ITAIC 2020 (ISSN 2693-2865)

Automated Detection of Lunar Craters Using

Deep Learning
2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC) | 978-1-7281-5244-8/20/$31.00 ©2020 IEEE | DOI: 10.1109/ITAIC49862.2020.9339179

Yutong Jia , Gang Wan , Lei Liu , Yitian Wu , Chenyang Zhang

School of Aerospace Information ,Space Engineering University of PLA, Beijing, China
963810890@qq.com, casper_51@163.com, drliu_uavrs@163.com, wuyitian1993@qq.com, 703931023 @qq.com
Corresponding Author: Gang Wan Email: casper_51@163.com

Abstract—In view of the problems of traditional lunar intelligent algorithm that introduces deep learning model
crater detection algorithm (CDA) needs artificial building to extract meteorite crater
crater, morphological characteristics, and the extraction
precision is not high and the retrieval speed slower. In this Traditional meteorite crater recognition algorithm
paper, based on the Image segmentation convolutional network relies on the construction and matching of data quality and
U-Net model, an automatic crater extraction method based on meteorite crater features. Its core idea is to construct more
the improved U-Net model is proposed. The features of craters accurate meteorite crater feature model and more efficient
on the lunar surface were parameterized by deep convolutional template matching algorithm.
neural network, and the residual block and multiple types of
dense skip connection were introduced into the convolutional In 2012, the convolution neural network was put
network in multi-scale sampling to further accelerate the forward, from the image segmentation method based on
convergence speed of the model and improve the detection neural networks became popular, Chen[1] using the
accuracy, thus realizing the intelligent detection of craters on significance test methods to analyze the impact craters in
the lunar surface. Finally, the accuracy of the algorithm is the CCD images of chang a one point of interest and
verified by selecting DEM on the lunar surface and craters significance, each crater points of interest and constitute a
marking data set. The experimental results show that the significant feature vector, using the SVM classifier to
Improved U-Net model can quickly and accurately detected construct impact craters and classification criterion of the
craters on the lunar surface. crater, which extract impact craters in the image. Based on
the U-Net model of image semantic segmentation in deep
Keywords—lunar crater detection; deep learning; residual learning, Silburt[2] proposed the DeepMoon model for the
block; dense skip connection; U-net recognition of craters on the lunar surface, and transferred
I. INTRODUCTION the model to the mercury surface for the recognition of
craters, and achieved good results. Lee[3] developed the
Craters are the most significant geomorphic form on DeepMars model by extending the DeepMoon model
the lunar surface. It is a kind of topographic mark structure to craters on the surface of Mars, enabling rapid
gradually formed by the impact of meteorites on satellites, identification of craters on the surface of Mars. Zheng
large asteroids or other bodies. It has a large number of Lei[4] used convolutional neural network (CNN) in deep
different shapes and presents annular crater structure of learning combined with custom meteorite crater sample
different sizes and uneven aggregation degree. It is of great database to identify large meteorite craters on the lunar
significance to study the evolution of the moon, the history surface.
of meteorite impacts in the solar system and the space
environment. First of all, the study of lunar craters is Compared with traditional meteorite crater
conducive to the in-depth understanding of the identification methods, the intelligent meteorite crater
characteristics of planetary topography and evolution. identification method based on deep learning can better
Second, studying the distribution of crater diameter- solve the problems such as crater radius difference and
frequency can assist in planetary geological and mineral complex shape characteristics. Due to the complex
resource exploration. In addition, studying the material of features of lunar craters, the identification of lunar craters
crater walls can help to explore the presence of water ice. based on U-Net architecture has the problems of low
accuracy rate and slow convergence rate. Therefore, this
Since the beginning of the Lunar and Mars exploration paper takes craters on the surface of deep space objects as
activities in the United States, the accurate and rapid the research object, and combines existing Lunar DEM
identification of impact craters has always been the focus observation products to construct an intelligent
of deep space exploration research. Many researchers have identification framework for craters on the surface of the
also proposed a series of algorithms to extract lunar craters. moon based on improved U-Net. Residual network and
According to the different design ideas of these algorithms, dense connection acceleration model are introduced to
their development can be roughly divided into two improve the detection accuracy. In addition, more
directions: the traditional algorithm that uses image reasonable post-processing operation also makes the
processing technology to identify meteorite crater and the

