Bachelor theses
Deep learning for detection of defects in sewer CCTV
Author
Ondřej Chládek
Year
2020
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
doc. Ing. Pavel Kordík, Ph.D.
Department
Summary
This thesis is focused on defect detection of CCTV inspection of sewerage. The aim is to create a model of deep neural network which can determine with high accuracy whether a given image contains defective canalization or not and thus to make it easier for technicians who would otherwise have to analyze the whole record manually. Trained model for defect detection managed to achieve 99\% performance on test set. For defect classification I suggest using two models, first would detect if given image contains defective canalization or not, second would classify which defect is present in the image. Previously mentioned model was used for defect detection. Classification accuracy on test set was 87\%.
Detection of anomalies in water main flow time series
Author
Dmytro Molokoiedov
Year
2020
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
doc. Ing. Pavel Kordík, Ph.D.
Department
Summary
The water distribution network is a complex system of pipes and nodes connected
to each other, through which water flows. One of the most important problems
in such networks is leakages. This paper presents several methods to solve this
problem. A benchmark was set for a more accurate evaluation of algorithms and
methods such as one-class SVM, Isolation forest, and LSTM were implemented.
The data used in this work was artificially created based on real historical data
using the LeakDB framework.
Spatio-temporal prediction using artificial neural networks
Author
Jiří Pihrt
Year
2021
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Ing. Jitka Hrabáková, Ph.D.
Department
Summary
This thesis applies the methods of spatiotemporal prediction to the task of predicting the motion of traffic participants for autonomous vehicles. A sequence of bird’s-eye view artificial rasterized frames are used as input to neural networks which are trained to predict the most probable continuation of the sequence. The Lyft Motion Prediction for Autonomous Vehicles dataset is explored and adapted for this task. Several neural network architectures, namely ConvLSTM, PredRNN, PhyDNet, and U-Net, are researched, trained and their results compared.
Detection of text in historical maps
Author
Adam Peňáz
Year
2023
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Ing. Mgr. Ladislava Smítková Janků, Ph.D.
Department
Summary
Text Detection is a challenging task, especially when it comes to detecting specific types of text such as map Nomenclatures on historical maps. This thesis proposes two CNN-based networks, PSENet and TextPMs, implemented in PyTorch for detecting map Nomenclatures, and evaluates their performance using the Nomenclature dataset. Additionally, the thesis explores the effectiveness of pre-trained models on the TotalText dataset for detecting nomenclatures in historical maps. Although pre-trained models did not yield promising results, the proposed methods achieved f-measure scores of 88.8% and 91.1%, respectively, demonstrating their suitability for the task. Overall, the thesis contributes to the field of historical map analysis by introducing effective methods for text detection in this challenging domain.
Quantification of uncertainty of precipitation nowcasting
Author
Pavel Chudomel
Year
2023
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
doc. Ing. Pavel Hrabák, Ph.D.
Department
Summary
Uncertainty quantification is an important aspect of deep learning. Since most of the models cannot explain their predictions, it is essential to express a measure of their uncertainty. In this thesis we discover a multiple ways of achieving this. We implemented MC dropout and quantile regression into the PhyDNet model, a state-of-the-art prediction model for spatio-temporal phenomena -- specifically, precipitation measurement in our case. MC dropout proves to be a flexible method, yet it requires meticulous configuration. Quantile regression offers remarkable coverage probabilities, but we found out that this method can underestimate the uncertainty in low precipitation density areas. We present an elaborate overview of these methods and provide readers with a valuable guide for selecting the most suitable method depending on their specific requirements and goals.
Segmentation of historical maps by deep learning
Author
Matěj Polák
Year
2023
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Ing. Mgr. Ladislava Smítková Janků, Ph.D.
Department
Summary
The topic of the bachelor's thesis is the use of deep learning in the task of semantic segmentation of historical maps. The theoretical part presents current approaches to image segmentation in the context of historical document analysis. It then introduces definitions from the fields of image processing, machine learning and applied mathematics. In the practical part, a dataset of historical maps is analyzed and used for implementing two different deep learning models (UNet and TransUNet). It is followed by the experimental section, in which the effects of various modifications of the models on their performance are observed. Finally, the results of the practical and experimental parts are analyzed in the Discussion section. The outputs of the thesis can further be applied in the study of acquiring knowledge from historical documents.
