Open Collections

Crop yield forecasting on the Canadian Prairies by satellite data and machine learning methods

University of British Columbia

The production of grain crops plays an important role in the economy of the Canadian Prairies and early reliable crop yield forecasts over large areas would help policy makers and grain marketing agencies in planning for exports. Forecast models developed from satellite data have the potential to provide quantitative and timely information on agricultural crops over large areas. The use of nonlinear modeling techniques from the field of machine learning could improve crop forecasting from the linear models most commonly used today. The Canadian Prairies consist of the provinces of Alberta, Saskatchewan and Manitoba and three of the major crops in this region are barley, canola and spring wheat. The agricultural land on the Canadian Prairies has been divided into Census Agricultural Regions (CAR) by Statistics Canada. A clustering model was applied to the crop yield data to group the CARs for the development of forecast models. The normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) derived from the Moderate Resolution Imaging Spectro-radiometer (MODIS), NDVI derived from the Advanced Very High Resolution Radiometer (AVHRR) and several climate indices were considered as predictors for crop yields. A correlation analysis between crop yield and the time series of each potential predictor was performed to determine which variables showed the most forecasting potential and at what time during the growing season their values were most correlated to crop yield. Various combinations of MODIS-NDVI, MODIS-EVI and NOAA-NDVI were used to forecast the yield of barley, canola and spring wheat. Multiple linear regression as well as nonlinear Bayesian neural networks and model-based recursive partitioning forecast models were developed using the various sets of predictors. The models were trained using a cross-validation method and the forecast results of each model were evaluated by calculating the skill score from the mean absolute error, with 95% confidence intervals for the skill scores calculated using a bootstrap method. The results were compared in an effort to determine the optimal set of predictors and type of forecast model for each crop.

Text

2013-10-17

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/45281

Atmospheric Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/