Classification of Inundation Level using Tweets in Indonesian Language

Abstract

Extreme flood events are expected to occur more frequently as climate change has yet to show signs of improvement. This has the potential to lead to higher rainfall and floods that would come more quickly. Early warning systems may sometimes fail to provide quick information when conditions in the field may not match to what is known in the information center, such as a malfunctioning water pump or a water level that has increased relatively quickly. Therefore, this study aims to provide an alternative source of information that may provide inundation level during flood condition based on tweets from Twitter. The proposed model is expected to provide output in the form of inundation level categories, namely “high”, “medium”, “low”, and “unknown”. 10-fold stratified cross validation with seven variations of classifiers were used to evaluate the model. The best relevance classification resulted in 90.6% accuracy (SVM Linear SVC), 89.05% average precision (SVM RBF), and 82.03% average F1-score (SVM Linear SVC) and average recall of 84.10% (Logistic Regression). The best classification results of inundation level resulted in accuracy (82.74%), average precision (85.44%) average recall (68.07%) and average F1-score (71.43%). All of them were obtained by using the SVM Linear SVC.