Qualitative analysis of random forests for evaporation prediction in Indian Regions

Abstract

The performance of logistic regression, discriminant analysis, and random forest has been compared for the prediction of evaporation of different regions of India during 2019 at ICAR-IARI, New Delhi . The present experiment was performed at Raipur (Chhattisgarh), Karnal (Haryana), Pattambi (Kerala) and Anantpur (Andhra Pradesh). Evaporation and other weather parameters are collected from the year 1985-2012, 1973-2005, 1991-2005 and 1958-2010 respectively. The performance of the techniques is compared using classification, misclassification, and sensitivity of the model along with the Receiver Operating Characteristics (ROC) curve and Area Under Curve (AUC) value. The combinations of variables as independent variables are used in two sets. In the first set, maximum & minimum temperature, relative humidity morning & evening, wind speed, rainfall, and bright sunshine hours are used. In the second set mean temperature, mean relative humidity, bright sunshine hours, and wind speed is used to see the effect on evaporation. It is found that more accuracy is obtained using the second set as predictors. The model validation accuracy is checked via running developed model on out of sample data, i.e. testing data (last three years). The study demonstrates that the random forest approach predict evaporation in a much better way than logistic regression, discriminant analysis. The random forest model can provide timely information for the decision-makers to make crucial decisions impacting due to evaporation conditions in India.