Machine learning-based methods for sea surface rainfall detection from CYGNSS delay-doppler maps

Because of its unique advantages of a short revisit period, cheap observation cost, and high spatial–temporal resolution, GNSS-Reflectometry (GNSS-R) technology has been used successfully in the field of ocean remote sensing. However, there is very limited research on rainfall detection (RD) using this technology. For this purpose, we aim to study the potential of spaceborne GNSS-R in RD Using Delay-Doppler Maps (DDMs) data collected by Cyclone GNSS (CYGNSS) satellites. First, a fast Non-Local Means (NLM) algorithm based on integral images is proposed for DDM denoising to enhance the quality of DDM data. Then, three GNSS-R observables [i.e., DDM average (DDMA), leading edge slope (LES), and trailing edge slope (TES)] derived from power DDM are calculated. In addition, because the RD method based on probability density function (PDF) observables threshold is greatly affected by factors such as geometry and sea state, we have proposed three new methods for spaceborne GNSS-R RD based on Support Vector Machines (SVM), Random Forests (RF) and Convolutional Neural Networks (CNN), respectively. These three methods have significant advantages in establishing multi-parameter models and can provide a strong alternative for spaceborne GNSS-R sea surface RD. To evaluate the RD performance of the proposed methods, the Integrated Multi-satellite Retrievals of Global Precipitation Measurements (GPM-IMERG) is used as reference data. The experimental results show that the proposed three methods are significantly better than the PDF method in precision, recall, and F1 score. Moreover, the RD accuracy of the proposed SVM and RF method is basically the same, while the proposed CNN method is significantly better than the other two methods. Especially compared with the PDF method, the RD accuracy is improved by more than 10%. Generally, when the wind speed is less than 10 m/s, our proposed CNN method can detect rainfall and achieve good detection performance with spaceborne GNSS-R data, which can be better than 78.5%, 83.8%, and 78.1% in precision, recall, and F1 score, respectively.