Advances in regional crop yield estimation over the United States using satellite remote sensing data

Abstract:

Since the dawn of earth observation imagery, particularly from systems like Landsat and the Advanced Very High Resolution Radiometer, there has been an overarching desire to regionally estimate crop production remotely. Research efforts integrating space-based imagery into yield models to achieve this need have indeed paralleled these systems through the years, yet development of a truly useful crop production monitoring system has been arguably mediocre in coming. As a result, relatively few organizations have yet to operationalize the concept, and this is most acute in regions of the globe where there are not even alternative sources of crop production data being collected.

However, the National Agricultural Statistics Service (NASS) has continued to push for this type of data source as a means to complement its long-standing, traditional crop production survey efforts which are financially costly to the government and create undue respondent burden on farmers. Corn and soybeans, the two largest field crops in the United States, have been the focus of satellite-based production monitoring by NASS for the past decade. Data from the Moderate Resolution Imaging Spectroradiometer (MODIS) has been seen as the most pragmatic input source for modeling yields primarily based on its daily revisit capabilities and reasonable ground sample resolution.

The research methods presented here will be broad but provides a summary of what is useful and adoptable with satellite imagery in terms of crop yield estimation. Corn and soybeans will be of particular focus but other major staple crops like wheat and rice will also be presented. NASS will demonstrate that while MODIS provides a slew of vegetation related products, the traditional normalized difference vegetation index (NDVI) is still ideal. Results using land surface temperature products, also generated from MODIS, will also be shown. Beyond the MODIS data itself, NASS research has also focused efforts on understanding a variety of data mining and modeling options and results strongly lean toward solutions of ensemble decision trees like Cubist and Random Forest. Those comparisons of what are seen as best will be also be shown. And finally, important model refinements accounting for temporal and spatial trends have also been considered and results will be presented.