Scholarlink

<p>Changepoint (CP) detection in climate time series is challenging due to seasonality, trends, and serial correlation. This study introduces a method that combines regression modeling with autoregressive error structures to detect abrupt regime shifts. The model breaks down the series into seasonal means, a linear trend, and shift offsets, while capturing seasonally varying autocorrelation using a periodic autoregressive process (PAR). Parameters are estimated via the iterative Cochrane–Orcutt method and Yule–Walker equations, and the model selection is guided by the Minimum Description Length principle (MDL). A multi-island genetic algorithm (GA) searches for optimal CPs and autoregressive (AR) orders. Validation on synthetic data shows the approach reliably detects CPs across varying shift magnitudes and autocorrelation levels. A homogenization procedure for the monthly series from a network of meteorological stations across a study region is also proposed and built on the aforementioned GA approach to minimize an MDL objective function.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Genetic-algorithm based changepoints detection and homogenization of precipitation series

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Youssef Saliba,
  • Alina Bărbulescu

摘要

Changepoint (CP) detection in climate time series is challenging due to seasonality, trends, and serial correlation. This study introduces a method that combines regression modeling with autoregressive error structures to detect abrupt regime shifts. The model breaks down the series into seasonal means, a linear trend, and shift offsets, while capturing seasonally varying autocorrelation using a periodic autoregressive process (PAR). Parameters are estimated via the iterative Cochrane–Orcutt method and Yule–Walker equations, and the model selection is guided by the Minimum Description Length principle (MDL). A multi-island genetic algorithm (GA) searches for optimal CPs and autoregressive (AR) orders. Validation on synthetic data shows the approach reliably detects CPs across varying shift magnitudes and autocorrelation levels. A homogenization procedure for the monthly series from a network of meteorological stations across a study region is also proposed and built on the aforementioned GA approach to minimize an MDL objective function.