|Topic:||Deep spatio-temporal models for satellite based Earth observation.|
|Supervisor:||doc. Boris Flach, Dr. rer. nat. habil.|
|Announce as:||Diplomová práce, Semestrální projekt|
|Description:||The research in this topic will contribute to a project funded by the European Space Agency in collaboration with Mapradix s.r.o. The project aims at spatio-temporal probabilistic neural models for classifying and detecting changes of land cover use in satellite images. The thesis will use the following types of initial data: medium spatial resolution and high temporal resolution satellite data as well as temporally sparse but very high spatial resolution satellite & airborne data with annotations. Its goal is to develop a spatio-temporal deep learning approach which consists of
(1) learning suitable feature representations by training 3D convolutional networks on the annotated part of the data.
(2) designing a recurrent approach for temporal processing of the features obtained from the previous step. This will require to enhance an existing approach (e.g. RNN, HMM) by capabilities to model seasonal variations as well as long term trends.
|Bibliography:||Alexandre Miot, Gilles Drigout, An empirical study of neural networks for trend detection in time series, arXiv:1912.04009|