Wiki for Crater Annotation in DeepGIS

Welcome to the DeepGIS crater wiki!

The Mars and Moon crater annotation DeepGIS tool is accessible through the following webpages.

• Mars DeepGIS
• Moon DeepGIS

Crater Annotation: General guidelines followed for labeling craters is simple: annotate all visible craters.

After creating a user account in DeepGIS, a user is provided with an image labeling page consisting of selected images from LROC NAC and Mars CTX images. There are a total of 1015 LROC NAC images stored in the Moon DeepGIS database and 505 Mars CTX images in the Mars DeepGIS database.

Please contact us, if you need access to admin/expert privileges.

Figure 1. Annotation page (upper half) of DeepGIS. The landing page provides a description of the instructions to annotate craters in mars and lunar images.

Figure 2. Annotation page (lower half) of DeepGIS. The lower half shows the downloadable image link. The B01_009866_1747_XN_05S358W in the image path represents the image id of Mars CTX image and tile_5120_40960.png follows the naming convention from how image tiles were created.

Figure 3. Zoom feature in the annotation page of DeepGIS. This figure shows a zoomed image of the crater from previous mars image. The zoom feature enables easy annotation of the smaller craters present in the image.

Right after login on the left hand side of the screen there will be categories to annotate: choose crater (default option in mars and moon DeepGIS). There are currently three annotation tools on the DeepGIS: circle, ellipse, and bezier as shown in the figure below:

Figure 4. Annotations Tools. Annotation tools on the left hand side of of the screen directly after logging into the DeepGIS website.

The “Circle” tool is used by selecting “Circle” under Annotation tools and then clicking and dragging circles that most accurately describes craters. The “Ellipse” tool acts in a very similar manner except there is an additional degree of freedom in the vertical direction; meaning the ellipse can be either shrunk or enlarged to fit the rim of the crater. The “Bezier” annotation tool is different in the sense that one has to use their cursor to click tiny dots around the crater’s rim until one has gone all the way around the crater. Only after connecting the final dot to the beginning dot does this complete the annotation process.

Figure 5. Annotation using bezier curves. Annotation of a zoomed in crater using “Bezier” tool in DeepGIS resulting in polygonal mask over the found crater.

The “MaskRCNN” button underneath Deep Learning Zoo option on the left enables users to select and download the dataset.

Figure 6. Show all annotations. This page shows the existing annotations on the mars images.

DeepGIS also enables its admin and staff users to edit and correct annotations made by other users through “Display and Edit” button. This allows us to examine and validate crater annotations through an expert evaluation process.

Figure 7. Display and edit existing annotations. This page shows the existing annotations on the mars images as gif file with and without annotations. The “edit” option will load the existing annotations overlaid on the same image and an expert can correct those annotations.

At the bottom of this page, there is an option to create a direct download link of the annotations. These annotations can be provided as an input to Google Colaboratory Notebook to train and predict craters. The direct link can be used an input for the notebook or the annotations can be uploaded to run the training. [https://github.com/DREAMS-lab/DeepGIS_deeplearning_zoo ](Jupyter Notebook Code)

Figure 8. Direct link to generate datasets. A dataset consisting of both image and its corresponding annotation can be generated in DeepGIS through the “Generate Dataset” button. The procedure follows two simple steps select the interested annotations and click generate dataset. This link can be directly to Google Colaboratory notebook, which can train a MaskRCNN model to predict the masks and categories in the images.

Figure 9. Google Colab notebooks. The open sourced training and inference colab MaskRCNN notebook for datasets generated through DeepGIS.

Additionally, we also provide an option “Predict using AI” that is can be used in future applications. This feature enables its users to correct and remove labels than draw labels from scratch. The craters with confidence above 98% is currently shown on the images. A present in the DeepGIS can be trained on newer annotations and result in better predictions. This feature is used by researchers in geological applications especially to detect rocks and plants in DeepGIS (flow.deepgis.org).

Figure 10. Predict using AI. The model trained in Google Colab can be uploaded and used to predict labels in DeepGIS. The image above shows 2 labels generated by Mask RCNN with more than 98% confidence.