Clustering Algorithm Comparison and Seasonal Validation for Daily Weather Patterns in Bima Regency

Abstract

Bima Regency in West Nusa Tenggara lies in a monsoonal region with sharp seasonal contrast, yet its daily weather typology has never been characterized quantitatively and its surface observation network is sparse. In comparable studies, algorithm choice usually rests on internal validity indices alone, so the physical meaning of the clusters is rarely demonstrated. This study identifies the daily weather typology of Bima Regency and compares four clustering algorithms using a framework combining internal validation, external validation, and a stability test. The data comprise 4,383 daily observations of six meteorological variables from NASA POWER for 2014 to 2025. Preprocessing covered a quality audit, a log transformation of rainfall, and standardization. The number of clusters was set to three from a composite score of validity indices. Four algorithms were tested: K-Means, Gaussian Mixture Model, Self-Organizing Map, and Hierarchical Ward. K-Means achieved the best internal performance with a Silhouette of 0.2748, a Calinski-Harabasz of 1825.39, and a Davies-Bouldin of 1.1993. Hierarchical Ward reached the highest seasonal agreement with a Cramer's V of 0.6324, only marginally above the 0.6321 of K-Means. Based on the weighted composite score, K-Means was selected as the main algorithm and proved highly stable, with a bootstrap Adjusted Rand Index of 0.9800. The typologies are wet monsoon at 45.40 percent of days, windy dry season at 31.46 percent, and hot-dry high-radiation conditions at 23.13 percent. The clusters therefore carry meteorological meaning and can support agricultural planning and drought preparedness.