Deep Learning based Bias Correction of Forecast Products with Fusion of Geographical Information

Accurate numerical weather prediction is a prerequisite for disaster prevention and mitigation. Based on the numerical ocean-atmosphere-wave coupling forecasting model developed by our team, artificial intelligence technology is introduced to correct prediction bias, addressing the systematic forecast error inherent in numerical models, and further enhancing the accuracy and reliability of our forecast products. This study aims to establish a bias correction model for the numerical forecast products and integrate multi-source geographic information, such as elevation, land cover, and soil type, to improve the forecast results. Convolutional Neural Networks (CNNs) effectively extract spatial features through the computational mechanisms of their convolutional modules, making them well-suited for tasks like meteorological forecast correction, where spatial correlations have a significant impact. We employ a Residual Convolutional Neural Network to efficiently extract the spatial features from numerical model forecast results, leading to improved correction performance compared to traditional correction methods. A spatiotemporal feature extraction module is utilized to adaptively detect terrain and surface features at different scales, addressing the extraction and fusion of multi-source heterogeneous features. We further evaluate and optimize the impact of terrain and land surface features at different resolutions on model correction effectiveness, maximizing prediction accuracy. This research will provide strong technical support for enhancing the precision of ocean-atmosphere-wave coupling forecasting products and offer reliable data assurance for disaster prevention and mitigation efforts.