Dataset transformation using hybrid method of polar-based cartesian and image filtering technique for annual rainfall clustering

Abstract

Clustering is categorized as unsupervised learning because there is no class label information available in grouping a dataset. For this reason, an assessment of the quality of the clustering results is critical. In general, two essential clustering parameters are the similarities between cluster members in a cluster and the cluster centers’ separation. Various approaches can be used to improve a clustering algorithm’s performance: raw data pre-processing, cluster center initialization techniques, objective function assignment, modification of specific steps, and others. The subject of this study is the pattern of rainfall every month during the year of the observation period obtained from the clustering process. This study aims to improve the performance of K-Mean Clustering through manipulation of raw data pre-processing into certain datasets. Image filtering technique is used to generate a dataset based on the relationship between neighboring rainfall values. Polar-based Cartesian data space transformation is used to generate a dataset based on a range of rainfall values for each month during the year of the observation period. Four scenarios have been used to test the performance of the proposed method. The study results show that the proposed method produces the highest performance ratio (54.79%) of all scenarios’ total average GOS (Global Optimum Solution). Meanwhile, increasing GOS to the original method also resulted in the highest increase in GOS ratio (63.68%) compared to other methods. Further studies will focus on the application of the proposed methods for improving the performance of SOM and Fuzzy C-Mean Clustering.