Machine learning modeling of water footprint in crop production distinguishing water supply and irrigation method scenarios
Zhibin Li,Wei Wang,Xiangxiang Ji,Pute Wu,La Zhuo
DOI: https://doi.org/10.1016/j.jhydrol.2023.130171
IF: 6.4
2023-09-18
Journal of Hydrology
Abstract:Crop water footprint (WF) calculation via physical process-based crop or hydrological models has basic requirements in terms of cost, time, and knowledge threshold. It is unsuitable for laypersons such as policymakers or cross-disciplinary researchers. Here, machine learning (ML) is an effective solution owing to low cost, high speed, and low processing difficulty. However, the feasibility and applicability of ML modeling for crop WFs at multiple spatial scales remain unverified. Here, we quantitatively evaluated the applicability, optimal model parameters and importance of the feature variables of four ML methods, including random forest (RF), light gradient boosting machine (LightGBM), generalized additive model (GAM), and artificial neural network (ANN), for constructing WF models for crop production at site and provincial scales. The models were tested through the case study for major grain crops (wheat, maize, soybean, and rice) in China over 2000–2019, distinguishing between irrigated and rainfed crops and between different irrigation methods. The findings show that the WF simulation accuracy for the irrigated scenario was higher than that for the rainfed scenario. The type of feasible ML model varied with crop types, variable combinations, water supply scenarios, and spatial scales. At the site scale, the RF model was generally suitable for simulating WF of maize, soybean, and rice, with R 2 values ranging from 0.7 to 0.78 under different irrigation scenarios; the LightGBM was more suitable for simulating wheat with R 2 values ranging from 0.76 to 0.82. The LightGBM was deemed to be the best model at the provincial scale, with some regions presenting a 50% lower root mean square error in the crop WF simulation than in RF. Variables related to spatio-temporal class feature were found to be the most important in both the RF and LightGBM simulations. This study provides a global methodological reference for different countries or regions to establish methods for rapid WF assessments in crop production with cross-disciplinary researchers or macro policymakers as the target users.
geosciences, multidisciplinary,water resources,engineering, civil