Center for Planetary Sciences / Department of Astronomy & Astrophysics / Canadian Institute for Theoretical Astrophysics University of Toronto
DeepMoon is a TensorFlow-based pipeline for training a convolutional neural network (CNN) to recognize craters on the Moon, and determine their positions and radii. It is the companion repo to the paper Lunar Crater Identification via Deep Learning, which describes the motivation and development of the code, as well as results.
The DeepMoon pipeline trains a neural net using data derived from a global digital elevation map (DEM) and catalogue of craters. The code is divided into three parts. The first generates a set images of the Moon randomly cropped from the DEM, with corresponding crater positions and radii. The second trains a convnet using this data. The third validates the convnet's predictions.
To first order, our CNN activates regions with high negative gradients, i.e. pixels that decrease in value as you move across the image. Below illustrates two examples of this, the first is a genuine DEM Lunar image from our dataset, the second is a sample image taken from the web.
DeepMoon requires the following packages to function:
- Python version 2.7 or 3.5+
- Cartopy >= 0.14.2. Cartopy itself has a
number of dependencies,
including the GEOS and Proj.4.x libraries. (For Ubuntu systems, these can be
installed through the
libgeos++-dev
andlibproj-dev
packages, respectively.) - h5py >= 2.6.0
- Keras 1.2.2 (documentation); also tested with Keras >= 2.0.2
- Numpy >= 1.12
- OpenCV >= 3.2.0.6
- pandas >= 0.19.1
- Pillow >= 3.1.2
- PyTables >=3.4.2
- TensorFlow 0.10.0rc0, also tested with TensorFlow >= 1.0
This list can also be found in the requirements.txt
.
Our train, validation and test datasets, global DEM, post-processed crater distribution on the test set, best model, and sample output images can be found on Zenodo.
Examples of how to read these data can be found in the
docs/Using Zenodo Data.ipynb
IPython notebook.
We use the LRO-Kaguya merged 59 m/pixel DEM. The DEM was downsampled to 118 m/pixel and converted to 16-bit GeoTiff with the USGS Astrogeology Cloud Processing service, and then rescaled to 8-bit PNG using the GDAL library:
gdal_translate -of PNG -scale -21138 21138 -co worldfile=no
LunarLROLrocKaguya_118mperpix_int16.tif LunarLROLrocKaguya_118mperpix.png
For the ground truth longitude / latitude locations and sizes of craters, we
combine the LROC Craters 5 to 20 km diameter dataset with the
Head et al. 2010 >= 20 km diameter one (alternate download
link). The LROC dataset was converted from ESRI shapefile to .csv.
They can be found under the catalogues
folder of the repo, and have had their
formatting slightly modified to be read into pandas.
During initial testing, we also used the Salamunićcar LU78287GT catalogue.
Each stage of DeepMoon has a corresponding script: run_input_data_gen.py
for
generating input data, run_model_training.py
to build and train the convnet,
and run_get_unique_craters.py
to validate predictions and generate a crater
atlas. User-defined parameters, and instructions on on how to use each script,
can be found in the scripts themselves.
We recommend copying these scripts into a new working directory (and appending this repo to your Python path) instead of modifying them in the repo.
Our model with default parameters was trained on a 16GB Tesla P100 GPU, however 12GB GPUs are more standard. Therefore, our default model may not run on many systems without reducing the batch size, number of filters, etc., which can affect final model convergence.
See docs/Using Zenodo Data.ipynb
for basic examples on generating sample
datasets, loading a pre-trained CNN and using it to make predictions on
samples.
- Ari Silburt - convnet architecture, crater extraction and post-processing silburt
- Charles Zhu - input image generation, data ingestion and post-processing cczhu
Copyright 2018 Ari Silburt, Charles Zhu and contributors.
DeepMoon is free software made available under the MIT License. For details see the LICENSE.md file.