Scholarlink

<p>Satellite Rainfall Products (SRPs) are vital for regions with limited ground-based observations offering high spatial and temporal resolution for precipitation monitoring. However, these datasets often contain systematic biases, necessitating correction for accurate hydrological applications. This study employs the Random Forest algorithm to reduce bias in three widely used SRPs: IMERG (Integrated Multi-Satellite Retrievals for GPM), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN), and Modern-Era Retrospective Analysis for Research and Applications Version 2 (MERRA-2). These SRPs were compared with observed rain gauge stations from the India Meteorological Department in the Kosi River basin. Contingency tests, cumulative distribution function, statistical metrics, box plot, and correlation matrix were employed at daily, monthly, seasonal, and annual time scales. Contingency test showed IMERG had the highest Probability of Detection (86.76%) and strong Critical Success Index (79.88%) followed by MERRA-2 and PERSIANN. The findings of the statistical analysis on a monthly scale showed that IMERG datasets are better than PERSIANN and MERRA-2 datasets with higher R<sup>2</sup>, KGE and lower RMSE values. IMERG also performed better at seasonal and annual scales making it suitable as rainfall input in a Transformer model for real-time water level prediction in flood forecasting. The model showed superior performance at 1-day lead time (R = 0.98, NSE = 0.96) with robust accuracy maintained even at 14-day lead time (R = 0.91, NSE = 0.82). Overall, in the absence of observed data the IMERG dataset proves to be a reliable alternative for hydrological and climatic studies.</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

From bias to forecast: advancing satellite rainfall accuracy and flood prediction with transformer modeling in the Kosi basin (India)

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Aditya Kumar Singh,
  • Vivekanand Singh,
  • Ajit Kumar,
  • Thendiyath Roshni

摘要

Satellite Rainfall Products (SRPs) are vital for regions with limited ground-based observations offering high spatial and temporal resolution for precipitation monitoring. However, these datasets often contain systematic biases, necessitating correction for accurate hydrological applications. This study employs the Random Forest algorithm to reduce bias in three widely used SRPs: IMERG (Integrated Multi-Satellite Retrievals for GPM), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN), and Modern-Era Retrospective Analysis for Research and Applications Version 2 (MERRA-2). These SRPs were compared with observed rain gauge stations from the India Meteorological Department in the Kosi River basin. Contingency tests, cumulative distribution function, statistical metrics, box plot, and correlation matrix were employed at daily, monthly, seasonal, and annual time scales. Contingency test showed IMERG had the highest Probability of Detection (86.76%) and strong Critical Success Index (79.88%) followed by MERRA-2 and PERSIANN. The findings of the statistical analysis on a monthly scale showed that IMERG datasets are better than PERSIANN and MERRA-2 datasets with higher R2, KGE and lower RMSE values. IMERG also performed better at seasonal and annual scales making it suitable as rainfall input in a Transformer model for real-time water level prediction in flood forecasting. The model showed superior performance at 1-day lead time (R = 0.98, NSE = 0.96) with robust accuracy maintained even at 14-day lead time (R = 0.91, NSE = 0.82). Overall, in the absence of observed data the IMERG dataset proves to be a reliable alternative for hydrological and climatic studies.