Conference Agenda

Shipping in the Arctic is a huge commercial operation. The presence of icebergs therefore poses a hazard to such operations. Of particular interest are icebergs within sea ice, and a need for automated detection methods. In this work, we utilise a convolutional neural network (CNN) for iceberg detection in fast ice environments. Fast ice is a type of sea ice that forms off of coastlines and remains attached to surrounding land or sea floor. This means that fast ice generally remains in place and is not affected by currents or wind. In Arctic seas, fast ice can extend down to 20 m and has a varied topology depending on environment. Fast ice can be distinguished from drift ice since it does not contain large cracks and fractures. We utilise Sentinel-1 SAR data acquired over Franz-Josef region for this work. Although icebergs show up clearly in optical data, the dependence on conditions such as cloud cover means training with optical data would lead to a less robust program. As such, SAR data is used alongside a Sentinel-2 optical dataset. This SAR data is split by horizontal (HH) and horizontal-vertical (HV) polarisation, with icebergs being clearer in HV polarisation. We also make use of a land mask from the Polar Geospatial Center which helped to aid the training process.

A detection filter used to identify icebergs was proposed by Marino et al (2016). This filter is known as the dual intensity polarisation ratio anomaly detector (iDPolRAD) and has been successfully used in a previous study by Soldal et al (2019) to separate, identify and detect icebergs in sea ice environments. For this work, an iDPolRAD filter is applied to the SAR data to produce training images for a YOLO v8 detection model.

We perform training for 50 epochs with a batch size of 16 and a learning rate of 0.001. For YOLO v8, precision, recall, F1 score and mean average precision (mAP) are used for evaluating the detection performance. Precision measures the ratio between true positives and any new detections (false positives), while recall measures the ratio between true positive and objects the model failed to detect (false negatives). The F1 score acts as a ratio between precision and recall and can be used to determine the optimum confidence score for the algorithm. The mAP score is defined as the total accuracy of the model and is found by taking the area under the Precision-Recall (PR) curve. For model evaluation, we obtained a precision of 0.759, recall of 0.706, F1 score of 0.732 and a mAP of 0.789, giving the model an accuracy of 79%. These results are acceptable for feasible operational use.

The main limitations of this work amount to a lack of an available automated iceberg training dataset, which was addressed by creation of a manual dataset and the continued lack of coverage which can be addressed by future SAR missions (ROSE-L and NISAR). It is hoped that our detection system can be further improved in the future for potential commercialisation.