Location: Adaptive Cropping Systems LaboratoryTitle: Mapping sub-field maize yields in Nebraska, USA by combining remote sensing imagery, canopy simulation models, and machine learning
|JEFFRIES, GRAHAM - Tufts University|
|GRIFFIN, TIMOTHY - Tufts University|
|NAUMOVA, E - Tufts University|
|KOCH, M - Tufts University|
|WARDLOW, B - Tufts University|
Submitted to: Precision Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/1/2019
Publication Date: 10/21/2019
Citation: Jeffries, G., Griffin, T., Fleisher, D.H., Naumova, E.N., Koch, M., Wardlow, B.D. 2019. Mapping sub-field maize yields in Nebraska, USA by combining remote sensing imagery, canopy simulation models, and machine learning. Precision Agriculture. https://doi.org/10.1007/s11119-019-09689-z.
Interpretive Summary: Our country’s food supply needs to be increased and made more secure to support a growing population. Precision agriculture is an approach that helps farmers obtain more consistent and higher yields. But better methods to link satellite maps of the field with predicted yields are needed to improve this type of practice. This study tested the suitability of different mathematical algorithms to predict field yields using several different satellite pictures from fields in Nebraska. The results identified the specific algorithms and imaging technology that gave the best yield predictions. In the future, these discoveries will be useful as part of image-based decision support systems to help reduce farm resource use and improve the sustainability of on-farm production.
Technical Abstract: Crop yield maps are valuable for many applications in precision agriculture, but are typically only available for sites and years where yields were recorded specialized equipment. A method for mapping sub-field crop yields from remote sensing imagery could lower the cost and increase the availability of crop yield maps. We tested an algorithm for creating maize (Zea mays L.) yield maps (10 m) without in situ observations. Crop simulation outputs were used to fit statistical models which then predicted yield using remote sensing imagery (Landsat and airborne hyperspectral). The method was validated using harvester yield monitor records for 21 site-years for irrigated and rainfed fields in eastern Nebraska, USA. We tested alternative specifications of the prediction algorithm, and the preferred method explained 67.7 percent of the variation in pixel-level yields across all fields (RMSE = 0.99; NRMSE = 0.08). Predictive performance was typically higher for rainfed sites and in locations with more than five observation years. Gradient boosted forest models outperformed linear regression models for predicting mean yields, but not for capturing within-field variation. Significant but correctable proportional bias in predictions was detected. Scalable crop yield mapping methods with remote sensing imagery show promise for precision agriculture applications.