Clustering for Regional Time Trend in the Nonstationary Extreme Distribution

Sungchul Hong, Jong‐June Jeon, Yongdai KimWater 2022 · DOI

TL;DR

This paper introduces a novel, distribution-free method for identifying and clustering regional time trends in nonstationary hydrological processes. By unifying parameter estimation and clustering, it offers a more robust approach to analyzing extreme events, as demonstrated by its superior performance in simulations and its application to South Korean precipitation data.

What Problem Does This Paper Address?

Estimating the properties of extreme hydrological events (like floods or droughts) is crucial for risk assessment and infrastructure design. However, these estimations often rely on the assumption that the data is stationary, meaning its statistical properties don't change over time. This assumption is increasingly challenged by climate change, which leads to nonstationary processes where trends and shifts are evident. Traditional regional frequency analysis (RFA) methods, which pool data from similar regions to improve estimates, struggle with nonstationary data because the concept of "similar regions" becomes ill-defined. Furthermore, existing de-trending methods are often distribution-specific and applied independently to each hydrological process, ignoring the spatial proximity of observation sites. This paper addresses the need for a method that can:

  1. Handle nonstationary hydrological processes without assuming a specific underlying distribution.
  2. Simultaneously identify and cluster regional trends, acknowledging that adjacent sites might share similar temporal patterns.
  3. Improve the accuracy and reduce the variance of trend estimations by leveraging spatial information.

Methods

The core of the proposed method lies in a distribution-free de-trending technique that integrates sparse regularization for parameter estimation and clustering. Here's a breakdown:

  • Nonstationary Model: The authors model the nonstationary process with a linear time-variant mean: $U_{tj} = \beta_j t + \epsilon_{tj}$, where $\beta_j$ is the trend parameter for site $j$, $t$ is time, and $\epsilon_{tj}$ is an error term. The goal is to estimate these $\beta_j$ values and group sites with similar $\beta_j$.
  • Distribution-Free De-trending: Instead of assuming a specific distribution for $\epsilon_{tj}$, the method focuses on the differences between observations, making it robust to various underlying distributions.
  • Sparse Regularization (Fused Lasso): The method employs a fused lasso penalty within a regression framework. This technique is powerful because it can simultaneously:
  • Shrinkage (Lasso): Drive some trend coefficients ($\beta_j$) exactly to zero, effectively identifying sites with no significant trend.
  • Grouping (Fused Lasso): Encourage adjacent or similar sites to have the same trend coefficient, thus performing the clustering. This is achieved by penalizing the difference between coefficients of neighboring sites.
  • Unified Framework: The key innovation is that parameter estimation (of $\beta_j$) and clustering (grouping of $\beta_j$) are performed simultaneously within a single optimization problem. This contrasts with traditional approaches where clustering is a separate, preceding step.
  • Recovering Procedure: To mitigate potential bias introduced by the regularization, a two-stage process is used: first, estimate clustered trends using fused lasso; second, refit the linear trends within each identified cluster to obtain more accurate estimates.
  • Model Selection (BIC): The Bayesian Information Criterion (BIC) is used to select the optimal tuning parameters for the regularization method, which in turn determines the number of clusters.

Datasets

The paper evaluates its method using two types of data:

  • Simulated Data: The authors generate synthetic data from a Generalized Extreme Value (GEV) distribution with varying trend parameters ($\beta_j$) to test the performance of their method against several benchmarks under controlled conditions.
  • Real-World Data: Annual maximum daily precipitation data from 60 meteorological observatories across the South Korean peninsula, collected from 1971 to 2021, is used for a case study. This dataset is publicly available.

Key Insights and Ideas

  • Simultaneous Clustering and Estimation: The most significant insight is the power of integrating trend estimation and clustering into a single framework. This avoids the propagation of errors from a potentially suboptimal pre-clustering step and allows the data to inform both aspects simultaneously.
  • Distribution-Free Robustness: By not relying on specific distributional assumptions, the method is broadly applicable to various hydrological processes, which is a major advantage in real-world scenarios where distributions might be unknown or complex.
  • Spatial Information for Trend Analysis: The fused lasso penalty effectively leverages the spatial proximity of observation sites. This is crucial because hydrological processes in geographically close areas are more likely to share similar underlying trends, even if their absolute values differ.
  • Improved Accuracy via Grouping: The simulation results show that by grouping sites with similar trends, the variance of the estimated coefficients is significantly reduced, leading to lower Mean Squared Error (MSE) compared to methods that treat each site independently.
  • Regional Patterns of Extreme Precipitation: The real-data analysis reveals distinct regional patterns in the trends of extreme precipitation across South Korea, with increasing trends in the east and south, and decreasing or stationary trends in the west. This highlights the spatial heterogeneity of climate change impacts.

Key Results

  • Simulation Performance: The proposed method consistently outperformed other benchmark methods (naive LSE, naive LAD, rank LSE, rank LAD) in terms of Mean Squared Error (MSE) in simulations. In one simulation, the proposed method achieved an average of 3.6 clusters, demonstrating its ability to group sites effectively.
  • South Korea Precipitation Trends: The analysis identified regional clusters of time trends in annual maximum daily precipitation.
  • The largest cluster (27 sites) in the southeastern coastal line showed an increasing trend (coefficient ~0.47).
  • Other clusters showed varying trends, with weaker trends in the west and midlands, and stronger positive trends in Jeju and Ulleung islands.
  • A notable finding was a negative trend in Ganghwa, located in the northwestern end.

Limitations and Future Work

  • Tuning Parameter Sensitivity: The authors acknowledge that the proposed estimator is severely dependent on the tuning parameter selection. A careless choice can lead to misleading results for the linear time trend. They explicitly state that developing or justifying a more robust tuning parameter selection method is a key area for future work.
  • Computational Complexity: While the ADMM method is mentioned as a way to handle large datasets, the optimization of non-differentiable functions can still be computationally intensive.
  • Further Analysis of Extreme Events: The paper suggests that further analysis is needed to detect patterns of extreme climate events, especially given the counter-intuitive finding of decreasing trends in southern sites, which contrasts with previous research on extreme event probability.
  • Generalization to Other Nonstationary Models: The current method focuses on linear time trends. Future work could explore extending it to other forms of nonstationarity.

Summary updated

Abstract

Since the estimation of tail properties requires a stationarity of observations, it is necessary to develop a de-trending method not dependent on underlying distributions for nonstationary hydrological processes. Moreover, de-trending has been independently applied to hydrological processes, even though the processes are observed in geometrically adjacent sites. This paper presents a distribution-free de-trending method for nonstationary hydrological processes. Our method also provides clustered regional trends obtained by sparse regularization in a general distribution. It aggregates the parameter estimation and clustering within a unified framework. In the simulation study, our proposed method has superiority over other compared methods with respect to MSE and variance of coefficients. In real data analysis, the clustered trends of the annual maximum precipitation in the South Korean peninsula are reported, and the patterns of the estimated trends are visualized.
BibTeX
@inproceedings{hong2022clustering,
  author = {Sungchul Hong and Jong‐June Jeon and Yongdai Kim},
  title = {{Clustering for Regional Time Trend in the Nonstationary Extreme Distribution}},
  booktitle = {Water},
  year = {2022},
  doi = {10.3390/w14111720},
}
Figure