Lunar sinuous rilles are thought to be lava channels or collapsed lava tubes formed during mare volcanism. A global distribution map cataloguing the rilles will help us in understanding the type of volcanism and lava flow viscosity, as well as detect potential lava tube entries and decide the safer landing spots for future moon missions. In this progressing project, using the SELENE KAGUYA multiband images and the survey conducted in 2013 that catalogued nearly 200 sinuous rilles on the Moon’s surface as our reference, we attempt to mark the sinuous rilles and use visualization techniques on Matlab to analyse how well our deep-learning networks work on the image data. Our sample of data includes rilles from Procellarum, Imbrium, Orientale, Serenitatis, Tranquillitatis, Nubium, Sinus Aestuum and Thomson rille basins. We create our gold-standard data using the MATLAB application created using GUI to create patches of various sizes. We work with deep-learning network architectures with and without the learnable parameters, adapting the architectures suitable for our purpose. The multiband images are trained on networks without the learnable parameters and the transfer learning (with the learnable) is performed on image data types mimicking RGB images. Our findings include which is the best network architecture and the better patch size of the image for the feature detection of up to 97.5% accuracy. We also bring forth topics and new visions to further this research in the outlook.
Lunar sinuous rilles are thought to be lava channels or collapsed lava tubes formed during mare volcanism. A global distribution map cataloguing the rilles will help us in understanding the type of volcanism and lava flow viscosity, as well as detect potential lava tube entries and decide the safer landing spots for future moon missions. In this progressing project, using the SELENE KAGUYA multiband images and the survey conducted in 2013 that catalogued nearly 200 sinuous rilles on the Moon’s surface as our reference, we attempt to mark the sinuous rilles and use visualization techniques on Matlab to analyse how well our deep-learning networks work on the image data. Our sample of data includes rilles from Procellarum, Imbrium, Orientale, Serenitatis, Tranquillitatis, Nubium, Sinus Aestuum and Thomson rille basins. We create our gold-standard data using the MATLAB application created using GUI to create patches of various sizes. We work with deep-learning network architectures with and without the learnable parameters, adapting the architectures suitable for our purpose. The multiband images are trained on networks without the learnable parameters and the transfer learning (with the learnable) is performed on image data types mimicking RGB images. Our findings include which is the best network architecture and the better patch size of the image for the feature detection of up to 97.5% accuracy. We also bring forth topics and new visions to further this research in the outlook.
AI BASED DETECTION OF LUNAR SINUOUS RILLES AND COMPARISION WITH MANUAL DETECTION METHODS
BANERJEE, DYUTIDEEPTA
2023/2024
Abstract
Lunar sinuous rilles are thought to be lava channels or collapsed lava tubes formed during mare volcanism. A global distribution map cataloguing the rilles will help us in understanding the type of volcanism and lava flow viscosity, as well as detect potential lava tube entries and decide the safer landing spots for future moon missions. In this progressing project, using the SELENE KAGUYA multiband images and the survey conducted in 2013 that catalogued nearly 200 sinuous rilles on the Moon’s surface as our reference, we attempt to mark the sinuous rilles and use visualization techniques on Matlab to analyse how well our deep-learning networks work on the image data. Our sample of data includes rilles from Procellarum, Imbrium, Orientale, Serenitatis, Tranquillitatis, Nubium, Sinus Aestuum and Thomson rille basins. We create our gold-standard data using the MATLAB application created using GUI to create patches of various sizes. We work with deep-learning network architectures with and without the learnable parameters, adapting the architectures suitable for our purpose. The multiband images are trained on networks without the learnable parameters and the transfer learning (with the learnable) is performed on image data types mimicking RGB images. Our findings include which is the best network architecture and the better patch size of the image for the feature detection of up to 97.5% accuracy. We also bring forth topics and new visions to further this research in the outlook.File | Dimensione | Formato | |
---|---|---|---|
Banerjee_Dyutideepta.pdf
accesso riservato
Dimensione
21.65 MB
Formato
Adobe PDF
|
21.65 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/65145