Exploring the Speculative Potential of Modeling NDVI Curves for Creating End-of-Year Yield Estimates

NASS creates crop yield forecasts throughout the growing season, using qualitative survey data and objective biophysical data regarding certain major commodity crops. In response to interest in improving timeliness and spatial detail, as well as concerns about costs and response rates, NASS researchers have begun studying the use of remote sensing techniques to create forecast estimates for crop yield. Using the MODIS Normalized Difference Vegetation Index (NDVI) 8-day composite and crop masks, NASS has standardized a remote sensing process for creating forecast indications for corn and soybeans. Yield forecasting with NDVI depends on knowing a crop's NDVI signature during critical dates. However, it is difficult to forecast either the peak NDVI for a growing season, or the final shape of the NDVI curve. The relationship between NDVI and yield is nonlinear, which adds to uncertainty.

The objective of this research was to explore the speculative potential for modeling the full-season and end-of-season NDVI curve for the corn crop in the United States. The study made use of MODIS NDVI 8-day composite data spanning 18 years, aggregated at the national and state levels. Modeled curves and actual end-of-year observed curves were compared retrospectively. Different curve modeling methods were contrasted, and transforms to the data, such as detrending and mathematical transforms, were explored.

The paper presents 1) the geospatial data used including the time series MODIS imagery and the crop masks which define the pixels of interest, 2) the crop yield forecasting methodology for creating in-season forecast indications, and 3) the comparative analysis of NDVI curve modeling, including a review of the performance of different techniques. The results showed the impact of applying various transforms and modeling techniques to create in-season NDVI curves for use of yield forecasting. A measure of the speculative potential for these modeled curves is provided, as well as a comparative analysis of the impact of detrending NDVI and yield numbers across years when creating in-season forecasts.