Scholarlink

Abstract <p>In climatology, air humidity is of fundamental importance due to water vapor being a key component of the climate system. However, the series of measurements of relative humidity in situ are sparse in space and time, especially in the early 20th century, which makes it difficult to analyze long-term climatic changes. In this paper, we propose a new approach to reconstructing data on near-surface atmospheric humidity over the ocean using machine learning methods. The research was based on the author’s DISO3 database, formed on the basis of carefully selected ship observations from the international ICOADS array. To account for regional and seasonal humidity patterns, the data was divided into <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(5^{\circ}\times 5^{\circ}\)</EquationSource> <!--BPhysMGU2570295Vostrikova-m1--> </InlineEquation> spatial cells with further identification of the seasons. The paper compares four machine learning models: linear regression, random forest, gradient boosting (CatBoost) and a fully connected neural network. The results of the study showed that the CatBoost model demonstrates the best accuracy of relative humidity reconstruction compared to other methods considered. The analysis of the significance of the input parameters confirmed the physical validity of the model, identifying key meteorological factors that significantly affect the accuracy of the reconstruction. Bootstrap uncertainty analysis showed the statistical stability of the results obtained. The study confirmed the promise of machine learning methods for reconstructing humidity data. The proposed approach makes it possible to effectively reconstruct gaps in historical data, extending the time frame of climate research into the past.</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

Reconstruction of Atmospheric Surface Relative Humidity over the Ocean from Concurrent Meteorological Measurements and Observations Using Machine Learning Methods

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • S. A. Vostrikova,
  • M. A. Krinitsky,
  • M. P. Alexandrova,
  • S. K. Gulev

摘要

Abstract

In climatology, air humidity is of fundamental importance due to water vapor being a key component of the climate system. However, the series of measurements of relative humidity in situ are sparse in space and time, especially in the early 20th century, which makes it difficult to analyze long-term climatic changes. In this paper, we propose a new approach to reconstructing data on near-surface atmospheric humidity over the ocean using machine learning methods. The research was based on the author’s DISO3 database, formed on the basis of carefully selected ship observations from the international ICOADS array. To account for regional and seasonal humidity patterns, the data was divided into \(5^{\circ}\times 5^{\circ}\) spatial cells with further identification of the seasons. The paper compares four machine learning models: linear regression, random forest, gradient boosting (CatBoost) and a fully connected neural network. The results of the study showed that the CatBoost model demonstrates the best accuracy of relative humidity reconstruction compared to other methods considered. The analysis of the significance of the input parameters confirmed the physical validity of the model, identifying key meteorological factors that significantly affect the accuracy of the reconstruction. Bootstrap uncertainty analysis showed the statistical stability of the results obtained. The study confirmed the promise of machine learning methods for reconstructing humidity data. The proposed approach makes it possible to effectively reconstruct gaps in historical data, extending the time frame of climate research into the past.