Training Dataset for the Machine Learning Approach in Glacier Monitoring Applying Sar Data

The study analysed the possibility of utilizing machine learning to determine glacier displacements. The obtained results were compared with the offset-tracking method available in SNAP software. The Jakobshavn glacier in Greenland served as the test site. Analyses were carried out using Sentinel-1 data during the period of August 1 to August 7, 2021. To generate a dataset for the selected part of the glacier, a synthetic training dataset comprising 4,500 samples was created. It was constructed by applying rotation in the range of ±30º and resizing within the range of ±10-20 pixels to the original patch. The final neural network (NN) consisted of 7 layers. The maximum displacement value is 250 m, corresponding to a velocity of 41 m/day. Notably, these maximum values are consistent with the results from offset-tracking. Nevertheless, the results in the slowly moving areas are not reliable because of the coarse resolution of the NN output.