Improving short-term rain prediction by using deep neural networks with advanced architecture
Author
Filip Miškařík
Year
2023
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Ing. Daniel Vašata, Ph.D.
Department
Summary
The goal of precipitation nowcasting is to give a precise short-term prediction of rainfall intensity in a local region. In recent years, this problem has seen wide adoption by many deep learning models for spatiotemporal prediction. This thesis focuses on improving one such model, PhyDNet, which disentangles physical dynamics from other, unknown, information in a two-branch architecture.
The branch for capturing the unknown phenomena uses a basic Conv\-LSTM model for general video prediction. In this thesis, we apply several modifications to the PhyDNet, ranging from small adjustments to the current architecture to experiments with swapping the ConvLSTM for different models, namely SA-ConvLSTM and PredRNN.
We present two promising alterations -- replacing ConvLSTM with a more complex PredRNN model and adding a skip connection to the branch that models physical dynamics. The results of our experiments show that these modifications can outperform the original network, especially for more long-term predictions.
Improving prediction of storm structure by spectral methods
Author
Adam Barla
Year
2023
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Ing. Daniel Vašata, Ph.D.
Department
Summary
Short-term weather forecasting (nowcasting) is crucial in predicting extreme weather events. This thesis focuses on structural biases that can be introduced into convolutional neural network predictions.
I investigated whether guided upsampling, used instead of transposed convolution, can improve the accuracy and reliability of weather forecasts. To this end, I trained UNet and Guided UNet (GUNet) models on radar images from a network of meteorological radars created by the OPERA radar program. I analyzed both models using performance indicators such as mean squared error (MSE), mean absolute error (MAE), and the structural similarity index (SSIM).
The results showed that the GUNet model slightly outperformed the UNet model regarding average MAE and MSE and demonstrated a better ability to capture higher frequencies in the Fourier spectrum of radar images. Moreover, the GUNet model achieved marginally better results on images with higher radar echo intensity, essential for predicting severe weather events.
The study suggests that the GUNet model can improve short-term weather predictions, and the results provide a basis for further research in this area.
Detection of cloud coverage using advanced deep learning methods
Author
Jaroslav Hradil
Year
2024
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Department
Summary
Accurate cloud detection is a key component of the preprocessing needed to prepare optical satellite imagery for various remote sensing applications. In recent years, numerous deep learning methods for cloud detection, such as convolutional neural networks or visual transformer, have been introduced, and this thesis aims to compare them.
In the thesis, current literature on cloud detection methods is reviewed, and the deep learning architectures U-Net and SegFormer are investigated in detail. They are implemented in the PyTorch framework, trained on the 38-Cloud dataset, which consists of Landsat-8 satellite imagery. Their performance is evaluated using accuracy, intersection over union (IoU) and F1 score.
The results show that SegFormer slightly outperforms the U-Net based architecture in terms of accuracy, while U-Net slightly outperforms SegFormer in the IoU and F1 score. Moreover, the U-Net based architecture has much shorter training time. This proves that even though being relatively simple and old, U-Net is still relevant and efficient architecture today.
The findings of this study can help other professionals choose the most suitable deep learning architecture for their use-case and provide a solid basis for further research.
Super-Resolution Enhancement of Weather Data Using Diffusion Models
Author
Jan-Matyáš Martinů
Year
2024
Type
Bachelor thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Ing. Miroslav Čepek, Ph.D.
Department
Summary
This thesis investigates the application of advanced deep-learning diffusion models, specifically SR3, SRDiff, and ResDiff architectures, for super-resolution of weather data. The primary focus is to evaluate these model's capability to enhance the resolution of meteorological variables from low-resolution inputs, a crucial aspect for an accurate weather forecasting and climate analysis.
Through experiments, conducted using the WeatherBench dataset, this work compares the performance of these models using a variety of validation metrics and explores enhancements through physics-based modifications and architectural improvements. The findings indicate that SRDiff, and ResDiff, further improved by incorporating physics-based filters, significantly outperform the traditional SR3 method, offering substantial improvements in capturing high-frequency details essential for accurate meteorological representations.
