报告摘要：In recent years,the use of machine learning algorithms for solving regression and classification problems has increased dramatically,including in the Earth and environmental sciences. Machine learning has become a dominant approach for predicting soil and agronomic properties from a wide range of explanatory variables. In this presentation, I will use the example of predicting maize yield in Ghana based on climate, soil, crop, management, and fertilizer data. A random forest regression model trained on more than 3,000 paired observations explained over 80% of the observed variability in maize yield. The trained model can also be used to evaluate how fertilizer application rate influences yield by generating so-called fertilizer–yield response curves. From these curves, we can determine the optimal fertilizer rate by identifying the rate that maximizes profit. I will explain how this can be done, but I will also raise the question of whether this approach is scientifically sound and valid, since it effectively uses an empirical, correlation-based model as if it were causal. I will not provide answers but aim to initiate a discussion. The key issue appears to revolve around achieving unconfoundedness among explanatory variables—a condition typically ensured through well-balanced experimental designs, but which may also be attainable through thoughtful combinations of observational data.

报告人：Gerard Heuvelink 教授 荷兰瓦赫宁根大学

报告人简介：

Gerard Heuvelink是ISRIC–World Soil Information的高级研究员、荷兰瓦赫宁根大学土壤地理与景观研究组的计量土壤学与数字土壤制图特聘教授。他的博士研究聚焦于地理信息系统空间建模中的误差传播，标志着其在空间不确定性分析领域科研生涯的开端，主要研究方向为土壤应用。2014年，他因推动计量土壤学发展而获得国际土壤科学联合会授予的理查德·韦伯斯特奖章；2020年，获得国际空间精度研究协会授予的彼得·伯勒奖章。他在地统计学、空间不确定性分析、计量土壤学、数字土壤制图以及农学与土壤科学机器学习领域发表了200余篇SCI论文，被科睿唯安评选为“全球高被引科学家”，其论文在Web of Science上被引用超过17,500次（H指数为55）。Gerard Heuvelink现任《European Journal of Soil Science》副主编，并担任《Geoderma》《Spatial Statistics》等五本SCI期刊的编委。多年来，他已指导45余名硕士研究生和25名博士研究生。2011年至2022年，他担任中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室客座教授。

主持人：葛咏 研究员

时  间：2025年10月17日（星期五）10:00 -12:00

地  点：A332会议室

主办单位：地理信息科学与技术全国重点实验室