Statistical and machine learning approaches to study weather-disease-mustard yield relationship under varying environmental conditions

Abstract

AbstractThe sowing window of mustard crop under Indian conditions, often varies from place to place and from year to year, creating a diverse set of environmental conditions available for crop growth and development. The present study examines the use of statistical and machine learning approaches for mustard yield prediction at eight sowing dates, using long-term (2006–2021) weather and disease data collected from the experimental fields of GB Pant University of Agriculture & Technology, Pantnagar, India. Descriptive statistics suggest that there is a drastic reduction in mustard yield when sowing is delayed after October 15. Cross comparison of models suggested that ANN followed by LASSO can be used for mustard yield prediction satisfactorily at most of the sowing dates. Among different modelling approaches, ANN approach resulted in a higher coefficient of determination (R2 ≈ 1), with a lower normalised root mean square error (nRMSE) during calibration (ranging between 0.05 and 2.08), as well as validation (ranging between 0.02 and 4.57) for different sowing dates, over other approaches. The Pearson correlation coefficients (r) were also determined for the observed and predicted yield. These values also showed ANN as the best performing model with correlation values ranging from 0.63 to 1.00, followed by LASSO (0.47–0.92), ENET (0.46–0.92), and Ridge (0.38–0.91). Small values ofrcorrespond to the late sowing dates DOS 6 and onwards. Machine learning approach ANN outperformed the shrinkage regression methods for most of the sowing dates with its lower error values and higher correlation coefficient values. From the above study it was concluded that machine learning approaches using weather indices and disease severity as the predictor can be effectively used for precise yield prediction of mustard crop under different environmental conditions of north India.