Scholarlink

<p>Accurate forecasting of the Indian summer monsoon (ISM) is critical for water resource management, agricultural planning, and disaster mitigation. In this study, a deep learning (DL) methodology is developed to assess whether pre-monsoon atmospheric features can reliably predict ISM precipitation. Several key parameters, e.g., sea surface temperature, skin temperature, 2-meter temperature, total column water vapour, and a convective index (vertically integrated moisture divergence) are integrated and merged to form a unified thermal variable that captures both marine and terrestrial influences. The domain-summed inland Indian region precipitation is analyzed over a test period from 2012 to 2024 using Cumulative Distribution Functions (CDF), spatial precipitation maps, frequency-of-exceedance curves, and grouped annual performance metrics. The CDF analysis reveals that the ERA5 reanalysis cumulative monthly precipitation (used as a reference data) for the test period has a median of approximately 1.2&#xa0;m and a mean of 1.3&#xa0;m. In comparison, the predicted distribution exhibits a lower median of approximately 1.1&#xa0;m and a mean of nearly 1.0&#xa0;m, indicating systematic underestimation, particularly in the upper tail, where the 90th percentile of ERA5 values reaches roughly 1.8&#xa0;m, compared to 1.5&#xa0;m for the predictions. Spatial maps reveal that, although the model accurately captures broad precipitation patterns along the Western Ghats and the Bay of Bengal, regional biases persist, with coastal areas often overestimated and inland regions underestimated. The frequency-of-exceedance analysis further indicates that the model underestimates the occurrence of heavy precipitation events (HPE). Grouped annual performance metrics, featuring an average correlation of 0.78 and a Kling–Gupta Efficiency (KGE) of 0.72, underscore the model’s moderate skill across varying monsoon conditions. The statistical analysis of model performance suggests that although there are associated specific systematic biases intrinsic to the model, overall, our integrated DL model effectively captures the general properties of monsoon precipitation and hence can be utilized for extended-range precipitation forecasting.</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

Deep learning-driven mapping of pre-monsoon features for Indian summer monsoon precipitation forecasting

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Deepak Singh Bisht,
  • Vivekananda Hazra,
  • Durgesh Nandan Piyush

摘要

Accurate forecasting of the Indian summer monsoon (ISM) is critical for water resource management, agricultural planning, and disaster mitigation. In this study, a deep learning (DL) methodology is developed to assess whether pre-monsoon atmospheric features can reliably predict ISM precipitation. Several key parameters, e.g., sea surface temperature, skin temperature, 2-meter temperature, total column water vapour, and a convective index (vertically integrated moisture divergence) are integrated and merged to form a unified thermal variable that captures both marine and terrestrial influences. The domain-summed inland Indian region precipitation is analyzed over a test period from 2012 to 2024 using Cumulative Distribution Functions (CDF), spatial precipitation maps, frequency-of-exceedance curves, and grouped annual performance metrics. The CDF analysis reveals that the ERA5 reanalysis cumulative monthly precipitation (used as a reference data) for the test period has a median of approximately 1.2 m and a mean of 1.3 m. In comparison, the predicted distribution exhibits a lower median of approximately 1.1 m and a mean of nearly 1.0 m, indicating systematic underestimation, particularly in the upper tail, where the 90th percentile of ERA5 values reaches roughly 1.8 m, compared to 1.5 m for the predictions. Spatial maps reveal that, although the model accurately captures broad precipitation patterns along the Western Ghats and the Bay of Bengal, regional biases persist, with coastal areas often overestimated and inland regions underestimated. The frequency-of-exceedance analysis further indicates that the model underestimates the occurrence of heavy precipitation events (HPE). Grouped annual performance metrics, featuring an average correlation of 0.78 and a Kling–Gupta Efficiency (KGE) of 0.72, underscore the model’s moderate skill across varying monsoon conditions. The statistical analysis of model performance suggests that although there are associated specific systematic biases intrinsic to the model, overall, our integrated DL model effectively captures the general properties of monsoon precipitation and hence can be utilized for extended-range precipitation forecasting.