Mars Crater Detection

Performing semantic segmentation on Martian satellite images to identify craters along with their classes.

Created Using: Python TensorFlow Keras OpenCV

I can say that this project has been the highlight of my masters degree. I published a research paper on this project and got to present it at the ACM SIGSPATIAL conference in Hamburg, Germany.

The code for this project is available at: github.com/HitanshShah/mars-crater-detection. Reach out to me via email if you have any queries related to this project.

Crater Detection and Crater Counting have historically been manual and tedious tasks. With the recent advancements in the field of Deep Learning and Computer Vision, and with the availability of higher computation power, it has been possible to automate this process. Almost all of the recent studies have used some or the other variation of a U-Net architecture for this task. We aimed to outperform these former State-of-the-Art methods by using a U2-Net architecture, which in simple words, is a U-Net made up of U-Nets at all of its levels. We were able to achieve our goal, however, this was a binary semantic segmentation, i.e., the output only said if a particular part of the image was a crater or not a crater.

Left: Binary Crater Segmentation using Edges of Craters as ground truth annotations. Right: Binary Crater Segmentation using Full Craters as ground truth annotations.

However, it must be noted that not all craters are of the same nature. Some craters millions and millions of years old, whereas some are relatively new. Some craters, being so old, might even be filled or covered by a fresher layer, which in turn could be from different sources, like volcanic sedimenation, settling of dust, and so on. Some craters might have been formed as a result of ejecta material from another primary impact crater. These factors become relevant when crater counting is used for tasks like Surface Age Estimation. Thus, it is important to not just identify craters, but also get some insights related to the nature of the crater. There had been no research done on this till date and my team and I were the first ones to come up with a proper pipeline for the same. We collaborated with our professors Dr. Hannah Kerner and Dr. Jacob Adler to get more domain specific knowledge and to convert this from a class project to a research project.

Pipeline of Binary and Multi-class semantic segmentation of Martian satellite images.
Multi-class Segmentation using Edges of Craters as ground truth annotations.

All of our research and model training was done on the THEMIS dataset. Even though consistent with all other related research, this dataset is relatively old and of lower resolution. The newer CTX dataset provides a much higher resolution (20 times) than THEMIS and should be the base of all future studies. However, since it is such a high quality dataset, it is computationally inefficient (and close to impossible) to train Deep Learning models of similar complexity purely on this dataset at the same scale. To prove that our model, even though trained on a lower quality dataset, was capable of segmenting the higher resolution CTX images, we performed 2 out-of-domain evaluations.

Pipeline for Out-of-Domain evaluation of Binary Crater Segmentation on CTX and DoMars16k datasets.

For detailed results, discussions and other information related to project, pls read my paper here. I am currently working on improving the performance of the multi-class segmentation by using various data augmentation techniques and balanced loss functions catered specifically towards imbalanced datasets.