scenarios (meteorological + soil data, soil-only, meteorological-only) were tested. Three other machine learning techniques, K-nearest neighbor (KNN), Adaptive Boosting (AdaBoost), and Multilayer Perceptron Neural Network (MLP), were used to compare with the newly developed model. Different performance metrics such as correlation coefficient (R), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) were used to compare different model's performance. Results showed that the CGO-XGBoost model outperformed conventional machine learning models. A comparison of different input scenarios revealed that combination 2 (Soil data only) gave the best results. Combination 3 (only meteorological data) performs weakest among input scenarios. The best model (Combination2 + CGO-XGBoost) achieved the highest accuracy (R = 0.9961, RMSE = 0.0185, MAPE = 2.20%), outperforming traditional methods. SHapley Additive exPlanations (SHAP) interpretability analysis indicates that soil moisture exerts the greatest impact on potato K_c. Field Capacity (FC) and Minimum temperature rank as the second and third most significant factors. The integration of SHAP values in the proposed solution improves the interpretability of the model, offering valuable insights into the environmental and soil factors affecting K_c predictions. The results showed that the proposed model can accurately predict K_c, demonstrating its potential to enhance water-use efficiency and support precision irrigation strategies.