The periphery of the polar caps is characterized by steep scarps. These were observed to undergo active erosion, involving avalanches and ice block falls. We wish to quantify the block fall events over space and time and identify the source areas from where the ice blocks originate. Repeat High Resolution Imaging Science Experiment (HiRISE) observations with image scales of up to 0.25 m/pixel enable an identification of fresh fallen blocks at the bottom of the scarps. However, a single HiRISE image covers very large areas and can have a size of several Gigabytes, which requires an automated approach.
We have developed a deep learning-driven change detection model to automatically detect the detached ice block from the steep scarps. The convolutional neural network simultaneously analyzes both the pre-detach and post-detach images. The supervised method requires over 10,000 co-registered image pairs for training, validation and testing. Thereafter, we can efficiently apply the well-trained model to multi-temporal HiRISE images for detecting block fall activity. The good performance and fast processing speed of our method allows us to monitor the entire scarp area, which is an important step in estimating the ongoing mass wasting and studying the evolution of the Martian polar scarps.
Contact: Shu Su