Use of crop simulation model and machine learning techniques to predict rice yield in Udham Singh Nagar district of Uttarakhand
Loading...
Date
2022-09
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
G.B. Pant University of Agriculture and Technology, Pantnagar, District Udham Singh Nagar, Uttarakhand. PIN - 263145
Abstract
The present study was conducted at C6 Block in the Norman E. Borlaug Crop Research Centre of G.B. Pant University of Agriculture and Technology, Pantnagar to predict rice yield during kharif season 2021. Calibration and validation were carried out using the CERES-Rice (version 4.8) model. The experiment was laid out with two prominent cultivars (HKR-47 and Pusa Basmati 1509), two transplanting dates (07th July and 17th July), and two fertilizers doses (F1= 150:60:40 and F2= 180:90:60) to calibrate the CERES-Rice model, so that model could be used for district level rice yield prediction. SMLR and ANN techniques were used to predict rice yield at the district level. Using genetic coefficients, the model was calibrated with the experimental dataset from 2021. The effectiveness of the genetic coefficient was verified for variety HKR-47 using the experimental dataset of 2018 and for variety Pusa Basmati 1509 using the experimental dataset of 2020. Two different approaches i.e., statistical and machine learning were used to predict rice yield at the district level. Experimental analysis suggested that HKR-47 and Pusa Basmati 1509 both varieties performed better when transplanted on 17th July as compared to 07th July. The performance of the model CERES-Rice was satisfactory for two transplanting dates, two fertilizer doses, and two varieties during the period of study for almost all crop characters. In case of calibration the % RMSE for observed and simulated data for emergence days panicle initiation, anthesis, physiological maturity, grain yield, Leaf Area Index and harvest index was found to be 25, 9.59, 3.46, 7.56, 17.77, 9.67, and 8.69, respectively for HKR-47. While in the case of Pusa Basmati 1509 the % RMSE for emergence days, panicle initiation, anthesis, physiological maturity, grain yield, Leaf Area Index, and harvest index was found to be 20, 6, 16.77, 2.30, 5.74, 15.17, and 7.59, respectively. In the case of validating the data for HKR-47, the % RMSE for observed and simulated data for anthesis days, physiological maturity, and grain yield was found to be 12.32, 10.84, and 8.82, respectively. While in the case of Pusa Basmati 1509 % RMSE for grain yield and harvest index was found to be 8.09 and 8.02, respectively. Stepwise multiple regression equations were developed for yield prediction by using 19 years of weather and observed yield data. A total of four models were developed among them model 4 gave the best results (R2 = 0.91 and RMSE = 3.14%). Yield prediction models for rice crop have been developed using stepwise multiple linear regression (SMLR) and the ANN approach. For calibration of the model 15 years (2001-2015) of weather and yield data and for validation 4 years (2016-2019) of weather and yield data were used. The SMLR model performed well with an R2, nRMSE, and MBE value of 0.89, 0.177 %, and -0.149 t ha-1 respectively for calibration and 0.545, 0.66% and -0.22 t ha-1 for validation. In the ANN method, the value of R2, nRMSE and MBE is 1, 0.012 % and 6.136383e-05 t ha-1 respectively for calibration and 0.99, 0.0092 % and -0.0011 t ha-1, respectively for validation. In between statistical and machine learning approaches, the machine learning approach performs better for rice yield prediction with the coefficient of correlation reaching (R2) almost 1.