Neural network models to predict the maturity of spring wheat in western Canada

Abstract

The Canadian Wheat Board (CWB) and the railways require accurate predictions of harvest dates to facilitate the movement of grain from the Canadian Prairies. Our objective was to develop neural network (NN) models to predict, 4–8 wk in advance, the maturity date of spring wheat (Triticum aestivum L.) across the Prairies with an accuracy of 0.5 wk. We used soil and climatic zone information, cultivar, CWB wheat class, seeding date, and the first 9 wk of weather after seeding to predict days to maturity in the standing crop from seeding date. Our first Prairie-wide model (MD1), derived from data at 76 locations over 1970 –1995, performed well (R2 = 0.82–0.83, average absolute error = 3.0–3.1 d), but used solar radiation inputs, which may not always be available. Therefore, a second model (MD2) was developed (R2 = 0.76–0.80, average absolute error = 3.3–3.4 d) without solar radiation inputs. A third model (MD3) was developed (R2 = 0.72, average absolute error = 3.8 d) specifically for Canada Western Red Spring (CWRS) wheat grown at dryland locations. The MD2 and MD3 models were tested on more recent (1996–1998) data, but their performance was poor (R2 = 0.26–0.33; average absolute errors = 5.2–6.9 d) because of differences in the weather data. Thus, a fourth model (MD4) was developed (R2 = 0.85–0.90; average absolute error = 2.7–2.8 d) using 1989–1998 crop data from 19 locations combined with more recent Environment Canada “Real Time” Weather (ECRTW) data. Considering the variation in soil types, cultivars, seeding dates, and weather conditions across the Prairies, the NN models gave accurate predictions of wheat maturity. Key words: Neural networks, model, wheat, maturity, weather, Canadian prairies