HR: 1340h
AN: B13D-1526    [Abstracts]
TI: County-Level Crop Yield Prediction Using Remote Sensing Data
AU: * Wagstaff, K L
EM: kiri.wagstaff@jpl.nasa.gov
AF: Jet Propulsion Laboratory, California Institute of Technology, Mail Stop 306-463, 4800 Oak Grove Drive, Pasadena, CA 91109, United States
AU: Roper, A
EM: alexr@ugcs.caltech.edu
AF: California Institute of Technology, 1200 E. California Blvd., Pasadena, CA 91125, United States
AU: Lane, T
EM: terran@cs.unm.edu
AF: University of New Mexico, MSC01 1130, 1 University of New Mexico, Albuquerque, NM 87131, United States
AB: Early estimates of crop yield, particularly at a fine scale, can inform precision agriculture efforts. The USDA National Agricultural Statistics Service (NASS) currently provides estimates of yield on a monthly basis for each state. These estimates are based on phone interviews with farmers and in-situ examination of randomly selected plots. We seek to provide predictions at a much higher spatial resolution, on a more frequent basis, using remote sensing observations. We use publicly available data from the MODIS (Moderate Resolution Imaging Spectroradiometer) instruments on the Aqua and Terra spacecraft. These observations have a spatial resolution of 250 m and consist of two spectral bands (red and infra-red) with a repeat period of 8 days. As part of the HARVIST (Heterogeneous Agricultural Research Via Interactive, Scalable Technology) project, we have created statistical crop yield models using historical MODIS data combined with the per-county yield reported by the USDA at the end of the growing season. In our approach, we analyze 100 randomly selected historical pixels from each county to generate a yield prediction for the county as a whole. We construct a time series for each pixel that consists of its NDVI (Normalized Difference Vegetation Index) value observed during each 8-day time period to date. We then cluster all pixels together to identify groups of distinct elements (different crops, bodies of water, urban areas, desert, etc.) and create a regression model for each one. For each crop of interest, the model that best predicts that crop's historical yield is selected. These models can then be applied to data from subsequent years to generate predictions for the future. We applied this approach to data from California and Kansas for corn and wheat. We found that, in general, the yield prediction error decreased as the harvest time approached. In California, distinctly different models were selected to predict corn and wheat, permitting specialization for each crop type. The best models from 2001 predicted yield for 2002 with a 10% (corn) and 23% (wheat) relative error three months before harvest. In Kansas, the 2001 models for corn and wheat were not well distinguished, providing good predictions for wheat (19% error three months before harvest) but poor predictions for corn (55% error three months before harvest). In post-analysis, we found that the 2001 pixel NDVI time series for Kansas are much more homogeneous than those for California, making it difficult to select crop-specific models. We are currently working on incorporating historical data from additional years, which will provide more diversity and potentially better predictions. We are also in the process of applying this technique to additional crops.
DE: 0402 Agricultural systems
DE: 0430 Computational methods and data processing
DE: 0480 Remote sensing
SC: Biogeosciences [B]
MN: 2007 Fall Meeting