Skip to main content

Is High Resolution Imagery Sufficient to Solely Explain Crop Yields?

corn cob


Remote sensing has provided a variety of methods to monitor crop growth, condition, and predict yields over the last several decades. One limiting factor, however, has been the spatial resolution of the imagery, the coarseness of which has limited effectiveness at smaller- and inner-field scales. Recent introductions of very high resolution imagery (from 10 - <1 m) combined with near daily observational capabilities has the potential to eliminate these limitations and open up new possibilities for crop yield monitoring. 


A recent study co-authored by NASA Harvest Partners Dr. Sergii Skakun, Dr. David Johnson, Dr. Eric Vermote, Dr. Jean-Claude Roger, and Dr. Belen Franch in collaboration with Dr. Natacha Kalecinski [UMD College Park] and Dr. Meredith Brown [UMD College Park] recently looked into just how useful these new imagery sources are. Using yield data collected from 30 corn and soybean fields in the state of Iowa, the authors explored the capability of four high resolution satellite sensors to explain the yield variability in the selected sites.


Distribution of corn and soybean fields in Iowa, USA.


The study investigated imagery from the Worldview-3 [1.25 m], Planet/Dove-Classic [3.25 m] sensors, as well as harmonized Landsat 8/Sentinel-2 (HLS) imagery [30 m]. The authors then extracted spectral data from this imagery, specifically surface reflectance and 8 different vegetation indices. Yield reference data was collected using combine harvesters equipped with GPS capable of recording dry volumetric yields approximately every 3 m along its track. 


Example of yield distribution over one of the study's field sites.


The team first explored the effect that spatial resolution has on capturing yield variability at field-scale. To do this they rescaled the 3 m yield reference data to 10 m, 20 m, and 30 m by averaging point yield data that fell within each coarser pixel. They then determined the Relative Efficiency, or how closely each rescaled map matched to the original 3 m map, by calculating the ratio between coincident yield values in the rescaled and original maps. This analysis found that only the original 3 m data was able to fully explain yield variability within each field. The 10 m, 20 m, and 30 m maps were able to only explain 86%, 72%, and 59% of the yield variability respectively. This suggests that as the ability to predict yield decreases, the coarseness of studied imagery increases. This would indicate that finer-scale resolution is necessary to achieve the most accurate crop yield estimates.


Dependence of the relative efficiency, the metric showing how well one can capture yield variability, on spatial resolution. These results were obtained by simulating yields from the harvester machinery at various spatial resolutions.


The authors also explored how the spectral and temporal resolutions of each studied imagery are able to explain yield within each field site. To do this, they matched the original 3 m yield reference data with corresponding imagery from each sensor. A linear regression model was run for each cloud-free image (and derived vegetation indices) and the corresponding yield data, producing normal and adjusted coefficients of determination (R2 and R2adj) and root mean square error (RMSE) values.


The analysis found that vegetation indices were overall unable to explain yield variability within both corn and soybean fields. While vegetation indices based on the green and red-edge bands were more successful (37% and 44% respectively) than red band based indices (19%), all indices were found to be inadequate. Surface reflectance, however, was shown to be more capable in its explanatory power. In accordance with the indices results, green, red-edge, and near-infrared surface reflectance bands were found to be the bands of greatest importance, however the best linear regression models were created by utilizing all available spectral bands.


Overall, the authors observed mixed performance of the satellite-derived models. While some modeled results showed very high R2 values of 0.88 (HLS) and 0.77 (Planet) for some fields; other fields had much lower R2 values of 0.21 (HLS) and 0.09 (Planet). One potential explanation for this gulf of success is the varying amounts of vegetation within each field. The authors note that fields with the highest yields (e.g. highest levels of vegetation) had lower R2 values, likely caused by a vegetation density-induced saturation of the reflectance values. Regardless of reason, this wide range of capability suggests that high resolution remotely sensed imagery is not sufficient to solely explain yield variability. The paper ends by noting that additional biophysical data, such as soil moisture and evapotranspiration rates, is needed to fully explain crop yields. 


The full paper can be read here.


News Date
Mar 25, 2021
Sergii Skakun, Keelin Haynes