Scholarlink

<p>This study investigates the performance of climate models in simulating rainfall extremes across West Africa and the potential of machine learning (ML) to improve model accuracy through bias correction techniques. Using daily precipitation data from Thirteen (13) CMIP6 models and CHIRPS observations (1983–2014). The study analyzed Consecutive Dry Days (CDD), Consecutive Wet Days (CWD), Very Heavy Precipitation days (R20mm), Total Precipitation (PRCPTOT), and Simple Daily Intensity Index (SDII) developed by ETCCDI to assess patterns in rainfall extremes and their trends. A back propagation algorithm (feed-forward neural network model) was trained to correct biases, and its output was evaluated against observations using Taylor’s diagram and Mann–Kendall trend analysis. The ML model captured the latitudinal rainfall pattern and intensity in SDII with a correlation coefficient of 0.93, central root mean square error CRMSE of 0.93, Nash–Sutcliffe Efficiency (NSE) of 0.91, and Mean Absolute Percentage Error (MAPE) of 8.54%, and PRCPTOT with a correlation coefficient of 0.81, central Root Mean Square Error (CRMSE) of 211.70, Nash–Sutcliffe Efficiency (NSE) of 0.72, and Mean Absolute Percentage Error (MAPE) of 12.43%, closely matching the observation. Trend analyses showed that rainfall changes across West Africa are generally weak and scattered, with dry days more evident along the Guinea coast and an increase in rainfall intensity observed in central and southern regions of the study area. While Global Climate Models GCMs often overestimate or underestimate trends, the ML model demonstrated improvement across most indices, outperforming both individual GCMs and the multi-model ensemble. These findings show the limitations of GCM outputs for regional studies and the potential of machine learning in bias correction of global climate models.</p> Graphical Abstract <p> This study presents the effort to use machine learning (artificial neural network) to correct biases in extreme precipitation indices over West Africa. The region depends heavily on rainfall for food production, water resources, and disaster management, yet its rainfall is highly variable and poorly captured by current climate models due to complex local processes and coarse model resolution. Thirteen (13) CMIP6 models and CHIRPS observational data were used to compute five extreme rainfall indices (CDD, CWD, PRCPTOT, SDII, and R20mm) for the period 1983–2014. These were then used as input into a back propagation-based Artificial Neural Network model, trained on 1983–2006 data and tested on 2007–2014. Results show that the ANN performed significantly better than raw CMIP6 models and their ensemble mean, especially for SDII (correlation of 0.93, CRMSE of 0.93), PRCPTOT (correlation 0.81, CRMSE 211.7), and CWD (correlation 0.82, CRMSE 6.08), closely capturing observed values. While the ANN underestimated very heavy rainfall days (R20mm), however, it was still more consistent than individual models, which generally overestimated precipitation values. For CDD, the ANN had a strong correlation of 0.91 and minimal bias. This highlights the potential of machine learning for improving climate model accuracy in regions with high vulnerability like West Africa.</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

Machine Learning for Bias-Correction of Extreme Precipitation Indices Over West Africa

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Isiaq Adebayo Okeyode,
  • Vincent Olanrewaju Ajayi,
  • Ayomide Victor Arowolo,
  • Michael Temitayo Odunmorayo,
  • Michael Chukwuemeka Ochei,
  • Ibraheem Ayomide Raji,
  • Ibrahim Adedapo Tella

摘要

This study investigates the performance of climate models in simulating rainfall extremes across West Africa and the potential of machine learning (ML) to improve model accuracy through bias correction techniques. Using daily precipitation data from Thirteen (13) CMIP6 models and CHIRPS observations (1983–2014). The study analyzed Consecutive Dry Days (CDD), Consecutive Wet Days (CWD), Very Heavy Precipitation days (R20mm), Total Precipitation (PRCPTOT), and Simple Daily Intensity Index (SDII) developed by ETCCDI to assess patterns in rainfall extremes and their trends. A back propagation algorithm (feed-forward neural network model) was trained to correct biases, and its output was evaluated against observations using Taylor’s diagram and Mann–Kendall trend analysis. The ML model captured the latitudinal rainfall pattern and intensity in SDII with a correlation coefficient of 0.93, central root mean square error CRMSE of 0.93, Nash–Sutcliffe Efficiency (NSE) of 0.91, and Mean Absolute Percentage Error (MAPE) of 8.54%, and PRCPTOT with a correlation coefficient of 0.81, central Root Mean Square Error (CRMSE) of 211.70, Nash–Sutcliffe Efficiency (NSE) of 0.72, and Mean Absolute Percentage Error (MAPE) of 12.43%, closely matching the observation. Trend analyses showed that rainfall changes across West Africa are generally weak and scattered, with dry days more evident along the Guinea coast and an increase in rainfall intensity observed in central and southern regions of the study area. While Global Climate Models GCMs often overestimate or underestimate trends, the ML model demonstrated improvement across most indices, outperforming both individual GCMs and the multi-model ensemble. These findings show the limitations of GCM outputs for regional studies and the potential of machine learning in bias correction of global climate models.

Graphical Abstract

This study presents the effort to use machine learning (artificial neural network) to correct biases in extreme precipitation indices over West Africa. The region depends heavily on rainfall for food production, water resources, and disaster management, yet its rainfall is highly variable and poorly captured by current climate models due to complex local processes and coarse model resolution. Thirteen (13) CMIP6 models and CHIRPS observational data were used to compute five extreme rainfall indices (CDD, CWD, PRCPTOT, SDII, and R20mm) for the period 1983–2014. These were then used as input into a back propagation-based Artificial Neural Network model, trained on 1983–2006 data and tested on 2007–2014. Results show that the ANN performed significantly better than raw CMIP6 models and their ensemble mean, especially for SDII (correlation of 0.93, CRMSE of 0.93), PRCPTOT (correlation 0.81, CRMSE 211.7), and CWD (correlation 0.82, CRMSE 6.08), closely capturing observed values. While the ANN underestimated very heavy rainfall days (R20mm), however, it was still more consistent than individual models, which generally overestimated precipitation values. For CDD, the ANN had a strong correlation of 0.91 and minimal bias. This highlights the potential of machine learning for improving climate model accuracy in regions with high vulnerability like West Africa.