Fertilization and soil management machine learning based sustainable agronomic prescriptions for durum wheat in Italy

The agricultural sector is challenged to produce more food with less available land. This challenge is compounded by increasing production input costs, which are further impacted by geopolitical changes and the adverse effects of climate change. Finding sustainable solutions to address these challenges is crucial. Machine learning can be a valuable tool in agriculture to optimize soil, nutrient, and crop management, helping to maximize food production on limited land and mitigate the effects of climate change. Machine learning models hold great promise in agriculture; however, their success is contingent on ease of use, accessibility for farmers, and the perceived utility of such models by the agricultural community. This work aimed to develop a meta-machine learning that allows the simulation of different combinations of soil and nitrogen management of durum wheat in Italy. This model aimed to assist decision makers and stakeholders in determining the most effective agronomic management based on attainable crop yield and potential income margins derived from the use of agronomic inputs. The meta-machine learning model was developed and tested across four sites located in the Italian regions of Marche and Basilicata. These sites featured different experimental designs, and durum wheat was grown for several years. The study involved the comparison of a total of eleven different nitrogen levels. The meta-machine learning was composed by linked classification and regression machine learning models. These components were trained using a multi-data source approach, which included data from remote sensing, crop phenology, soil chemical properties, weather data, soil management, and nitrogen levels, to predict durum wheat yield. The classification task employed a Random Forest model with an accuracy of 0.98, kappa of 0.96 and recall of 0.98 for predicting the crop phenology while the yield prediction task was performed by a Neural Network model with an R squared of 0.90, Root Mean Square Error of 0.59, Mean Absolute Error of 0.45 and Mean Absolute Percentage Error of 0.17. The variable importance analysis was conducted to identify the most important covariates that allow to improve the model’s accuracy. This analysis revealed that temperature, precipitation, NDVI (Normalized Difference Vegetation Index), and nitrogen input are the most important factors. The meta-model was used to run simulations of 30 different combinations of soil management and fertilization levels. These simulations aimed to identify the most effective agronomic strategy for each of the farm sites. The no tillage management have been found to result in increased grain yield. The Marginal Fertilizer Yield Index was used to determine the optimal nitrogen application for the crop. The potential transferability to field conditions of the model is facilitated by its utilization of publicly available spatial datasets, which can enable the broader application of the meta-model.