Scholarlink

<p>This study developed two one-dimensional variational technique-like machine learning (1DVAR-like ML) methods utilizing the Japan Meteorological Agency Meso-Scale Model (MSM) data to improve water vapor profile estimation from ground-based microwave radiometers (MWRs). The first method corrects systematic errors in the MSM (ML-CM), while the second incorporates MSM forecasts as explanatory variables in ML training (ML-MT). Validation using seven months of data from Tsukuba demonstrated that both methods significantly reduced root-mean-square error across all altitudes compared to conventional ML methods trained on reanalysis data. ML-CM effectively corrected MSM systematic biases, while ML-MT showed superior accuracy in the lowest atmospheric layer (below 500&#xa0;m). Case studies confirmed that the proposed methods can reproduce water vapor inversions and sharp vertical gradients. However, ML-CM’s performance for capturing temporal variation was limited by its high dependency on the MSM. In contrast, ML-MT effectively combined MWR observations with the MSM vertical structure as a guide. This enabled accurate estimations even when the MSM was erroneous, capturing short-term variations at 1-minute intervals that 3-hourly MSM updates could not resolve. Consequently, we achieved 1DVAR-like ML methods with much higher accuracy than conventional techniques and lower retrieval costs than physical 1DVAR estimation for water vapor profile.</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

1DVAR-Like Machine Learning for Water Vapor Profile Retrieval Using Ground-Based Microwave Radiometer and Numerical Weather Prediction Data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月10日

查看全文
  • Kentaro Araki,
  • Yuya Takashima

摘要

This study developed two one-dimensional variational technique-like machine learning (1DVAR-like ML) methods utilizing the Japan Meteorological Agency Meso-Scale Model (MSM) data to improve water vapor profile estimation from ground-based microwave radiometers (MWRs). The first method corrects systematic errors in the MSM (ML-CM), while the second incorporates MSM forecasts as explanatory variables in ML training (ML-MT). Validation using seven months of data from Tsukuba demonstrated that both methods significantly reduced root-mean-square error across all altitudes compared to conventional ML methods trained on reanalysis data. ML-CM effectively corrected MSM systematic biases, while ML-MT showed superior accuracy in the lowest atmospheric layer (below 500 m). Case studies confirmed that the proposed methods can reproduce water vapor inversions and sharp vertical gradients. However, ML-CM’s performance for capturing temporal variation was limited by its high dependency on the MSM. In contrast, ML-MT effectively combined MWR observations with the MSM vertical structure as a guide. This enabled accurate estimations even when the MSM was erroneous, capturing short-term variations at 1-minute intervals that 3-hourly MSM updates could not resolve. Consequently, we achieved 1DVAR-like ML methods with much higher accuracy than conventional techniques and lower retrieval costs than physical 1DVAR estimation for water vapor profile.