Dynamic prediction method for cotton yield based on COSIM model in Xinjiang

Abstract: Xinjiang is the largest cotton producing area in China accounting for more than 50% of the total cotton production in China. So the accuracy of the prediction of cotton production in Xinjiang is particularly important. Based on calibration and validation of cotton growth model COSIM, in this paper, we used a dynamic prediction model for cotton yield forecast and focused on solving the problem of the unknown climatic data substitution during the prediction period. In the process of prediction, the model read the climatic data day by day. For predicting the growth, development and yield of cotton by the dynamic prediction model, in this study, we substituted the measured climatic data in the recent 50, 30, 20, 10, and 5 years for the unknown climatic data from forecasting day to harvest day, respectively. Meanwhile, the climatic data measured in the year was input into the model before forecasting day. In this way, the cotton yield and development could be predicted day by day. To test the reliability of the method, an experiment with 5 different sowing date (April 10th, April 20th, April 30th, May 10th, May 20th) was designed in 2011 at Wusu, Xinjiang (44°43′ N，84°67′ E). Each treatment was replicated 3 times. The cotton was harvested on September 10th, September 15th, September 21th, September 29th and October 5th, respectively. During the experiment, the growing stage of the cotton was recorded. The leaf area and biomass were determined. These parameter values were input into the COSIM model for cotton lint yield prediction. The model reliability was evaluated by comparing the simulated and measured values of lint yield and growing stages. For the simulation, the climatic data measured in 2011 was used. The results showed that the root mean square error (RMSE) of the cotton growing from emergence to flowering stage was 2.2-5.9 d. The determination coefficient was 0.99. For the lint yields simulations, the RMSE was 165.9 kg/hm2. It indicated that the model was reliable in simulating cotton development and lint yield. Based on experimental results of treatment 1 (sowing date was April 20th), we selected the best substitution one for the unknown climatic data from the 5 schemes (climatic data of the recent 50, 30, 20, 10, and 5 years) and then validated by the results from the other treatments. The results showed that the for the randomly selected 7 predicting time (April 1st, May 1st, June 1st, July 1st, August 1st, September 1st, October 1st), the standard deviation of the measured and predicted lint yield of the 5 schemes from 50 to 5 years' climatic data was 171, 123, 82, 86 and 106 kg/hm2, respectively. The predicting accuracy was above 87% compared with the measured values and above 83% compared with the simulated values for the lint yields. Among them, the accuracy in the predicting time after the sowing date was above 93%. Based on the predicting accuracy and the standard deviation, the best scheme was the 10 years' climatic data substation scheme. The validation of the best scheme using the results from the other treatments showed that predicting accuracy could reach 81.3%-99.6%, indicating the reliability of the best scheme for cotton lint yield prediction. Compared with a single station forecasting, the regional forecasting of cotton yield is more important to national macro-control. In a large region, cotton is not sowing on the same day but during a time period. Therefore, in predicting the regional cotton yield, the effect of sowing time should be taken into consideration. As a case, this study only does the forecast once a month. In practice, the daily dynamic forecast would be realized.