The role of topography, soil, and remotely sensed vegetation condition towards predicting crop yield

Authors

Trenton E. Franz, University of Nebraska-LincolnFollow
Sayli Pokal, University of Nebraska-LincolnFollow
Justin P. Gibson, CropMetrics, North Bend, NE
Yuzhen Zhou, University of Nebraska - LincolnFollow
Hamed Gholizadeh, University of Nebraska-LincolnFollow
Fatima Amor Tenorio, University of Nebraska-LincolnFollow
Daran Rudnick, University of Nebraska-LincolnFollow
Derek M. Heeren, University of Nebraska-LincolnFollow
Matthew F. McCabe, King Abdullah University of Science and TechnologyFollow
Matteo Ziliani, King Abdullah University of Science and TechnologyFollow
Zhenong Jin, University of Minnesota - Twin CitiesFollow
Kaiyu Guan, University of Illinois Urbana-ChampaignFollow
Ming Pan, Princeton UniversityFollow
John Gates, CropMetrics, North Bend, NEFollow
Brian Wardlow, University of Nebraska - LincolnFollow

Date of this Version

2020

Citation

Published in Field Crops Research 252 (2020) 107788

DOI: 10.1016/j.fcr.2020.107788

Comments

Copyright © 2020 Elsevier B.V. Used by permission.

Abstract

Foreknowledge of the spatiotemporal drivers of crop yield would provide a valuable source of information to optimize on-farm inputs and maximize profitability. In recent years, an abundance of spatial data providing information on soils, topography, and vegetation condition have become available from both proximal and remote sensing platforms. Given the wide range of data costs (between USD $0−50/ha), it is important to understand where often limited financial resources should be directed to optimize field production. Two key questions arise. First, will these data actually aid in better fine-resolution yield prediction to help optimize crop management and farm economics? Second, what level of priority should stakeholders commit to in order to obtain these data? Before fully addressing these questions a remaining challenge is the complex nature of spatiotemporal yield variation. Here, a methodological framework is presented to separate the spatial and temporal components of crop yield variation at the subfield level. The framework can also be used to quantify the benefits of different data types on the predicted crop yield as well to better understand the connection of that data to underlying mechanisms controlling yield. Here, fine-resolution (10 m) datasets were assembled for eight 64 ha field sites, spanning a range of climatic, topographic, and soil conditions across Nebraska. Using Empirical Orthogonal Function (EOF) analysis, we found the first axis of variation contained 60–85 % of the explained variance from any particular field, thus greatly reducing the dimensionality of the problem. Using Multiple Linear Regression (MLR) and Random Forest (RF) approaches, we quantified that location within the field had the largest relative importance for modeling crop yield patterns. Secondary factors included a combination of vegetation condition, soil water content, and topography. With respect to predicting spatiotemporal crop yield patterns, we found the RF approach (prediction RMSE of 0.2−0.4 Mg/ha for maize) was superior to MLR (0.3−0.8 Mg/ha). While not directly comparable to MLR and RF the EOF approach had relatively low error (0.5–1.7 Mg/ha) and is intriguing as it requires few calibration parameters (2–6 used here) and utilizes the climate-based aridity index, allowing for pragmatic long-term predictions of subfield crop yield.