Predictability of Naïve Bayes classifier for lahar hazard mapping by weather radar

Abstract

The aim of this study is to develop lahar hazard vulnerability as a warning system by introducing radar-rainfall observation to data mining technique of Naïve Bayes Classifier (NBC). NBC is used to estimate lahar occurrences based on the posterior probability of rainfall, topographic factor, soil moisture, and soil type as predictors. Rainfall intensity and working rainfall were obtained from a weather radar. The soil moisture is derived from SMAP satellite imagery. A river on Mount Merapi, a very active volcano in Indonesia, was selected as the target basin. Observed rainfall and recorded lahar events in Gendol River from October 2016 to February 2018 were divided into a training dataset and a testing dataset. Qualitative evaluation through visual assessment of the hazard map product reveals that the model could estimate the occurrences of lahar. The performance of the model in terms of accuracy, Brier score, and quantitative dichotomous quality indices showed a reasonable skill. The study suggests that the NBC technique is advantageous for estimating lahar occurrences that are displayed on hazard maps. This work is expected to contribute to debris flow hazard mitigation by the data mining approach in volcanic regions.