Artificial Intelligence-Based Spatiotemporal Clustering for Identifying Homogeneous Climatic Zones in West Java
TL;DR
Distinguishing the evaluation metrics of the 54-zone candidate stage from the 50-zone cartographic structure of the spatial refinement stage clarifies the analytical sequence and supports a more precise interpretation of cluster quality, climatic homogeneity, and final zone boundaries.
Abstract
Homogeneous Climatic Zones (ZOM) support operational seasonal services in Indonesia, particularly across regions where rainfall changes sharply over short distances. This study evaluates a climatic regionalization framework that combines Uniform Manifold Approximation and Projection (UMAP) and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) using BMKG gridded rainfall data across West Java at approximately 5 km resolution for the 1991–2020 period. Each of the 1,214 grid cells was represented by a 36-dasarian climatological rainfall profile. A search across 360 hyperparameter combinations yielded 324 executable configurations, followed by spatial absorption of transition grids and layered evaluation using the Silhouette Score, Heterogeneity Index (HI), and operational tractability. Scenario 60 (n_neighbors = 5, min_dist = 0.01, min_cluster_size = 10, min_samples = 5) produced 54 initial zones with a Silhouette Score of 0.5633 and an HI of 7.239, compared with an HI of 8.464 for the 41-zone operational baseline. Subsequent spatial refinement produced a corrected final map containing 50 sequential ZOM. Distinguishing the evaluation metrics of the 54-zone candidate stage from the 50-zone cartographic structure of the spatial refinement stage clarifies the analytical sequence and supports a more precise interpretation of cluster quality, climatic homogeneity, and final zone boundaries.
