Figure 2

Training, testing, and validation of neural network. Neural network was trained using total rainfall from November to January (RainF_NovDecJan), an average of maximum temperature for the October and November months (MaxT_JanFeb) along with information about variety (Variety) and soil type (Soil_number) as input, and DRR disease incidence data (Supplementary File S1) as output. The entire dataset was divided into 70% training set and 30% testing set. Neural network (NN) was trained in R using neuralnet algorithm with backpropagation method. The best fit in NN training was obtained with one hidden layer having three nodes. Network topology for the training and linear regression between actual and predicted for the training set are shown in (a) and (b), respectively. Numbers in connecting lines from input layer to hidden layer and from hidden layer to output layer represent weights used in the model and number connecting blue circles are biases. Linear activation function was used for the network training. Validation of the trained neural network was performed with DRR disease incidence from the current (2019–2020) field experiment study and with DRR incidence data of location-2 from Sinha et al. (2019) research article. Validation data set included DRR disease incidence data from severe combined stress treatment plots. Input and output data used for validation is in Supplementary File S1. A comparison between actual (red) and predicted DRR disease incidence (green) is shown as line graph (d). RMSE, r and R² is showing root mean square error, correlation coefficient and coefficient of determination, respectively. DRR incidence was categorized into High DRR (disease incidence > 30%) and Low DRR (disease incidence < 30%) and confusion matrix was created with actual and predicted DRR incidence for the validation set. The prediction accuracy was calculated.