Optimized Use of Engineered Features from GNSS Reflectometry for Prediction of Significant Wave Heights

GNSS interferometric reflectometry is a well-established technique in Ocean Remote Sensing that can be used for the retrieval of sea surface characteristics. In particular, the evaluation of interference patterns in GNSS signal-to-noise ratio (SNR) observations allows for an estimation of the Significant Wave Height (SWH), e.g. by relating the SWH to the attenuation that is typically present in the oscillating interference pattern for increasing elevation angle of the signal-emitting GNSS satellite.

Recently, we developed new machine learning methods for the analysis of GNSS SNR observation data obtained from the research platform FINO 2 in the Baltic Sea. The core element thereof is the extraction of various engineered features from SNR interference patterns  by means of kernel regression and clustering techniques. The various engineered features were used as input for the prediction of the SWH with supervised machine learning models (artificial neural networks, bagged regression trees, linear models). In a case study, these predictions provided a remarkable improvement in accuracy compared to predictions which solely use a common feature stemming from the aforementioned attenuation in the SNR interference pattern.  

However, an optimized extraction of information from the various and partially redundant engineered features for the prediction task is desirable, aiming at the reduction of model complexity without reducing predictive performance. This goal is successfully addressed in the present work by applying a forward selection scheme and a principal component analysis for the set of available  engineered features. The usage of the engineered features can also be optimized by tuning the hyperparameters of complex supervised machine learning models used for the SWH prediction. Such a tuning is performd by means of a grid search for a random forest model applied to the engineered features. This optimization represents an advancement of the application of the bagged regression trees with an improvement in accuracy of the respective SWH predictions.

The improved methods for SWH prediction at FINO2 are outlined and the impact of the involved optimizations concerning the use of the engineered features is evaluated and discussed in detail in a case study.