978-1-7281-5244-8/20/$31.00 ©2020 IEEE 1419

Authorized licensed use limited to: Carleton University. Downloaded on May 27,2021 at 22:54:56 UTC from IEEE Xplore. Restrictions apply.
detection model can effectively reduce false negative III. IMPROVED NETWORK MODEL
errors.
A. Residual Network
II. U-NET NETWORK STRUCTURE Residuals networks generally precede ResNet, which
U-Net is an image segmentation algorithm proposed by is a convolutional neural network proposed by He
Ronneberger[5] in 2015, which was originally mainly used Kaiming[6] when he participated in ILSVRC image
in semantic segmentation of medical images. Compared recognition competition in 2015. Its biggest advantage is
with other deep learning frameworks, U-Net has the its unique residual transmission mechanism. This
advantages of fewer required labeling samples and innovation can well solve the phenomenon of "gradient
constant multi-scale characteristics.In the extraction of disappearance" that may occur when the convolutional
meteorite craters, Silburt proposed the identification network has too many layers in deep learning. It has been
framework of U-Net craters and applied it to the extraction widely used in the deep level network convolution model
of meteorite craters on the lunar surface. Its network and has achieved good target extraction accuracy and
structure is shown as follows: model convergence speed.

Fig. 1. The brief architecture of UNet used in DeepMoon.

(a)Two-tier Resnet (b) Three-layer Resnet

The convolution layer is used to extract the Fig. 2 . Residual block.

characteristics of the input data, ReLu nonlinear activation
function is to ensure network, Dropout layer is on the Fig. 2 is the basis for ResNet34 and ResNet101,
training samples in order to avoid the model fitting, and consisting of the identical connection path and the residual
maximum pooling is further high-dimensional feature path.Left in the residual path by two 3 x3 convolution layer,
extracting, and reduce the number of network parameters, batch normalization and rectified linear units (ReLu)
the sampling is to decode the high-dimensional activation function, and then the addition to output the
characteristics, finally the sigmoid activation function is a results of two paths.In the figure on the right, the
binary classification in order to achieve the craters convolution of the first 1x1 reduces the 256 dimensional
recognition. channel to 64 dimensional, and then recovers through the
The advantage of U-Net lies in that its convolutional convolution of 1x1 at the end, the main purpose is to
neural network uses down-sampling and up-sampling reduce the number of parameters, thus reducing the
internally, which greatly reduces the memory amount of computation.It is worth noting that skip joins do
consumption of model parameters and calculation. In not introduce additional arguments or computational
addition, its cross-hierarchy feature channel further complexity.
preserves the texture information of the original image. B. Dense Skip Connection
However, due to the large difference in the craters'
In[7] introduced the concept of skip join and applied it
geomorphologic features and radii on the lunar surface, the
for the first time In semantic segmentation full convolution
sample data at the poles of planetbodies still have the error
of projection deformation even after the orthographic network. Before long, IQ [5]proposed a U-NHW architecture
correction. Due to the fact that U-Net did not fully consider for semantic segmentation of medical images on the basis
the extraction of overlapping and intersecting meteorite of skip connection. Jumping join helps restore the ability
craters in the post-processing of crater identification, the to achieve full spatial resolution, thus making the full
feature information of crater geomorphic edge may be lost convolution approach suitable for semantic segmentation.
through convolution of crater samples, Moreover, the Jump connections have been further applied to modern
model has too much redundancy in training pictures, neural network structures, such as residual networks and
which leads to the slow and low accuracy of network dense networks, which facilitate gradient flows and
training and was not taken into account during the improve the overall performance of classification
networks.
training.

