Improving rainfall retrieval accuracy using cross-Modal deep learning: Merging wifi with commercial microwave link

Abstract

Commercial microwave links (CMLs) have attracted increasing attention in meteorology and hydrology as a cost-effective method for rainfall retrieval. Nevertheless, the typical single-link and high-frequency structure of CMLs often limits the accuracy of rainfall retrieval. This study proposes a cross-modal deep learning model that merges channel state information (CSI) from WiFi with received signal level (RSL) from CMLs for high-precision rainfall retrieval. First, a fast dynamic time warping algorithm is applied to align the RSL and CSI data. The CSI data are then denoised and filtered using wavelet denoising and Savitzky-Golay filtering. At the same time, the RSL data undergo feature engineering, including gradient extraction, rolling statistics, and wavelet-based features. A gated cross-modal attention mechanism is then utilized for feature-level fusion, and a dynamic weighted network adaptively adjusts the contribution of each modality. Finally, a decision-level fusion module with consistency regularization generates the final rainfall intensity estimates. Experimental results demonstrate that the proposed model achieves a pearson correlation coefficient of 0.9541, a root mean square error of 0.0271 mm/h, a coefficient of variation of 1.8948, and a relative bias of -30.4%, confirming its effectiveness in improving rainfall retrieval accuracy through WiFi-CML data fusion.