Subfield maize yield prediction improves when in-season crop water deficit is included in remote sensing imagery-based models

Abstract

In-season prediction of crop yield is a topic of research studied by several scientists using different methods. Seasonal forecasts provide critical insights to different stakeholders who use the information for strategic and tactical decisions. In this study, we propose a novel scalable method to forecast in season subfield crop yield through a machine learning model based on remotely sensed imagery and data from a process-based crop model on a cumulative crop drought index (CDI) designed to capture the impact of in-season crop water deficit on crops. To evaluate the performance of our proposed model, we used 352 growers' fields of different sizes across the states of Michigan, Indiana, Iowa, and Illinois, with 2520 respective yield maps generated by combine harvesters equipped with precise high-resolution yield monitor sensor, over multiple years (from 2006 up to 2019). We obtained high resolution digital elevation model, climate, and soil data to execute the SALUS model, a process-based crop model, to calculate the CDI for each field used in the study. We used Landsat Analysis Ready Dataset (ARD) products generated by USGS as image source to calculate the green chlorophyll vegetation index (GCVI). We found that the inclusion of the CDI in remote sensing-based random forest models substantially improved in-season subfield corn yield prediction. The addition of the CDI in the yield prediction model showed that the greatest improvements in predictions were observed in the driest year (2012) in our case study. The proposed approach also showed that the subfield spatial variations of corn yield are better captured with the inclusion of CDI for most fields. The earliest prediction in the growing season with GCVI and CDI together outperformed the latest prediction with GCVI alone, highlighting the potential of CDI for predicting spatial variability of maize yield around grain filling period, which is on average close to two months before typical crop harvest in the US Midwest.

Plant phenotyping relevance

リモートセンシング画像と機械学習を用いて、圃場内のトウモロコシ収量という植物形質を予測する手法を提案・評価しており、形質推定法が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.