1420

Authorized licensed use limited to: Carleton University. Downloaded on May 27,2021 at 22:54:56 UTC from IEEE Xplore. Restrictions apply.
 As shown in Fig. 3, the improvement of U - Net disc area and the background. p k ,i  > 0,1@ , refer to the
network, not only increases the new short jump connection
(orange line), also contains the original U - Net in the long probability that the pixel point is predicted as A kind, that
jump connection part (blue dotted line), a new generation is, the probability value output by the last Softmax layer of
of dense jump links can closely transfer characteristic the network. g k ,i  ^0,1` , refer to the tag value of type K
information in multiple residual module, make the crater of pixel point i.
characteristic information more accurate and fast
convergence. C. Evaluation Index
Adam optimization with back-propagation are used
for training with the batch size of a single image and
learning rate of 0.00001. The trained network is then
applied for crater detection on 100 test images. The
common recall and precision rates calculated as follows
are used to evaluate the carter detection performance.
Tp
Precission=
Fig. 3. Improved network structure Tp  Fp
(1)
It should be pointed out that the padding operation is Tp
adopted in the experiment, because the input data is the Recall
Tp  Fn
same size as the output data except the depth after
convolution operation. In addition, this experiment also where T p , F p , and Fn represent the number of true
introduces the learning rate to solve the problem of
network convergence. positive, false positive, and false negatives respectively.
D. Experimental Results
IV. LUNAR CRATER EXTRACTION EXPERIMENT
In this experiment, the experimental results of
A. Experimental Data relatively dense meteorite craters with overlapping
Lunar Sphere observations include DEM data products meteorite craters were selected for demonstration. We
based on lunar observations by the Lunar Reconnaissance have employed our network to detect craters on test-
Orbiter (LRO) and Kaguya, and lunar craters annotated images.The final experimental results are shown in Fig. 5
with Povilaitis[12] and Head naissance Orbiter Lunar and Fig. 6.
DEM product data is fixed in size, and square image input
is generally used for deep learning in image segmentation
applications[14].

(a) Moon DEM Image (b) Ground-Truth Target Mask

Fig. 4. Sample of Datast

B. Loss Function
Semantic segmentation is still the problem of pixel
classification, and the most commonly used loss function
is the cross entropy function. In this paper, the DicLoss
function is used to replace the traditional cross entropy loss
function˖
N

K
2wk ¦ p k , i g k , i
LDice =1-¦ N
i 1
N
(2)
(c)Unet Model Predict (d)Improve Unet Model Predict
k 1
¦p
i 1
k,i  ¦ g k,i
i 1
Fig. 5 Crater Detection Result Comparison on Moon region I

It can be seen from Fig. 5 that the original UNet model

Where, N refer to the number of pixels. K refer to the is able to identify individual meteorite craters well, but
number of categories and is set as 2, including the optic partially overlapping and intersecting meteorite craters
and larger meteorite craters cannot be identified. However,

1421

Authorized licensed use limited to: Carleton University. Downloaded on May 27,2021 at 22:54:56 UTC from IEEE Xplore. Restrictions apply.
the improved model in this paper introduces difference respectively on the crater data set, and the loss cost is
network and jump connection, which can identify this type 0.0907 and 0.0213 respectively. It should be pointed out
of meteorite craters. that the improved U-NHW model converges faster from the
improvement of accuracy.

