Classification of Disaster Impact Severity in North Maluku Using Random Forest Based on Spatial-Temporal Characteristics

Abstract

Disaster impact severity is typically assessed only after casualty and damage data have been collected, even though the spatial-temporal characteristics of an event are known from the outset. This study aims to construct a disaster severity index using multi-component Principal Component Analysis (PCA) and to classify it using only spatial-temporal characteristics, deliberately excluding impact variables as predictors to avoid target leakage. The dataset comprises 239 disaster events in North Maluku Province (2021-2025) sourced from DIBI-BNPB. The severity index was built through log1p transformation, standardization, KMO (0.757) and Bartlett's Test (chi- square=1010.81; p<0.001), and Kaiser-criterion PCA yielding four principal components (cumulative variance 64.44%). The severity class was divided into four quartile-based classes and validated through a monotonic relationship with mean affected/displaced victims and a significant ANOVA test (F=17.574; p<0.001). Classification used five spatial-temporal features (year, month, week, regency/city, disaster type) under a temporal split scheme, comparing Random Forest, XGBoost, and a Dummy Classifier. Random Forest achieved the best performance with a Macro-F1 of 0.146 on the 2025 test set, outperforming the baseline (0.041) though the absolute value remains low. An ablation test revealed that a temporal-only feature scenario achieved the highest Macro-F1 (0.180), indicating limitations in the categorical encoding scheme. Sensitivity analysis confirmed the results are robust to changes in the class scheme (Macro-F1 difference of 0.047). The main contribution of this study is a validated severity-index construction methodology and early empirical evidence on the limitations of spatial- temporal features for early disaster severity classification.