JSC - Jülich Supercomputing Centre
Area of research:
Diploma & Master Thesis
Contract time limit:
Background:The cross-sectional team deep learning at the Jülich Supercomputing Centre (JSC) is conducting basic and applied research in the field of adaptive multi-layer neural network architectures that learn complex tasks from very large amount of unprocessed data. One of the tackled use cases is the processing and analysis of remote sensing images. Acquired data from sensors on-board aircraft and satellite platforms cannot be directly used by the applications. The interpretation of remote sensing images is not straightforward, and it requires a powerful yet highly accurate processing scheme to extract reliable and valuable information. Analytical methods such as machine learning and deep learning need be exploited to derive this value.
Project Description:The goal of the project is to design a deep network for the fusion of disjoint reflectance bands that have different spatial resolution and have been acquired by the same sensor. The considered data are the one acquired by the Sentinel-2 mission , which includes a constellation of two satellites that collect multispectral bands of 10m, 20m and 60m spatial resolution for visible, near infrared (NIR) and shortwave infrared (SWIR). The network will be developed for fusing the whole bands to obtain finer spatial resolution versions of the coarse bands.
Benefits:Direct access to high performance multi-GPU systems equipped with the state-of-the-art of deep learning frameworks (i.e., TensorFlow ).
Your Tasks:Revise Generative Adversarial Networks (GANs) .Reproduce the results of one proposed GAN network .Adapt the network for Sentinel-2 data and optimize the hyperparameters.Scale-up the training process.
Material:The candidate should understand what is Pansharpening and its application for remote sensing images. In short, due to physical constrains of satellite sensor instruments, a single sensor cannot acquire images that have both high spatial and spectral resolutions. This limitation can be solved with Pansharpening, that is a set of data fusion approaches which are used to fuse the panchromatic (i.e., high spatial resolution) with the multispectral images (i.e., high spectral resolution) . Pansharpening methods can be grouped into 4 categories: component substitution, multiresolution analysis, model, and super-resolution.
 sentinel.esa.int/web/sentinel/missions/sentinel-2  www.tensorflow.org  arxiv.org/abs/1406.2661  arxiv.org/abs/1805.03371  openremotesensing.net/wp-content/uploads/2015/02/IEEE_TGRS_2015_vivone_pansharpening.pdf  www.youtube.com/watch
Your Profile:Master’s student in Computer Science or Electrical/Computer Engineering Medium knowledge of Machine Learning and Deep Learning Medium/advanced knowledge of PythonEnglish proficiency is expected
Information required: Curriculum vitae (max 3 pages)List of previous projects/experiences where Machine Learning and Deep Learning algorithms have been adopted.
We offer:Financial support is provided by JSC for the duration of the thesis (i.e., 6 months with the possibility of extension to 9 months).