Automatic detection of lunar sub-km craters via deep learning
Riho Ito*, Ryosuke Nakamura*, Chikatoshi Honda†
*National Institute of Advanced Industrial Science and Technology (AIST), †The University of Aizu
Crater detection
High accuracy impact crater
databases are needed for estimating
generated age of the body’s surface
with crater chronology.
• Crater chronology is a method for
estimating generated age of surface from
crater size-frequencies.
• Accuracy of crater chronology depends
on accuracy of crater information
(geoinformation and diameters).
Resolution of exploration data were
improved. Sub-km craters can be
found from these high-resolution
imageries.
•
Resolution of SELENE Terrain Camera
Ortho (TCO) imagery is about 7.4
m/pixel.
SELENE’s Terrain Camera (TC) Ortho imagery:
Divided a image into 3 regions.
(2 regions: For learning, 1 region: For evaluation)
Crater Database:
290~1120m in diameter craters. (measured manually)
North
2) Resize all patches to same size.
p=15
p=25
98
p: patch size
96
CNN classified that
there is a crater at
center of a patch or
not.
In results of both
patterns, F-measure is
over 95 %.
94
92
90
%
Recall
Precision
F-measure
Result of crater detection from TC ortho imagery
3) Augmentation (rotation)
Output：15pixel
Feature of predicted craters:
High contrast or shadow regions.
●Crater
In result of p=25 model, miss labels
were decreased than p=15 model.
➡ Context information is
important for distinguishing
non-craters from craters.
Learning
Input
Dataset
Probability
crater
CNN
Output：25pixel
0.7
non-crater 0.3
Miss Labels
1 km
1 km
1 km
©JAXA/SELENE
Multiscale craters detected by changing resolution of input images. (p=25 model)
Detection
Input
Resize
Input image
49.3m/pix
24.7m/pix
16.4m/pix
12.3m/pix
9.9m/pix
Manually detecting sub-km crater take an
immense amount of time because number
of crater is increase exponentially with
decreasing diameter.
Output
24.7m/pix
Probability image
CNN
2 km
crater
2 km
©JAXA/SELENE
non-crater
Truth
Output
12.3m/pix
3. Experiments
2 km
Fully Connected
Convolution
Convolution
Convolution
Network architecture Datasets
Input
• By designing the neural network
architecture that has more multi layers,
deep learning can represent the data
more abstract.
• From supervised data, deep learning can
learn optimal features that are needed for
feature extraction process and data
identification process.
• Especially, Convolutional Neural Network
(CNN) keeps high performance in the
field of computer vision.
• [1] suggest a possible beneficial effect of
CNN at martian craters classification
task.
Accuracy of all regions
TCO_MAP_02_N24E333N21E336SC
To develop algorithms for detecting
craters automatically is important.
Deep Learning
label
Center
©JAXA/SELENE
Negative
Positive
True Negative False positive
Negative
(TN)
(FP)
False Negative True Positive
Positive
(FN)
(TP)
TP
TP
Recall =
Precision =
TP + FN
TP + FP
2 ∗ Precision ∗ Recall
Fmeasure =
Precision + Recall
Cropping size vary
with crater diameter.
South
100
prediction
1) Crop patches.
Out
To detect sub-km craters from
high-resolution lunar imagery by
using deep learning.
Result of patch based classification
Supervised data
Softmax
Purpose
4. Results
2. Method
Fully Connected
1. Introduction
Patch size
Prepared 2 patterns patches.
Training
South,
Center
North,
South
Test
North
Center
25pixel
15pixel
Comparing
results
North,
Center
South
label
noncrater
crater
noncrater
crater
noncrater
crater
Training Test
Patches Patches
3204
411
3204
411
3000
462
3000
462
3492
339
3492
339
2 km
5. Conclusion
• In patch based classification task, recall, precision, Fmeasure were over 90%.
• In detection task, CNN detected multiscale crater position.
However, there were any miss labels.
• For applying crater chronology, crater diameter is
important as well as position. So, we have to improve the
method for estimating crater diameter.
6. Future works
Considering supervised data.
• Suitable patch size and context area.
• Add other lunar data (e.g. DTM).
Considering network architectures.
• Separate a network to 2 processes.
1) Detect position of craters
2) Estimate crater diameter.
Comparing other algorithm [2].
[1] J. P. Cohen, H. Z. Lo, T. Lu and W. Ding, “Crater detection via convolutional neural networks,” 47th Lunar and Planetary Science Conference, Mar 2016.
[2] S. Yamamoto, T. Matsunaga, R. Nakamura, Y. Sekine, N. Hirata and Y. Yamaguchi, “Rotational Pixel Swapping Method for Detection of Circular Features in Binary Images,”
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 53, NO. 2, Feb 2015.
E-mail:[email protected]