Ocean swell height estimation from spaceborne GNSS-R data using hybrid deep learning model

Global navigation satellite system reflectometry (GNSS-R) has emerged as a pivotal remote sensing (RS) technology, widely utilized for retrieving crucial oceanic parameters such as wind speed, sea surface height, and sea ice detection. However, the retrieval of ocean swell height remains an underexplored area within this domain. The complexity of constructing multivariate regression models for swell height retrieval poses a significant challenge, particularly in contrast to existing empirical models. For this purpose, this article proposes a novel deep learning (DL) hybrid model, namely Multi-scale Conv-BiLSTM, which combines multi-scale convolution and bidirectional long short-term memory (BiLSTM) networks for the first time to retrieve ocean swell height using spaceborne GNSS-R data. This innovative hybrid model comprises three fundamental modules: a multi-scale feature extraction module, a feature relationship inference module based on BiLSTM network, and a manual extraction of multiple feature parameters module encompassing GNSS-R variable and auxiliary variable. Specifically, the multi-scale feature extraction module leverages deep convolutional neural network (DCNN) to extract spatial features surrounding the specular reflection point (SP) from the two-dimensional matrix of the bistatic radar scattering cross-section (BRCS) image and effective scattering area. Subsequently, the feature relationship inference module employs the BiLSTM network to engage in the inference process between feature relationships. This module excels in considering critical information associated with temporal characteristics, effectively capturing preceding and subsequent information. Validation was conducted using ERA5 and WaveWatch III (WW3) data by comparingthe proposed Multi-scale Conv-BiLSTM model against seven traditional machine learning (ML) models, including support vector machine (SVM), decision tree (DTR), light gradient boosting machine (Lightgbm), eXtreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), gradient boosting decision tree (GBDT) and DCNN. The results show that when ERA5 is used as reference data, the proposed Multi-scale Conv-BiLSTM model achieves a reduction in root mean square error (RMSE) by 23.67%, 28.63%, 9.77%, 8.91%, 28.50%, 16.46%, and 15.05% compared to the SVM, DTR, Lightgbm, XGBoost, AdaBoost, GBDT, and DCNN models, respectively. When WW3 is used as reference data, the proposed Multi-scale Conv-BiLSTM model exhibits an improvement in RMSE by 35.99%, 36.62%, 25.49%, 24.74%, 41.37%, 30.82%, and 29.61% compared to the SVM, DTR, Lightgbm, XGBoost, AdaBoost, GBDT, and DCNN models, respectively.