Decision-oriented computational framework for drought hazard-state classification with a California case study

This study develops a decision-oriented computational framework for multi-level drought hazard-state classification under class imbalance and class-boundary ambiguity. Using station-level climate observations from the Global Historical Climatology Network-Daily (GHCND) and operational drought labels from the U.S. Drought Monitor (USDM), the framework is evaluated through a case study of 48 monitoring stations across California from 2010 to 2023. An initial six-class formulation, comprising No Drought, D0 abnormally dry conditions, and four drought severity levels (D1 to D4), is assessed using several machine learning models, including logistic regression , support vector machine , k-nearest neighbors , multilayer perceptron , and random forest . Results reveal systematic misclassification near the transitional No Drought–D0 boundary, reflecting class imbalance, categorical proximity, and the inherent ambiguity of early drought conditions. To address this challenge, a two-stage classification framework is proposed. In the first stage, a decision tree model separates No Drought from drought-relevant D0–D4 conditions for operational screening. In the second stage, a random forest model further classifies routed drought-relevant records into D0–D4 hazard states. Additional validation experiments, including a persistence baseline, end-to-end two-stage evaluation, temporal holdout, spatial holdout, feature ablation, feature importance, transition-week diagnostics, and missing-data auditing, are conducted to assess the robustness and limitations of the framework beyond random-split accuracy. Under the expanded stratified random split, the single-stage RF achieved slightly higher aggregate accuracy (0.9739 versus 0.9684) and weighted F1-score (0.9740 versus 0.9686) than the end-to-end two-stage framework. The value of the two-stage formulation lies instead in its explicit separation of screening and D0–D4 hazard-state classification, stage-specific diagnosis, and decision-structured interpretation. Although California serves as the case study, the proposed framework is formulated to support adaptation to other regions and can support nowcasting-oriented drought hazard assessment, infrastructure resilience screening, regional monitoring, and risk-informed resource allocation where comparable monitoring data are available.