Weather based forecasting of sterility mosaic disease in pigeonpea (Cajanu cajan) using machine learning techniques and hybrid models

Abstract

Modelling incidence of sterility mosaic disease (SMD) on pigeonpea [Cajanus cajan (L.) Millsp.] for four locations [S K Nagar (Gujarat), Gulbarga (Karnataka), Rahuri (Maharashtra) and Vamban (Tamil Nadu)] was carried out based on field data sets generated during six kharif seasons [2011-16]. Mean seasonal incidence amongst all locations was on the decline during recent periods (0.5-5.3%) over past decades (9.8-12.8%). Correlation analyses of SMD incidence with weather parameters lagged one and two weeks indicated spatial differences for the variables besides their significance. While Max T (ºC) lagged by one week alone was significantly positive with SMD at Gulbarga (KA), Vamban (TN) had negative significance of rainfall (mm/week) and rainy days. S K Nagar (GJ) and Rahuri (MH) had shown opposite effects of both morning and evening RH (%) of both one and two lagged weeks. Support vector regression (SVR), artificial neural network (ANN) models and their combination with autoregressive integrated moving average (ARIMA) models applied for prediction of SMD incidence across locations revealed performance of hybrid models in general to be better based on the evaluation criteria of root mean square error (RMSE). ARIMASVR> ARIMA-ANN>SVR>ANN was the order of prediction accuracies at S K Nagar (GJ), Gulbarga (KA), and Vamban (TN). At Rahuri (MH), individual models performed better over their hybrids with ARIMA. While application of hybrid model of SVR-ARIMA is applicable under situations of SMD seasonal mean severity exceeding 1%, SVR model proves better for mean seasonal disease incidence in decimal values less than one.

Keywords

• ANN
• ARIMA
• Pigeonpea
• SMD
• SVR
• Weather variables