Scholarlink

<p>Accurate monthly rainfall prediction is vital for agriculture, water resource planning, and disaster preparedness. The goal of the study is to create an interpretable machine learning (ML) framework for monthly rainfall forecasting in four cities in Chhattisgarh, India: Ambikapur, Balrampur, Jagdalpur, and Sukma, utilizing climatic data from the National Centers for Environmental Prediction (NCEP). Three ML models, Support Vector Regression (SVR), Decision Tree Regressor (DTR), and Random Forest Regressor (RFR) are trained on 63 years of data (1948–2010) and evaluated on data from 2011 to 2017. Among these, the DTR consistently outperforms others across all locations, achieving R-Square (R<sup>2</sup>) values of 0.9952 (Ambikapur), 0.9527 (Balrampur), 0.9025 (Jagdalpur), and 0.9784 (Sukma); Mean Absolute Error (MAE) as low as 0.0786 mm; Root Mean Squared Error (RMSE) down to 0.3317 mm; and PBIAS ranging from –6.68% to 1.92%. Explainability is incorporated using SHapley Additive exPlanations (SHAP), which reveal precipitable water (ncep_pr_wtr) as the dominant contributor for Ambikapur, Balrampur, and Sukma, and vorticity at 500 hPa (ncep_p5_z) for Jagdalpur. A novel reverse verification approach is proposed by retraining the DTR model using only the most influential SHAP-identified feature. This reduced model still achieves R<sup>2</sup> above 0.93 at all stations, with RMSE ranging from 0.792 mm to 1.483 mm and PBIAS within ±5%, confirming that high accuracy can be maintained with significantly lower input complexity. A paired t-test confirms the statistical significance of DTR’s superiority. This work contributes a scalable, interpretable, and computationally efficient approach to rainfall prediction, suitable for operational deployment in resource-constrained environments.</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

Explainable AI reverse verification approach for monthly rainfall prediction in Chhattisgarh, India

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Rupesh Naik,
  • Babita Majhi

摘要

Accurate monthly rainfall prediction is vital for agriculture, water resource planning, and disaster preparedness. The goal of the study is to create an interpretable machine learning (ML) framework for monthly rainfall forecasting in four cities in Chhattisgarh, India: Ambikapur, Balrampur, Jagdalpur, and Sukma, utilizing climatic data from the National Centers for Environmental Prediction (NCEP). Three ML models, Support Vector Regression (SVR), Decision Tree Regressor (DTR), and Random Forest Regressor (RFR) are trained on 63 years of data (1948–2010) and evaluated on data from 2011 to 2017. Among these, the DTR consistently outperforms others across all locations, achieving R-Square (R2) values of 0.9952 (Ambikapur), 0.9527 (Balrampur), 0.9025 (Jagdalpur), and 0.9784 (Sukma); Mean Absolute Error (MAE) as low as 0.0786 mm; Root Mean Squared Error (RMSE) down to 0.3317 mm; and PBIAS ranging from –6.68% to 1.92%. Explainability is incorporated using SHapley Additive exPlanations (SHAP), which reveal precipitable water (ncep_pr_wtr) as the dominant contributor for Ambikapur, Balrampur, and Sukma, and vorticity at 500 hPa (ncep_p5_z) for Jagdalpur. A novel reverse verification approach is proposed by retraining the DTR model using only the most influential SHAP-identified feature. This reduced model still achieves R2 above 0.93 at all stations, with RMSE ranging from 0.792 mm to 1.483 mm and PBIAS within ±5%, confirming that high accuracy can be maintained with significantly lower input complexity. A paired t-test confirms the statistical significance of DTR’s superiority. This work contributes a scalable, interpretable, and computationally efficient approach to rainfall prediction, suitable for operational deployment in resource-constrained environments.