SCCS Colloquium - Mar 20, 2019
|Date:||Mar 20, 2019|
|Time:||14:00 - 15:00|
Note: This is an irregular slot, this time at 02.07.023.
Yuanze Chen: Lithological Classification Based on Convolutional Neural Networks (CNNs) using multi-sensor data
This is a Master's Thesis submission talk. Yuanze is advised by Thomas Huckle and Dr. rer. nat. Melanie Brandmeier (ESRI Germany).
Deep learning has been used successfully in computer vision problems, e.g. image classification, target detection and many more. By using deep learning in conjunction with ArcGIS, a model with advanced convolutional neural networks (CNN) is implement for lithological classification in the Mount Isa region (Australia). The area is ideal as there is only sparse vegetation and besides freely available Sentinel-2 and ASTER data, several geophysical datasets are available from exploration campaigns. By fusing the data and thus covering a wide spectral range as well as capturing geophysical properties of rocks, we aim at achieving high classification accuracy. This approach can be used for other applications that aim at classifying land covers.
In this study, an end-to-end deep learning model is developed using Keras and TensorFlow, which consists of convolutional, pooling and de-convolutional layers. This model is inspired by the family of U-net architectures, where low-level feature maps (encoders) are concatenated with high-level ones (decoders), which enables precise localization. This type of network architecture is especially designed to effectively solve pixel-wise classification problems, which appropriate for lithological or land cover classification. The spatially re-sampled multi-sensor remote sensing data with different bands and geophysical data are stacked up into an image cube, which is the input for the model. The connection between ArcGIS and the deep learning libraries is achieved by using the Python API for ArcGIS and implementing the workflow into Jupyter Notebooks.
The U-net based model classifies each pixel of the multi-band imagery into different types of rocks according to a defined probability threshold. The overall classification accuracy of the U-net based model is 74.82%, which is 61.25% higher than the Random Forest Classifier. Moreover, integration of ASTER and Sentinel-2A imagery covers a wide spectral range and captures geophysical properties of rocks perfectly, which achieves sufficient bandpasses for lithological classification and gives the best lithological mapping.
Keywords: CNNs, Deep Learning, Lithological Classification, GIS
David Schneller: Evaluation of Graph Partitioning Algorithms for Load Balancing of Earthquake Simulations
This is a Bachelor's Thesis submission talk. David is advised by Carsten Uphoff.
Load balancing in finite element methods can often be achieved by transforming the mesh into a graph which is then partitioned. As the so-called graph partitioning problem is computationally hard to solve exactly, many heuristics and libraries have been developed for it, and are publicly available. We examine the partitioning libraries ParMETIS, PT-SCOTCH and ParHIP for using them to load balance the computations of the earthquake simulation program SeisSol. For this, we present an overview over common partitioning libraries, and then present an interface for the afore-mentioned libraries which is integrated into the mesh-reading library PUML. We close with an evaluation of ParMETIS, PT-SCOTCH and ParHIP, mostly for the case that we have vertex weights, but no edge weights, conforming with the currently used graph weighting model in SeisSol.
Keywords: graph partitioning, load balancing, SeisSol