This thesis underscores the potential of integrating artificial intelligence with meteorological science to advance weather prediction capabilities, setting a foundation for future improvements in deep-learning diffusion models for weather data super-resolution.
Master theses
Improving deep learning precipitation nowcasting by using prior knowledge
Author
Matej Choma
Year
2022
Type
Master thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Mgr. Petr Novák, Ph.D.
Department
Summary
Deep learning methods dominate short-term high-resolution precipitation nowcasting in terms of prediction error. However, their operational usability is limited by difficulties explaining dynamics behind the predictions, which are smoothed out and missing the high-frequency features due to optimizing for mean error loss functions. This thesis summarizes our progress in addressing these issues.
Firstly, we present Intensity Classification Loss to improve the prediction of severe rainfall. The model is trained to predict the probability of precipitation with an intensity over 40 dBZ as a secondary output, which is compared to binary ground truth. Experiments have shown that this approach helps predict severe rainfall but does not predict precipitation with higher intensities than the selected threshold.
Secondly, we experiment with hand-engineering of the advection-diffusion differential equation into a PhyCell to introduce more accurate physical prior to a PhyDNet model that disentangles physical and residual dynamics. Results indicate that while PhyCell can learn the intended dynamics, training of PhyDNet remains driven by loss optimization, resulting in a model with the same prediction capabilities.
Short-Term Precipitation Forecasting from Satellite Data Using Machine Learning
Author
Jiří Pihrt
Year
2023
Type
Master thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
doc. Ing. Kamil Dedecius, Ph.D.
Department
Summary
Geostationary meteorological satellites are a source of global and frequent weather observations, but they do not directly observe precipitation. We research existing methods for inferring and forecasting rainfall from satellite data. The aim of this thesis is to predict high resolution precipitation radar observations up to 8 hours ahead from larger context but lower resolution multi-spectral geostationary satellite images. We develop a novel deep learning model for this task, utilizing the U-Net and PhyDNet neural networks. We name it WeatherFusionNet, as it fuses three different ways to process the satellite data; predicting future satellite images, estimating precipitation in the input sequence, and using the input sequence directly. To train and test it on real data, we participate in the NeurIPS Weather4cast 2022 competition, which provides spatially and temporally aligned satellite imagery and target precipitation radar data. WeatherFusionNet achieved first place in the Core challenge of the competition. We further experiment with several different models, try including static data in the input, and compare our model with a direct radar-to-radar model.
Improving neural cellular automata by incorporating physical dynamics
Author
František Koutenský
Year
2023
Type
Master thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Mgr. Alexander Kovalenko, Ph.D.
Department
Summary
This thesis presents the Physically Informed Neural Cellular Automaton (PINCA) model, a novel approach combining Neural Cellular Automata with the integration of physical dynamics. PINCA seeks to uncover governing equations from minimal datasets by learning differential operators through convolutional filters. A case study involving leopard coat patterns demonstrates the model's efficacy in deriving a unique reaction-diffusion system. Additionally, an exploration into the hidden states of Growing Neural Cellular Automata aims to shed light on cell specialization phenomena. The thesis comprises a literature review and the development and application of the PINCA model. This work bridges artificial intelligence and physics, offering new tools for modeling and understanding complex systems.
Incorporating Spatial Information in Deep Learning Models for Weather Prediction
Author
Dominik Chodounský
Year
2024
Type
Master thesis
Supervisor
Mgr. Petr Šimánek
Reviewers
Mgr. Alexander Kovalenko, Ph.D.
Department
Summary
Accurate weather prediction is critical for numerous applications, ranging from agriculture to disaster management and prevention. Many works that attempt to use deep learning models to generate weather forecasts fail to address spatial patterns and localized phenomena, which are typical of atmospheric variables. This thesis explores various approaches to building deep learning models that
account for the inherent prediction uncertainty by modelling the forecasts as probabilistic. Furthermore, it considers several methods of encoding spatial information in an attempt to leverage the correlation between a location's geographic profile and the weather patterns it observes. By incorporating learned
hex2vec embeddings of H3-regionalized map data in a Temporal Fusion Transformer, we have been able to show significant improvement in the prediction accuracy as well as demonstrate the potential for explainability with the use of this setup.