Image to image transition models for Earth Observation data
Organisatsiooni nimi
OÜ KappaZeta
Kokkuvõte
The emergence of large quantities of satellite data and the development of machine learning methods are enabling several new and exciting applications, that previously could only be dreamt of. Moreover, thanks to the systematic global data acquisition of Copernicus programme, the applications are also feasible for operational use.
Image to image deep learning models (implemented with e.g. generative adversarial networks (GANs) or convolutional neural networks (CNNs)) are one of the key technologies here. Transforming a non-intuitive synthetic aperture radar (SAR) image into a more conventional optical satellite image look, is one interesting direction currently in development in KappaZeta. With similar architecture one could also model the biomass maps from SAR images, when the optical image is missing (due to cloud cover), which would have enabled more direct measurement. The modelled biomass maps have high value for various agricultural applications. Artificial increase of spatial resolution (known as super resolution) has also been proven to be effective – based on a medium resolution satellite image (e.g. 10 m) one could model a high resolution image (1-3 m).
The questions that would be studied in the thesis:
1.\tHow to develop a model, which is accurate (the modeled image is as similar as possible to the target or ground reference)?
2.\tHow to develop a model, which is fast to fit and predict?
3.\tHow to improve the accuracy of the model with minimum training data?
Image to image deep learning models (implemented with e.g. generative adversarial networks (GANs) or convolutional neural networks (CNNs)) are one of the key technologies here. Transforming a non-intuitive synthetic aperture radar (SAR) image into a more conventional optical satellite image look, is one interesting direction currently in development in KappaZeta. With similar architecture one could also model the biomass maps from SAR images, when the optical image is missing (due to cloud cover), which would have enabled more direct measurement. The modelled biomass maps have high value for various agricultural applications. Artificial increase of spatial resolution (known as super resolution) has also been proven to be effective – based on a medium resolution satellite image (e.g. 10 m) one could model a high resolution image (1-3 m).
The questions that would be studied in the thesis:
1.\tHow to develop a model, which is accurate (the modeled image is as similar as possible to the target or ground reference)?
2.\tHow to develop a model, which is fast to fit and predict?
3.\tHow to improve the accuracy of the model with minimum training data?
Lõputöö kaitsmise aasta
2021-2022
Juhendaja
Kaupo Voormansik
Suhtlemiskeel(ed)
eesti keel, inglise keel
Nõuded kandideerijale
The candidates should have basic software development skills and know the fundamentals of machine learning and deep learning. Be able to self-organize, work independently and in a team. Interest towards Earth Observation and satellite technology is required, but prior knowledge is not. KappaZeta can provide the basic training needed to get going.
Tase
Bakalaureus, Magister
Kandideerimise kontakt
Nimi
Kaupo Voormansik
Tel
Vaata lähemalt