A Hybrid ConvNeXt, Dual-Transformer, and U-Net Architecture for Direct Satellite-Based Precipitation Estimation from MSG/SEVIRI

Abstract

Accurate quantification of precipitation is critical for anticipating extreme cli mate events such as droughts and flash floods, yet ground rain gauges and weather radars in Northern Algeria suffer from limited spatial coverage and high maintenance costs. Satellite remote sensing via Meteosat Second Generation (MSG/SEVIRI) multispectral imagery offers optimal spatio-temporal coverage, but converting indirect cloud-top properties into direct rainfall rates remains challenging. To overcome the limitations of standard machine learning and conventional convolutional neural networks in modeling long-range atmospheric dependencies, this study proposes a novel end-to-end hybrid deep learning architecture, combining ConvNeXt, Dual-Transformer, and U-Net models, for direct pixel-wise precipitation regression without intermediate classification. The model leverages a ConvNeXt backbone with large-kernel depthwise convolutions for local spatial feature extraction, a Dual-Transformer module utilizing parallel spatial and spectral attention to capture non-local cloud structures, and a U-Net decoder with multi-scale skip connections for high-resolution rainfall reconstruction. Trained and validated against reference target fields constructed from weather radar and rain gauge fusion covering 2006–2012, the proposed ConvNeXt + Dual-Transformer + U-Net framework demonstrates superior performance over traditional machine learning baselines (RF, MLP, SVM) and standard CNN architectures (U-Net, ConvNeXt + U-Net). The model effectively reduces heteroscedastic error dispersion, achieving Pearson correlation coefficients (R) of 0.78 for instantaneous estimates, 0.84 for 3 h accumulations, 0.89 for daily totals, and 0.96 at the weekly scale, alongside lower root mean square errors (RMSE = 12.4 mm/day) and a minimal bias (0.1 mm/week). These results confirm that integrating global attention mechanisms with convolutional backbones enhances the modeling of complex convective rainfall, providing a robust operational tool for continuous satellite-based precipitation retrieval over Northern Algeria.