BTT EAGER: Improving Crop Yield Prediction by Integrating Machine Learning with Process-Based Crop Models

预测作物产量是应对粮食安全新挑战的核心，特别是在全球气候变化的时代。目前，机器学习和作物建模是最常用的产量预测方法之一。该奖项支持基础研究将机器学习和作物模型的优势联合收割机结合起来。机器学习算法将用于预测中间植物性状，然后将其输入作物模型，以预测不同环境和田间管理实践中的谷物产量。EAGER项目的概念和执行都依赖于多个学科的合作，包括高通量表型分析、对象识别、机器学习、优化、计算机模拟和作物建模。如果成功的话，这项研究预计不仅可以提高准确性，还可以提高产量预测模型的可解释性，这将为下游研究和发现提供许多机会。跨学科的努力将提高跨学科的科学和工程教育的影响，同时为所有学生提供一个合作和包容的环境，从事尖端的研究活动。潜在的产量预测是生物学的重大挑战之一：了解表型是如何由基因型，环境，以及它们的相互作用决定的。机器学习算法能够基于基因型信息以合理的精度预测作物表型，但大多数模型具有黑箱结构，其结果难以解释。另一方面，作物模型通过提供作物生长周期不同阶段性状与环境条件之间相互作用的明确解释，提供了对表型变异原因的生物学见解，但性状测量数据的收集和模型系数的校准是劳动密集型的，耗时的，并且成本高。所提出的方法是一个嵌套模型。将训练深度学习算法，以根据基因型和经验测量的性状数据预测叶片出现率。训练数据将从通过采用新型表型分析技术的田间实验获得的植物叶片图像中提取。接下来，所得到的预测性状和环境数据将被输入作物模型以预测产量。如果证明有效，这种方法可以应用于研究其他植物性状，以提高作物产量预测。这个奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的知识价值和更广泛的影响审查标准进行评估的支持。

Predicting crop yield is central to addressing emerging challenges in food security, particularly in an era of global climate change. Currently, machine learning and crop modeling are among the most commonly used approaches for yield prediction. This award supports fundamental research to combine the strengths of machine learning and crop models. Machine learning algorithms will be used to predict intermediate plant traits, which will then be fed into a crop model to predict grain yields across different environment and field management practices. Both conception and execution of this EAGER project depend on collaborations across multiple disciplines, including high-throughput phenotyping, object recognition, machine learning, optimization, computer simulation, and crop modeling. If successful, this research is expected to improve not only accuracy but also interpretability of yield prediction models, which will open numerous opportunities for downstream research and discoveries. The interdisciplinary effort will enhance the impact of science and engineering education across disciplines, while providing a collaborative and inclusive environment for all students to engage in cutting edge research activities.Underlying yield prediction is one of the grand challenges of biology: understanding how phenotype is determined by genotype, environment, and their interactions. Machine learning algorithms are able to predict crop phenotype to reasonable accuracy based on genotype information, but most models have a black box structure and their results are hard to interpret. On the other hand, crop models offer biological insights into causes of phenotypic variation by providing explicit explanations of the interactions between traits and environmental conditions in different phases of the crop growth cycle, but the collection of trait measurement data and calibration of model coefficients are labor intensive, time consuming, and costly. The proposed approach is a nested model. Deep learning algorithms will be trained to predict leaf appearance rate from genotype and empirically measured trait data. Training data will be extracted from images of plant leaves obtained via field experiments that employ novel phenotyping technique. Next, the resulting predicted traits and environment data will be fed into the crop model to predict yield. If proven effective, this approach can be applied to study other plant traits to improve crop yield prediction.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.