Purpose

This research aims to develop a meta-machine learning model to optimize soil and nitrogen management for durum wheat in Italy. It addresses the challenges of increased food production on limited land amidst rising input costs, geopolitical changes, and climate change. The goal is to aid decision-makers in achieving maximum crop yield and income margins through effective agronomic strategies.

Methods

The study developed a meta-machine learning model, integrating classification and regression models, and tested it at four sites in Marche and Basilicata, Italy, over several years. The model incorporated data from remote sensing, crop phenology, soil chemical properties, weather data, soil management, and nitrogen levels. A Random Forest model was used to classify crop phenology, while a Neural Network model predicted yield. Eleven nitrogen levels were compared across these sites.

Results

The Random Forest model achieved an accuracy of 0.98, kappa of 0.96, and recall of 0.98 for predicting crop phenology. The Neural Network model for yield prediction had an R squared of 0.90 and a Root Mean Square Error of 0.59 t ha-1. Key factors identified for model accuracy were temperature, precipitation, NDVI, and nitrogen input. Simulations of 30 soil management and fertilization combinations revealed that no-tillage management increased grain yield. The Marginal Fertilizer Yield Index determined optimal nitrogen application.

Conclusions

The meta-machine learning model accurately predicted durum wheat yield and identified effective agronomic strategies, demonstrating the potential for broader application in field conditions. The model offers a promising approach to sustainable agriculture and climate change mitigation by utilising publicly available spatial datasets.

Graphical abstract

中文翻译：

---

意大利硬粒小麦基于机器学习的施肥和土壤管理可持续农艺配方

 目的

这项研究旨在开发一种元机器学习模型，以优化意大利硬粒小麦的土壤和氮管理。它解决了在投入成本上升、地缘政治变化和气候变化的情况下在有限的土地上增加粮食产量的挑战。目标是帮助决策者通过有效的农艺策略实现作物产量和收入最大化。

 方法

该研究开发了一种元机器学习模型，集成了分类和回归模型，并在意大利马尔凯和巴斯利卡塔的四个地点进行了数年的测试。该模型整合了来自遥感、作物物候、土壤化学特性、天气数据、土壤管理和氮水平的数据。随机森林模型用于对作物物候进行分类，而神经网络模型则用于预测产量。对这些地点的十一个氮水平进行了比较。

 结果

随机森林模型在预测作物物候方面的准确度为 0.98，kappa 为 0.96，召回率为 0.98。用于产量预测的神经网络模型的 R 平方为 0.90，均方根误差为 0.59 t ha-1。影响模型准确性的关键因素是温度、降水、NDVI 和氮输入。对 30 种土壤管理和施肥组合的模拟表明，免耕管理提高了粮食产量。边际肥料产量指数决定了最佳施氮量。

 结论

元机器学习模型准确预测了硬粒小麦产量并确定了有效的农艺策略，展示了在田间条件下更广泛应用的潜力。该模型通过利用公开的空间数据集，为可持续农业和减缓气候变化提供了一种有前途的方法。

 图形概要