Enhancing Coastal Dew Point Prediction: A Hybrid Deep Learning Framework for Southeastern Bangladesh

The dew point temperature is one of the main indicators of atmospheric humidity, and its accurate prediction is a vital factor in dynamic coastlands to help attain climate resilience, agriculture, and human health. This study proposes a new hybrid deep learning architecture, the Convolutional Neural Network-Residual Long Short-Term Memory-Concatenation Network (CNN-ResLSTM-ConNet) model that can be used to improve the dew point forecast in the Sitakunda coast area in southeastern Bangladesh. The model is a synergistic combination of the Convolutional Neural Networks (CNNs) to obtain spatial-level features, Long Short-Term Memory (LSTM) networks to consider the complex temporal dependencies, Residual Networks to overcome the problem of vanishing gradient, and the final multilayer perceptron (MLP) to optimize the dense layer. The proposed model was compared to the traditional time-series (TS) models, namely Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing State Space Model (ETS) and Trigonometric Box-Cox ARMA Trend Seasonal Model (TBATS); machine learning (ML) models, which are Support Vector Regression (SVR), Random Forest Regression (RFR), Extreme Gradient Boosting (XGBoost) and Prophet; seasonally adjusted TS and ML models; and deep learning (DL) models such as CNN, LSTM, MLP and Gated Recurrent Unit (GRU) based on 43 years (1981–2024) of daily data from the Bangladesh Meteorological Department (BMD). The CNN-ResLSTM-ConNet model performed better than other models in predictive ability with a RMSE of 1.063, a MAE of 0.735, a MAPE of 3.650, and a MASE of 0.868. The results show that the hybrid framework can accurately predict both short-term changes and long-term climate patterns and ensures a more accurate forecast which makes it a strong and dependable model for predicting dew point and planning for climate change in coastal Bangladesh.