@article{SteiningerAbelZiegleretal.2023, author = {Steininger, Michael and Abel, Daniel and Ziegler, Katrin and Krause, Anna and Paeth, Heiko and Hotho, Andreas}, title = {ConvMOS: climate model output statistics with deep learning}, series = {Data Mining and Knowledge Discovery}, volume = {37}, journal = {Data Mining and Knowledge Discovery}, number = {1}, issn = {1384-5810}, doi = {10.1007/s10618-022-00877-6}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-324213}, pages = {136-166}, year = {2023}, abstract = {Climate models are the tool of choice for scientists researching climate change. Like all models they suffer from errors, particularly systematic and location-specific representation errors. One way to reduce these errors is model output statistics (MOS) where the model output is fitted to observational data with machine learning. In this work, we assess the use of convolutional Deep Learning climate MOS approaches and present the ConvMOS architecture which is specifically designed based on the observation that there are systematic and location-specific errors in the precipitation estimates of climate models. We apply ConvMOS models to the simulated precipitation of the regional climate model REMO, showing that a combination of per-location model parameters for reducing location-specific errors and global model parameters for reducing systematic errors is indeed beneficial for MOS performance. We find that ConvMOS models can reduce errors considerably and perform significantly better than three commonly used MOS approaches and plain ResNet and U-Net models in most cases. Our results show that non-linear MOS models underestimate the number of extreme precipitation events, which we alleviate by training models specialized towards extreme precipitation events with the imbalanced regression method DenseLoss. While we consider climate MOS, we argue that aspects of ConvMOS may also be beneficial in other domains with geospatial data, such as air pollution modeling or weather forecasts.}, subject = {Klima}, language = {en} } @phdthesis{Somody2023, author = {Somody, Joseph Christian Campbell}, title = {Leveraging deep learning for identification and structural determination of novel protein complexes from \(in\) \(situ\) electron cryotomography of \(Mycoplasma\) \(pneumoniae\)}, doi = {10.25972/OPUS-31344}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-313447}, school = {Universit{\"a}t W{\"u}rzburg}, year = {2023}, abstract = {The holy grail of structural biology is to study a protein in situ, and this goal has been fast approaching since the resolution revolution and the achievement of atomic resolution. A cell's interior is not a dilute environment, and proteins have evolved to fold and function as needed in that environment; as such, an investigation of a cellular component should ideally include the full complexity of the cellular environment. Imaging whole cells in three dimensions using electron cryotomography is the best method to accomplish this goal, but it comes with a limitation on sample thickness and produces noisy data unamenable to direct analysis. This thesis establishes a novel workflow to systematically analyse whole-cell electron cryotomography data in three dimensions and to find and identify instances of protein complexes in the data to set up a determination of their structure and identity for success. Mycoplasma pneumoniae is a very small parasitic bacterium with fewer than 700 protein-coding genes, is thin enough and small enough to be imaged in large quantities by electron cryotomography, and can grow directly on the grids used for imaging, making it ideal for exploratory studies in structural proteomics. As part of the workflow, a methodology for training deep-learning-based particle-picking models is established. As a proof of principle, a dataset of whole-cell Mycoplasma pneumoniae tomograms is used with this workflow to characterize a novel membrane-associated complex observed in the data. Ultimately, 25431 such particles are picked from 353 tomograms and refined to a density map with a resolution of 11 {\AA}. Making good use of orthogonal datasets to filter search space and verify results, structures were predicted for candidate proteins and checked for suitable fit in the density map. In the end, with this approach, nine proteins were found to be part of the complex, which appears to be associated with chaperone activity and interact with translocon machinery. Visual proteomics refers to the ultimate potential of in situ electron cryotomography: the comprehensive interpretation of tomograms. The workflow presented here is demonstrated to help in reaching that potential.}, subject = {Kryoelektronenmikroskopie}, language = {en} } @article{MuellerLeppichGeissetal.2023, author = {M{\"u}ller, Konstantin and Leppich, Robert and Geiß, Christian and Borst, Vanessa and Pelizari, Patrick Aravena and Kounev, Samuel and Taubenb{\"o}ck, Hannes}, title = {Deep neural network regression for normalized digital surface model generation with Sentinel-2 imagery}, series = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, volume = {16}, journal = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, issn = {1939-1404}, doi = {10.1109/JSTARS.2023.3297710}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-349424}, pages = {8508-8519}, year = {2023}, abstract = {In recent history, normalized digital surface models (nDSMs) have been constantly gaining importance as a means to solve large-scale geographic problems. High-resolution surface models are precious, as they can provide detailed information for a specific area. However, measurements with a high resolution are time consuming and costly. Only a few approaches exist to create high-resolution nDSMs for extensive areas. This article explores approaches to extract high-resolution nDSMs from low-resolution Sentinel-2 data, allowing us to derive large-scale models. We thereby utilize the advantages of Sentinel 2 being open access, having global coverage, and providing steady updates through a high repetition rate. Several deep learning models are trained to overcome the gap in producing high-resolution surface maps from low-resolution input data. With U-Net as a base architecture, we extend the capabilities of our model by integrating tailored multiscale encoders with differently sized kernels in the convolution as well as conformed self-attention inside the skip connection gates. Using pixelwise regression, our U-Net base models can achieve a mean height error of approximately 2 m. Moreover, through our enhancements to the model architecture, we reduce the model error by more than 7\%.}, language = {en} }