Prediction of agricultural drought for the Canadian prairies using climatic and satelllite data
Wheat export is a significant component of the Canadian economy. In normal (nondrought) years, the export is as high as 30 million tonnes, but it is reduced to about 20 million tomes in drought years. This significant reduction in exports not only reduces direct profits but may also upset export targets and prices that are set in advance, if droughts are not accurately predicted. In this thesis, prediction of agricultural drought is attempted from both long-term and short-term perspectives. The long-term prediction refers to predicting wheat yield (production per unit area) prior to wheat planting; and, under the short-term prediction, wheat yield is estimated around harvesttime. Predictive analysis was performed on five crop districts of Saskatchewan (1b, 3bn, 4b, 6a, and 9a) using climate data (monthly and daily tempera ure and precipitation) from rune weather stations. In addition, Normalized Difference Vegetation Index values generated from NOAA (National Oceanic and Atmospheric Administration)/AVERR (Advanced Very High Radiometric Resolution) satellite data were used. The long-term prediction was made by fitting various time series techniques (trend, moving average, exponential smoothing, and autoregressive integrated moving average) to the yield series in a district. The technique providing minimum prediction-error was selected. The short-term prediction was made in both qualitative and quantitative forms. The qualitative prediction was attempted using the error correction procedure of pattern recognition. The. quantitative prediction involved modification of the computer program currently being used by the Canadian Wheat Board (CWB) to estimate wheat yield. The CWB program employs only monthly and precipitation and determines a drought index for a weather station. A hybrid model that employs daily climate data and a NDVI-based variable was developed. Among Various NDVI-based variables, the average NDVI during the entire growing period was found to be the best predictor of yield in the case of district 3bn. For the remaining four districts, the average NDVI during the heading stage was the most reliable predictor. The commencement and termination of the heading stage were determined using a biometeorological time scale model that required planting dates, daily maximum and minimum temperatures and the photoperiod. When evaluated, the hybrid model was found to have significantly higher predictive capability than the model currently in use; the values of r2 were 0.79, 0.96, 0.83, 0.95, and 0.39 (in the case of hybrid model) as opposed to 0.20, 0.71, 0.57, 0.58, and 0.00 (in the case of the current model) for the districts 1b, 3bn, 4b, 6a, and 9a, respectively.