SCCS Colloquium - Feb 5, 2020

From Sccswiki
Jump to navigation Jump to search
Date: February 5, 2020
Room: MI 01.09.014
Time: 15:00 - 16:00

Sven Hingst: Shared-Memory Parallelization of a Parallel Combination Technique Framework

Bachelor's thesis submission talk, in German. Sven is advised by Michael Obersteiner.

Numerically solving high dimensional partial differential equations(PDEs) is compu- tationally difficult due to the curse of dimensionality. The sparse grid combination technique partially mitigates this problem. A highly parallel framework implement- ing this technique is the distributed combigrid module of SG++ which is currently only parallelized with message passing. Combining message passing with shared memory parallelization often yields better performance. In this thesis shared-memory parallelization is added to the distributed combigrid module. This allows users of the framework to use hybrid parallelization in their pde-solvers. Speedups up to 3 were measured, in many cases moderate speedups can be seen but for some cases minor slowdowns occurred in comparison to the original version. These slowdowns should still enable using hybrid parallelization for users of the framework and make it worthwhile.

Keywords: Sparse Grids, OpenMP

Nikolaos Ioannis Bountos: Subpixel Classification of Anthropogenic Features Using Deep Learning on Sentinel-2 Data

Master's thesis submission talk. Nikolaos is advised by Prof. Thomas Huckle

The classification of specific urban features is important to monitor and manage the growth of settlements. In this work, we investigate the performance of different deep learning architectures for subpixel classification on Sentinel-2 data using labels derived from UAV images for the mentioned kind of features. We trained different deep learning models based on state of the art architectures, such as DeepLabv3 and U-Net. We investigate early and late fusion approaches as well as the behavior and contribution of extra multispectral bands for the improvement of the performance of the models that simply use the RGB channels. Furthermore, we propose a data augmentation method based on acquiring images on the same area from different times of the year in order to improve the models’ generalization. We provide extensive quantitative evaluation of our methods as well as visual experiments. Additionally, we compare the visual results with traditional methods such as SVM and Random Forests.

Keywords: Deep Learning, subpixel, classification, remote sensing