TABLE 1. COMPARISON OF MODEL TEST RESULTS

UNet Improve UNet

Iterations
P Loss P Loss
1000 0.579 0.1842 0.591 0.2561
2000 0.621 0.1652 0.652 0.1926
3000 0.786 0.1501 0.798 0.1663
4000 0.851 0.1302 0.882 0.1356
5000 0.894 0.0922 0.931 0.0421
(a) Moon DEM Image (b) Ground-Truth Target Mask
V.CONCLUSIONS
The identification and extraction of craters on the lunar
surface is an important research field in the field of lunar
exploration, and deep learning, with its superior image
processing and analysis capabilities, has become the
mainstream in the field of image processing and analysis
of various disciplines. Based on existing Lunar DEM data,
combined with deep learning image semantic
segmentation U-Net model and ResNet framework, this
paper proposes an improved U-Net automatic crater
(c)Unet Model Predict (d)Improve Unet Model Predict extraction model on the lunar surface, and greatly reduces
Fig. 6 Crater Detection Result Comparison on Moon region II. the false positive of detection results through improved
recognition post-processing technology. The experimental
The upper left (a) in the figure shows that the original results show that the improved U-Net model can better
DEM data is linearly converted into 8-bit grayscale image. identify lunar craters and their exact locations, and has a
The figure (b) on the right is a data set of meteorite craters faster convergence rate and better mobility than the U-Net
in the corresponding region. This data set represents model, which can be used for the automatic extraction of
meteorite craters in the form of the longitude and latitude lunar craters and provide support for the landing of lunar
of the crater center and the radius of the crater, so the probes. The follow-up work will consider integrating
figure is shown as a circle. The blue circle in the lower left multi-source deep-space detection data, such as synthetic
figure (c) is the result of moon crater extraction using aperture radar (SAR) observation data and high-resolution
UNetAnd the red circle in (d) on the right for the optical image observation data, etc., and combining with
improved UNet model in this paper, the results of crater more detailed meteorite crater manual annotation data and
extraction. edge detection algorithm, etc. to further improve the
Model after the crater model identification in the post- identification rate and extraction speed of meteorite
treatment process, this part is mainly based on threshold surface meteorite crater.
to remove some of the smaller craters crater radius, and REFERENCES
also by computing the center of craters overlap distance
[1] Cheng Y , Johnson A E , Matthies L H , et al. Optical landmark
and the interpolation of the sum of the radius of each other, detection for spacecraft navigation[C]// Aas/aiaa Space Flight
and then judge is crater contains or intersect relations, Mechanics Meeting. 2003.
scope of the intersection of smaller craters shall be [2] Ari S , Mohamad A D , Chenchong Z , et al. Lunar Crater
preserved and intersecting range larger crater is likely Identification via Deep Learning[J]. Icarus, 2018:S001910351830
same area after repeated sampling to identify repeated as a 1386-.
result, should be deleted. [3] Lee C . Automated crater detection on Mars using deep learning[J].
2019.
It can be seen from Fig. 6 that the original UNet can
[4] Zheng Lei, Hu Weiduo, Liu Chang. Large Crater Identification
well identify individual meteorite craters, but the partially Method Based On Deep Learning [J].Journal of Beijing University
overlapping and intersecting meteorite craters are deleted of Aeronautics and Astronautics, 2020, 046(005):994-1004.
in the post-processing, and the false negative error is large. [5] Ronneberger O , Fischer P , Brox T . U-Net: Convolutional
However, the improved model proposed in this paper still Networks for Biomedical Image Segmentation[J]. 2015.
achieves good recognition results in the area with dense [6] He K , Zhang X , Ren S , et al. Deep Residual Learning for Image
craters. Recognition[C]// IEEE Conference on Computer Vision & Pattern
Recognition. IEEE Computer Society, 2016.
The experimental results show that after 5000 [7] Evan Shelhamer, Jonathan Long, Trevor Darrell. Fully
iterations of training, the UNet model and the improved Convolutional Networks for Semantic Segmentation[M]. IEEE
UNet model reach the accuracy rate of 89.4% and 93.4% Computer Society, 2017.

1422

Authorized licensed use limited to: Carleton University. Downloaded on May 27,2021 at 22:54:56 UTC from IEEE Xplore. Restrictions apply.
[8] Jin Y , He F , Liu S , et al. Small Scale Crater Detection based on
Deep Learning with Multi-Temporal Samples of High-Resolution
Images[C]// 2019 10th International Workshop on the Analysis of
Multitemporal Remote Sensing Images (MultiTemp). 2019.
[9] Ali-Dib M , Menou K , Jackson A P , et al. Automated crater shape
retrieval using weakly-supervised deep learning[J]. Icarus, 2020,
345:113749.
[10] Meng D , Yunfeng C , Qingxian W . Method of Passive Image
Based Crater Autonomous Detection[J]. Chinese Journal of
Aeronautics, 2009, 22(3):301-306.
[11] Joseph Paul Cohen, Wei Ding. Crater detection via genetic search
methods to reduce image features[J]. Advances in Space Research,
2014.
[12] Povilaitis R Z , Robinson M S , Bogert C H V D , et al. Regional
Resurfacing, Secondary Crater Populations, and Crater Saturation
Equilibrium on the Moon[C]// Lunar & Planetary Science
Conference. Lunar and Planetary Science Conference, 2017.
[13] Iii J W H , Fassett R I , Kadish R J , et al. Global Distribution of
Large Lunar Craters: Implications for Resurfacing and Impactor
Populations[J]. ence, 2010, 329(5998):1504-1507.
[14] LOLA Team, Kaguya Team, 2015. LRO LOLA and Kaguya
Terrain Camera DEM merge 60N60S 512ppd (59m).
https://astrogeology.usgs.gov/search/map/Moon/LRO/LOLA/Lun
ar_LRO_LrocKaguya_DEMmerge_60N60S_512ppd.

1423

Authorized licensed use limited to: Carleton University. Downloaded on May 27,2021 at 22:54:56 UTC from IEEE Xplore. Restrictions apply.