Collins, MichaelYekkehkhany, Bahareh2022-10-042022-10-042022-09-19Yekkehkhany, B. (2022). Sea wind vector estimation using C-band full-polarimetric SAR data (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.http://hdl.handle.net/1880/115344https://dx.doi.org/10.11575/PRISM/40344This research used neural network (NN) and random forests (RF) models to estimate sea wind speed and direction using synthetic aperture radar (SAR) data. We used RADARSAT (RS)-2 C-band single look complex (SLC) fine quad-polarimetric data and buoy measurements. After data preparation, SAR data are paired with their relevant buoy observation. Then, SAR data parameters expected to be impacted by sea wind are generated. While the spatial resolution of our RS-2 data is 4.7 × 5.1 m, we cropped each image to a chip of 512 × 512 pixels centred by its related buoy and averaged the parameters over the image chip. Therefore, this study’s estimated wind speed and direction resolution is approximately 2.45 × 2.65km. Then, parameters are separated to train and test data by repeated k-fold cross-validation (CV). Also, successive halving random search CV is used to tune NN and RF hyperparameters. To estimate wind speed, least absolute shrinkage and selection operator (LASSO) feature selection determined the HV polarization normalized radar cross section (NRCS) (σ0H V ) and the real part of the correlation coefficient between HV and VH polarization channels (ℜ(ρHV V H)) as models inputs. The bias, root mean square error (RMSE), and correlation coefficient (CC) between the buoy measured and estimated wind speed by NN are 0.08 m/s, 1.96 m/s, and 0.81, and by RF are 0.01 m/s, 1.94 m/s, and 0.82, respectively. The machine learning models are given all SAR parameters as their inputs in estimating wind direction. First, some data bucketing of wind speed bins of 5 m/s and incidence angle bins of 11◦ is done. Finally, the models’ evaluations are based on their performance on each data bucket and the whole dataset by calculating the weighted average of all the data buckets. Then, the bias, RMSE, and CC between the buoy measured and estimated wind direction by NN are −0.69◦, 31.25◦, and 0.58, and by RF are −0.03◦, 25.73◦, and 0.77, respectively. Finally, a permutation feature importance is applied to the trained wind direction models. The imaginary parts of the correlation coefficient between cross- and co-polarization channels, ℑ(ρHHHV ), ℑ(ρHHV H), ℑ(ρV V HV ), and ℑ(ρV V V H), play significant roles in building both NN and RF models.engUniversity of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.Earth ObservationSynthetic Aperture RadarMachine LearningNeural NetworkRandom ForestsPolarimetric ParametersSatellite ImageryWind EstimationMarine ApplicationOceanographyRemote SensingArtificial IntelligenceEngineering--EnvironmentalEngineering--Marine and Oceandoctoral thesis