A Machine Learning Framework for Urban Drought Risk Assessment under Future Climate Scenarios: A Case Study of Gangneung City, South Korea

A Machine Learning Framework for Urban Drought Risk Assessment under Future Climate Scenarios: A Case Study of Gangneung City, South Korea

Yunseo Bae¹, Yujin Jung¹, Jiyoon Choi¹, Younghun Lee² Sangchul Lee^1,2*

 

¹ Division of Environmental Science & Ecological Engineering, Korea University, Seoul 02841, Republic of Korea

² Department of Environmental Science & Ecological Engineering, Korea University, Seoul 02841, Republic of Korea

 

* Corresponding author: Sangchul Lee (slee2024@korea.ac.kr)

 

Abstract:

In 2025, a national disaster was declared in the Republic of Korea due to severe drought, particularly Gangneung City in Gangwon State. In urban areas, drought risk is shaped not only by meteorological conditions but also by anthropogenic factors. However, conventional drought assessments largely rely on climatic or hydrological indices and often fail to reflect these socio-infrastructure factors. Recently, machine learning (ML) is widely adopted due to its ability to capture complex, nonlinear interactions among various factors. Accordingly, this study develops a ML-based framework to reproduce historical drought and to predict future urban drought risk under climate change scenarios in Gangneung City. Drought occurrence data from 2016 to 2025 were classified into five stages (Normal, Attention, Caution, Warning, and Severe) and used as multi-class target variables. Input data included meteorological (precipitation, temperature, humidity, wind speed, and evapotranspiration), topographic (DEM-based elevation, slope, aspect, watershed characteristics, and land cover), and anthropogenic variables (water supply infrastructure, population, and tourism activity). All input variables were spatially aggregated to administrative units, ensuring consistency with the spatial resolution of the observed drought occurrence data. An AutoML approach was applied to compare multiple classification algorithms and to identify the optimal model. Model performance was evaluated using time-aware validation strategies, including a temporal train–test split and time-series cross-validation. SHAP analysis was also employed to interpret the relative importance of key drought drivers. Future drought risk was projected by applying meteorological inputs derived from SSP2-4.5, SSP3-7.0, and SSP5-8.5 scenarios to the trained model, while other factors were assumed to remain. Administrative-unit-based drought occurrence probabilities were analyzed for the near (2030–2050), mid (2051–2060), and far future (2080–2100). In addition, hypothetical policy-oriented scenarios were explored by modifying anthropogenic variables, such as water leakage rates and tourism demand, to assess the sensitivity of drought risk to management assumptions. The findings from this study would demonstrate the ML-based framework is efficient to predict urban drought risk, supporting region-specific drought mitigation and climate adaptation strategies.

 

Key words: machine learning, urban drought, anthropogenic factors, drought risk mapping, climate change

 

Acknowledgement

Following are results of a study on the "Convergence and Open Sharing System "Project, supported by the Ministry of Education and National Research Foundation of Korea