Skip to content

Latest commit

 

History

History
51 lines (38 loc) · 2.91 KB

README.md

File metadata and controls

51 lines (38 loc) · 2.91 KB
Raw

Lunar impact craters ages estimation

Requirements:

Tested on

  • python 3.6
  • pytorch 0.4.0
  • pretrainedmodels 0.7.4

Data preparing

image data preparing

For image data, the DOM images (gray images) of the aged craters and the identified craters are resized to 256*256.

attribute data preparing

For attribute data, total 78 attribute data, morphological features with the corresponding stratigraphic information,of craters were extracted for age estimation model. 40 morphological features of craters were calculated with Chang’E data referring to the lunar impact craters database published by LPI[1]; 38 stratigraphic attributes of craters were extracted from the 1:5,000,000 Lunar Geologic Renovation (2013 edition) produced by the U.S. Geological Survey. What needs illustration is that the value of 38 stratigraphic attributes refer to the percent of the stratum in the crater. The details of morphological features and the stratigraphic attributes can be found in feature_description.csv. What needs illustration is that the first 40 features indicate the morphological attribute and the last 38 features indicate thr stratigraphic attribute. The morphological features refer to the lunar impact craters database published by LPI[1], and some age-independent features (like, Latitude, Longitude, et. al.), repetitive features (like, Diameter [km] and Radius [km], Radius [m]), and some incomplete features are removed. The stratigraphic attributes refer to the 1:5,000,000 Lunar Geologic Renovation (2013 edition) produced by the U.S. Geological Survey.

Experiment detail

We use Adam optimizer with learning rate schedule. The inital learning rate is decayed by the factor of 0.2 at 4/5 of entire epochs. The basic parameters are listed below:

  • Learning rate = 0.0003
  • Epochs = 10
  • Batch size = 32
  • Weights regularization = 0.0001

NOTE: 12 CNN models are optional [Resnet50, Resnet101, Resnet152, Senet, se_Resnet50, se_Resnet101, se_Resnet152, se_Resnext101, Polynet, Inceptionv3, DPN68b and Densenet201].

To Run

Training

To train age estimation model with Resnet50.

python train_moon_age_estimation.py -a=resnet50 -m=mt -b=32 --gpu=0 --lr=0.0003 --epochs=10

Testing

To test age estimation model with Resnet50.

python test_moon_age_estimation.py -a=resnet50 -m=mt -b=32 --gpu=0

Predicting

To predict the age of detected craters using the trained age estimation model with Resnet50.

python pred_dete_moon_age_estimation.py -a=resnet50 -m=mt -b=32 --gpu=0

Age Estimation's Flow

age estimation's flow

References

[1] Tarvainen, Antti, and Harri Valpola. "Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results." Advances in neural information processing systems. 2017.
[2] https://github.com/CuriousAI/mean-teacher