A Fourier-Guided Machine Learning Framework for District-Level Dew Point Comfort Forecasting Across 64 districts of Bangladesh (2025–2027)

Dew point, an absolute measure of atmospheric moisture, directly indicates when humid conditions become physiologically oppressive, yet it remains underutilized in operational heat-health forecasting for tropical monsoon nations. In Bangladesh, where 64 districts experience distinct moisture regimes, prolonged exposure to dew points exceeding 24°C impairs evaporative cooling, leading to heat strain even at moderate ambient temperatures. This study presents a machine learning framework for forecasting district-wise dew point comfort levels three years ahead.

Using 45 years (1980–2025) of meteorological data from Visual Crossing, we engineered cyclical features via Fourier series expansion to capture seasonal moisture periodicities. A Random Forest ensemble, selected after comparative trials against XGBoost and LightGBM, where all were trained on Fourier-enhanced features, was used to predict daily dew points. The model was validated using 5-fold cross-validation and produced recursive multi-step forecasts for 2025–2027 across all 64 districts. Dew point comfort was categorized using National Weather Service thresholds: oppressive (>24°C), muggy (21–24°C), humid (16–20°C), and pleasant (10–15°C).

Model performance was robust, with test R² values ranging from 0.844 (Cox's Bazar) to 0.943 (Netrokona) and 5-fold CV R² from 0.852 to 0.932. Over the three-year forecast period (2025–2027), coastal and south-central districts exhibited the highest cumulative burden of oppressive-plus-muggy days: Bhola (771 days), Bagherhat (756), Gopalganj (744) and Habiganj (723). By contrast, northern districts such as Kurigram recorded 666 such days. The Random Forest model demonstrated consistent generalization, with cross-validation accuracy within 1–2 points of test accuracy across all divisions.

These findings enable district-specific early warnings for dew point-driven heat stress, independent of temperature-based indices. Policymakers can use the district risk rankings to prioritize cooling center investments, issue labor advisories for outdoor workers, and provide humanitarian aid. A publicly accessible web dashboard is under development to translate forecasts into actionable local alerts.