Vision Transformer-Based Rainfall Detection from Nautical X-Band Radar Data

Due to the sensitivity of the X-band signal to rain, rain may affect the precision of oceanic parameters retrieval from X-band marine radar images. In this paper, a vision transformer (ViT)-based approach is proposed to detect rain in X-band radar data, allowing the recognition of rain-contaminated radar images. Given its ability to capture long-range dependencies in images and model global context effectively, ViT is considered as a promising alternative to convolutional neural networks (CNNs) for radar image classification tasks. Each radar image is first preprocessed and then separated into patches. They are subsequently flattened and embedded to form a sequence that is supplied to the transformer encoder block. Then, the outputs from the transformer encoder are aggregated to obtain the final classification result. The data in this study were acquired using a shipborne Decca marine radar system from south-southeast of Halifax, Canada. The real-time precipitation information was provided by a Non-Acoustic Data Acquisition System (NADAS) installed on the ship. The experiment results demonstrate that the ViT-based approach achieves a relatively superior rainfall recognition precision of 98.5% compared with the support vector machine (SVM)-based approach.