In-Season Probabilistic Crop Yield Forecasting: Integrating Agro-Climate, Remote Sensing, and Phenology Data

• Activity Number: 251
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Nathaniel Newlands*+ and David Sebastian Zamar
• Companies: Agriculture and Agri-Food Canada and Agri-Environmental Services Branch, Agriculture and Agri-Food Canada
• Keywords: Agriculture ; Bayesian ; Climate ; Forecasting ; Risk ; Uncertainty

Abstract

Statistical models help to provide decision-makers with an improved ability to spatially identify and assess, with enhanced foresight, potential risks and vulnerability of natural resources to climate variability and extremes. They also enable integration of diverse geospatial information together with its uncertainty for operational real-world application. We showcase a Bayesian method for sequential forecasting of the yield of major crops grown across the Canadian Prairies, Western Canada. This method incorporates robust least angle regression followed by robust cross validation for variable-selection, Markov chain Monte Carlo (MCMC) sampling for added spatial correlation support, and forms a joint probability distribution using the random forests algorithm for non-parametric modeling of future observable variables. We explore the relative improvement of candidate agro-climate, remote-sensing, and phenology indices on the overall accuracy of in-season forecasts (updated on a monthly basis) at two different spatial resolutions. Preliminary findings from cross-validation on spring wheat yield indicate a gain of 10% when involving net-difference vegetation index (NDVI) as a spatial index of crop yield potential and net model accuracy of 89%.