Comparative evaluation of na ̈ive–bayes and k–nearest neighbor classifiers to improve prediction of short-term precipitation

Abstract

This research carried out a comparative evaluation of K - Nearest Neighbour and Naive - Bayes classifiers to improve prediction of short-term precipitation in Numerical Weather Prediction (NWP) models. Aiming at providing alternatives to imperfections under the current NWP physical process methods namely representation of many points of the earth at high resolution, the dynamic nature of the atmosphere, errors in the initial conditions, inefficiency due to sparsely distributed data points. These short comings require large amounts of computing resources to effectively stimulate and predict weather changes. One of the many unsuccessful interventions is upgrade of hardware components to cope up with the requirements for high computer processing speed, storage capacity and need to continue increasing the resolution at which NWPs run while keeping power consumption with in the reasonable limits. Machine learning KNN and Naive Bayes scalability capabilities enable the NWP model to handle growing amounts of work while keeping computational cost within the reasonable limits. Research studies have demonstrated the capabilities of machine learning models to replace weather and climate models that are based on the physical processes and the basic equations of motion. In this research, input parameters are Maximum and Minimum temperature, Relative humidity, Dewpoint temperature, Dry and Wet bulb temperatures and Rainfall amount six hourly interval observation datasets for 03 (three) years from Uganda National Meteorological Authority. The observation datasets are for Kawanda, Namulonge and Masindi weather stations. These stations are within the same climatic influence of River Mayanja. Measurable metrics being Accuracy, F1 Score and Mathew correlation coefficient (MCC). F1 score and MCC experimental studies offers performance analysis. The data imbalance remedy prone to weather observation datasets was addressed using the Synthetic Minority Over Sampling Techniques (SMOTE) algorithm. Research findings have found the Naive Bayes robust in the prediction of short-term precipitation for both dry and wet season. KNN has high accuracy, low F1 score and MCC. This research asserts the KNN being weak in generation of independent datasets, thus not fit to be used in a weather prediction problem. These research findings can be taken on further into implementation of scalable NWPs for weather and climate products using minimal computation resources, lowering the overall budget constrains for National weather services across the globe, ultimately increasing access to accurate weather forecast, accelerating growth of other sectors of the economy like Agriculture, Transport, Construction, Manufacturing and Tourism.