اسرا نيايش

A Comparative Deep Learning Framework for Climate Variable Prediction in Southeast Iran an‎d the Oman Sea: LSTM, CNN-LSTM, Autoencoder, U-Net+STN, an‎d Attention Mechanisms

Subseasonal-to-seasonal prediction over southeastern Iran an‎d the adjacent Oman Sea is hampered by steep terrain, sharp lan‎d–sea thermal contrasts, an‎d long climate memory. We benchmark five deep-learning models—LSTM, CNN–LSTM, a dilated temporal-convolutional autoencoder (AE-TCN), a U-Net with a Spatial Transformer (U-Net+STN), an‎d a Transformer-style Attention encoder—trained on ERA5 (train/val/test = 1940–2010/2011–2017/2018–Feb 2025) to forecast 1, 3, an‎d 6-month horizons on separate lan‎d an‎d sea grids. Verification uses RMSE/MAE, ACC/R², Taylor synthesis, an‎d event-detection metrics (POD, FAR, CSI). We find a clear lan‎d–sea dichotomy: over complex terrain, U-Net+STN best preserves orographic gradients an‎d yields the lowest centered RMSE for 2-m temperature, while AE-TCN improves short-lead precipitation. Over the open sea, Attention (an‎d CNN–LSTM) maintains anomaly phase at 3–6 months an‎d lowers RMSE by ~20%, increasing CSI through a favorable POD–FAR balance. These results support regime-aware deployment: spatial heads for lan‎d, temporal-attention heads for sea, to provide actionable S2S guidance.

S2S prediction , land–sea contrast , southeastern Iran , Oman Sea , ERA5 , U-Net+STN , Attention , LSTM , AE-TCN , CNN–LSTM , extremes (POD/FAR/CSI)