Scholarlink

Abstract <p>Heatwaves threaten public health and economies across tropical regions like Bangladesh, yet traditional threshold-based forecasting fails to capture their complex spatiotemporal dynamics. We present HWNet-BD, a novel hybrid deep learning model that combines Convolutional Neural Networks for spatial feature extraction, Long Short-Term Memory networks for temporal modeling, and Transformer attention mechanisms for refined prediction. The model learns from historical patterns in nine meteorological features to forecast heatwave events one day in advance—a critical window for activating emergency response protocols and protecting vulnerable populations. Evaluated on an independent held-out test year (2025), HWNet-BD achieves 93.71% accuracy and 96.36% ROC-AUC, significantly outperforming baseline approaches. Sensitivity analysis across multiple threshold definitions confirms model robustness. This data-driven early-warning capability provides operational decision-makers with actionable lead time to mobilize resources, issue public advisories, and reduce heat-related mortality in climate-vulnerable regions.</p> Graphical Abstract <p>The graphical abstract illustrates the complete workflow of the HWNet-BD framework, designed to predict heatwave events in Bangladesh using multivariate ERA5 reanalysis data. The top panel visualizes the study area and the rigorous data processing pipeline, which transforms hourly meteorological data (2014–2025) through cleaning, feature engineering (calculating heat index, humidity, and anomalies), and temporal resampling, alongside SMOTE techniques to address the scarcity of heatwave samples. The center panel details the proposed hybrid deep learning architecture, demonstrating the sequential flow of data through a Convolutional Neural Network (CNN) block for spatial feature extraction, a Long Short-Term Memory (LSTM) network for modeling 7-day temporal dependencies, and a Transformer encoder that applies attention mechanisms to refine predictions before the final classification head outputs a binary “Heatwave” or “Non-Heatwave” result. The bottom panel displays extensive performance evaluations, including training and validation curves, confusion matrices for the held-out 2025 test set that confirm high sensitivity, and ROC/Precision-Recall curves showcasing a 96.36% AUC. A comparative analysis plot highlights HWNet-BD’s superior performance against baseline models such as GRU, LSTM, and ConvLSTM across accuracy, F1-score, and recall metrics. </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

HWNet-BD: A Hybrid Deep Learning Model for Accurate Heatwave Detection in Bangladesh Using Multivariate ERA5 Reanalysis Data

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Abdu Naim,
  • Xia Jingming,
  • Suaibia Tasnim

摘要

Abstract

Heatwaves threaten public health and economies across tropical regions like Bangladesh, yet traditional threshold-based forecasting fails to capture their complex spatiotemporal dynamics. We present HWNet-BD, a novel hybrid deep learning model that combines Convolutional Neural Networks for spatial feature extraction, Long Short-Term Memory networks for temporal modeling, and Transformer attention mechanisms for refined prediction. The model learns from historical patterns in nine meteorological features to forecast heatwave events one day in advance—a critical window for activating emergency response protocols and protecting vulnerable populations. Evaluated on an independent held-out test year (2025), HWNet-BD achieves 93.71% accuracy and 96.36% ROC-AUC, significantly outperforming baseline approaches. Sensitivity analysis across multiple threshold definitions confirms model robustness. This data-driven early-warning capability provides operational decision-makers with actionable lead time to mobilize resources, issue public advisories, and reduce heat-related mortality in climate-vulnerable regions.

Graphical Abstract

The graphical abstract illustrates the complete workflow of the HWNet-BD framework, designed to predict heatwave events in Bangladesh using multivariate ERA5 reanalysis data. The top panel visualizes the study area and the rigorous data processing pipeline, which transforms hourly meteorological data (2014–2025) through cleaning, feature engineering (calculating heat index, humidity, and anomalies), and temporal resampling, alongside SMOTE techniques to address the scarcity of heatwave samples. The center panel details the proposed hybrid deep learning architecture, demonstrating the sequential flow of data through a Convolutional Neural Network (CNN) block for spatial feature extraction, a Long Short-Term Memory (LSTM) network for modeling 7-day temporal dependencies, and a Transformer encoder that applies attention mechanisms to refine predictions before the final classification head outputs a binary “Heatwave” or “Non-Heatwave” result. The bottom panel displays extensive performance evaluations, including training and validation curves, confusion matrices for the held-out 2025 test set that confirm high sensitivity, and ROC/Precision-Recall curves showcasing a 96.36% AUC. A comparative analysis plot highlights HWNet-BD’s superior performance against baseline models such as GRU, LSTM, and ConvLSTM across accuracy, F1-score, and recall metrics.