Modeling days suitable for fieldwork using machine learning, process-based, and rule-based models

Context: Prediction of days suitable for fieldwork is important for understanding the potential effects of climate change and for selecting machinery systems to improve efficiency in field operations and avoid soil damage. Yet, we lack predictive models to inform decision-making at scale.Objective: We filled this knowledge gap by developing and testing five new workability models.Methods: One model follows soil moisture-based methods (APSIM), one uses simple rain and temperature thresholds, and three follow machine learning techniques (Random Forest, Decision Table, Neural Network). We parameterized the models using USDA survey data from Iowa, USA and evaluated their temporal and spatial prediction over twelve US Corn Belt States and different time periods using multiple statistical indexes and sensitivity analysis. The models operate at a 5-arcminute resolution.Results and conclusions: Results indicated that the simple rule model, the Decision Table model, and the process -based model predicted field workable days with an agreement index of 0.88, 0.86, and 0.84, respectively for the testing datasets (n = 22,671), and hence were deemed sufficient for future use. The selected models are better suited for large timespan evaluations of workability (monthly to annual, normalized root mean square error, nRMSE = 8 to 15%)) than weekly predictions (nRMSE = 21%). The machine learning models tended to cluster their predictions around a mean value and were about 50% less responsive to precipitation than the process -based or rule-based models. We concluded that simple approaches are more robust to be applied at scale than complex approaches with many data input requirements.Significance: The developed models enhance our capacity to predict climate change impacts on workability, a valuable indicator for decision-making and overall sustainability.