| Summary: | Polar Mesospheric Summer Echoes (PMSE) are strong coherent radar echoes that occur in the 80 to 90 km altitude range of the mesosphere during the Arctic summer months. These echoes are of significant interest to the space physics community as they provide insight into changes that occur in the atmosphere. To better understand these changes, large datasets of PMSE echoes need to be analysed. In this study, we aimed to develop a deep learning model that could segment PMSE signal data for analysis on larger EISCAT VHF datasets. For the task, we employed a UNet and a UNet++ architecture and tested how pretrained weights from other source domains perform. Next, different loss functions were tested and last the novel object-level augmentation method ObjectAug was employed with other image-level augmentation methods to increase model performance and reduce potential overfitting due to a small training dataset. The results indicate that using randomly initialized weights was the better option for the PMSE target domain and that the use of different loss functions only had a small impact on model performance. When using image- and object-level augmentation the best performing model was reached. It was also seen that there exist inconsistencies in the PMSE signal groundtruth labels. Dividing the inconsistencies into two categories: Granular and Coarse, it was seen that using object-level augmentation had a significantly higher performance on the Granular labelled PMSE signal samples. Overall, our study indicates that the best performing model can be used to segment PMSE for larger datasets or as a supportive tool for further labelling of PMSE signal data.
|
|---|