Bachelor Thesis

Student: Gloria Vergés
Supervisors: Dr. Lanqing Huang and
Prof. Irena Hajnsek
Time period: 2023 Spring Semester

The aim of this bachelor thesis was to investigate whether information (features) derived from multi-frequency SAR imagery could enhance sea ice classification using a random forest classifier (RFC). The "Ice Chart" product (U.S. National Ice Center, 2020) served as ground truth to label the data. The question was researched by using Sentinel-1 (C-band) and TanDEM-X SAR data (X-band). The Sentinel-1 and Tan-DEM-X acquisitions overlapped in three distinct sectors within the regions of the Amundsen and Ross Sea in Antarctica, with a temporal gap of approximately 30 minutes. Sixteen X-band features and three C-band features were retrieved and categorized as backscattering, interferometric, polarimetric and topographic based features. The data processing involved image rescaling, merging the high resolution features, as well as co-locating the features with the ground truth and labelling each pixel accordingly. The results shows that the X-band features based model scored a significantly better accuracy value (0.895) than the model based solely on the C-band features (0.784). Two out of the three C-band features were ranked under the top 10 most important features, indicating that C-band features are important.

Nevertheless, the inclusion of C-band features lowered the accuracy score, which suggest that the inclusion of C-band features does not improve prediction accuracy despite their importance.

Student: Robin Gisin
Supervisor: Dr. Lanqing Huang and Prof. Irena Hajnsek
Time period: 2023 Spring Semester

This thesis studies the potential of using machine learning models with multi-frequency synthetic aperture radar (SAR) imagery to estimate the elevation of sea ice. Random Forest Regression, a widely used machine learning model, is applied to predict the ice freeboard and snow freeboard using X-band data from the TanDEM-X mission and C-band data from the Sentinel-1 mission. The study uses data on the ice freeboard obtained from the CryoSat mission as the target variable, which is spatially and temporally filtered for alignment with the SAR data. Before applying the Random Forest Regression the two target variables are compared. The natural relationship that the snow freeboard tends to be greater because of the additional snow can be confirmed. The results from the model for the ice freeboard are not very significant, as insufficient data leads to a dependency on the randomness. Also, the RMSE is approximately 50% of the average ice freeboard. For the snow freeboard a clear positive correlation of 0.76 for the mutli-frequency input and a RMSE of 20% is achieved. The model shows that the combination of multi-frequency SAR imagery can improve the result for predicting the snow freeboard, as one of the top three features is the backscattering intensity HH from C-band.