Flood Prediction with Optimized Attributes and Clustering

Abstract

An emergency is a situation that poses an immediate risk to health, life, property, or environment. Most emergencies require urgent intervention to prevent a worsening of the situation. So, it is always better to predict the emergency before its happening and to take action for optimizing the loss. In this work, we tried to predict the flood by analysing the monthwise rainfall index of a particular area. First, we tried to find the months which have more contributions towards predicting the flood. For this, we used Particle Swarm Optimization (PSO) as feature selection technique and then applied classification algorithms such as J48 and Random Forest (RF). The experimentation was done for both without and with feature selection on the considered dataset. The results obtained without feature selection indicate that 70.34% and 78.81% of data are correctly classified and with feature selection 66.10% and 76.27% respectively in J48 and RF. Then we removed the class attribute from the dataset to see the effect of results when the class is not available and we applied K-mean and Density Based clustering techniques on the same dataset. It was observed from the results that Kmean with manhattan distance approach and Density Based clustering without feature selection classifies accurately 72.03% and 72.88% of data respectively. Similarly, when Kmean and Density Based clustering were used with feature selection, it was found that Kmean and Distanced Based clustering result in correct classification of 70.03% and 68.64% of data. We had also compared the model building time for both classification and clustering techniques using without and with feature selection. It was noticed that although the accuracy percentage was decreased with feature selection in both the cases, however, the model building time was reduced by 29%, 50%, 78%, and 60% in case of j48, RF, K-Mean, and Density Based techniques respectively.