Jakub Dvořák

I'm currently a GIS analyst/technician at the Global Change Research Institute of the Czech Academy of Sciences, working on wildfire modelling.

So far, I have been testing wildfire modelling software, automating workflows and building simple websites.

Grassland monitoring using SAR in complex terrain is not fully understood and may come with challenges related to topography and sensor geometry. To explore these potential challenges, my thesis detected mowing events using a high-resolution DEM for precise coregistration and terrain correction of Sentinel-1 SAR imagery. Effect of local incidence angle on detection accuracy from interferometric coherence was also explored.

Detection accuracies in this thesis were higher than in previous studies when only considering SAR detections. The improvement was most likely caused by counting detections from individual orbits to assess the certainty of each detection. The results suggest that the current Sentinel-1 coherence processing techniques are suitable for mowing detection even in mountainous terrain, and further developments should focus on other aspects of the system, such as SAR-optical fusion and the detection algorithm.

E-TRAINEE is an open e-learning course on time series analysis developed by experts at four universities: Charles University (CZ), Heidelberg University (DE), University of Innsbruck (AT) and University of Warsaw (PL). The course is aimed primarily at MSc and PhD students of Remote Sensing.

During the project, I developed content for modules 1, 2 and 4 (mostly exercises), course-wide functionality (such as javascript-based quizzes) and led the course release team, which among other things migrated the repository from gitlab to github and automated website generation.

To reduce training data needs for Deep Learning, two approaches are most common – either pretraining models on larger datasets or augmenting the available training data. However, these commonly-used strategies are not enough for land cover classification in RS.

Our goal was to classify trees and shrubs from aerial orthoimages in the treeline ecotone of Krkonoše Mountains National Park, and Jeseníky Protected Landscape Area, Czechia. Instead of training a model on a smaller, human-labelled dataset, we semiautomatically created training data using an ancillary normalised Digital Surface Model (nDSM) and image spectral information. This approach can complement existing techniques, trading accuracy for a larger labelled dataset while assuming that the classifier can handle the training data noise. Weakly supervised learning on a CNN led to 68.99% mean Intersection over Union (IoU) and 81.65% mean F1-score for U-Net and 72.94% IoU and 84.35% mean F1-score for our modified U-Net on a test set comprising over 1000 manually labelled points. Notwithstanding the bias resulting from the noise in training data (especially in the least occurring tree class), our data show that standard semantic segmentation networks can be used for weakly supervised learning for local-scale land cover mapping.

For now, the main scientific output of this project is a conference paper co-authored by me, Markéta Potůčková and Václav Treml, which I presented at the 2022 ISPRS Congress in Nice.

Utilization of convolutional neural networks (CNN) has been growing rapidly in many fields, including remote sensing. At the same time, several textbooks and online learning materials have appeared. What is not so frequent or missing, are easy-to-use tools enabling practical experimentation with different designs of CNNs on actual remote sensing data. This tool, implemented in Python, helps users understand 1D, 2D, and 3D (spectral, spatial and spectro-spatial) CNN architectures for classification of hyper- or multispectral images, while presenting a straightforward framework for building more complex networks.

Target audience for our tool are MSc and PhD students, researchers and practitioners from public sector and industry in fields related to remote sensing and computer vision dealing with CNNs at a beginner/intermediate level.

CNN Compare was awarded 3rd place at CATCON (Computer-assisted teaching contest), where I presented it during the 2022 ISPRS Congress in Nice. It has since been incorporated into module 4 of E-TRAINEE.

I and my friend Michal Kolář have been giving Intro to GIS classes to students at Lauder secondary school. Our classes have both a theoretical and practical components using QGIS.

The students experiment with a series of simple spatial analyses, they also learn about data acquisition, analysis and visualisation.

As part of a research project studying long-term hydrological changes in upper Úpa basin in Krkonoše mountains, Czechia. My team analysed land cover change in the basin based on the Landsat/Sentinel data acquired between 1984 and 2022.

My main contribution towards this project was creating yearly composites and deriving features which served as input for classification. Initially, I built a procedure to create the composites in GEE, but for some more problematic years, the results were not satisfactory and further data manipulation in GEE was impractical. So to avoid that challenge, I created the final composites with a series of Python and R scripts.

The R scripts for compositing were based on an exercise in the Earth Observation course I took at Humboldt University, Berlin. My scripts were subsequently used by MSc students at Charles University and the method is included in the final project report.

Deep Learning has become widely used in the remote sensing community, especially as a classifier. First part of this thesis described deep neural networks commonly used for remote sensing classification and their various applications.

In the practical portion of my thesis, I used two deep convolutional Encoder-Decoder networks – U-Net and its proposed adaptation "KrakonosNet". They were used to perform a sematic segmentation of spruce trees and dwarf pine shrubs in the tree line ecotone of the Krkonoše Mountains, Czechia, based on aerial orthoimagery.

Resulting classification were compared to other common classifiers, namely Maximum Likelihood, Random Forest, and a Support Vector Machine. Both U-Net and KrakonosNet significantly outperformed the other classifiers on this dataset, and KrakonosNet was more succesful than a basic U-Net.

Prague ZOO leads european reintroduction efforts of Przewalski horses into mongolian steppes. The horses have so far been reintroduced into the Protected Area Gobi B, but the area often gets serious snowstorms, locally known as "white dzud". In the winter of 2009/10, the dzud was exceptionally devastating, as the local population decreased from 138 to 49 horses.

Therefore, Prague ZOO and Charles University started looking for new areas where Przewalski horses could flourish. My team focused on assesing the potential threat of the dzud, by building a 'dzud index' based on satellite and interpolated climatologic data. My role in the project was mostly focused on programming (GEE) and formulating the index. Our work is being extended by a bachelor thesis, which will improve the index and present it using a GEE app. After the thesis is finished, I will come back into the project to do some code refactoring, and co-author a final paper.

Currently (12/2023), Prague ZOO and local stakeholders are conducting in-situ monitoring of two potential sites pre-selected with our method. Plans are to start reintroducing horses to the area in the next year or so.

Cartographers like to come up with new ways of converting the globe to a flat map and a lot of those more obscure projections are not implemented in GIS. To display these projections, I developed a ArcPy script which projects a vector layer based on a mathematical function defined in either cartesian or polar coordinates.

I created this tool as a student project for a "Programming for GIS" course during the second year of my bachelor's. It can be used either as a standalone script, or be imported into ArcMap Toolbox.

This map introduces Czechia and its surroundings to readers of "Funerary Practices in the Czech republic" by Olga Nešporová, Institute of Ethnology, Czech Academy of Sciences. The book was published by Emerald Publishing with the map being printed in grayscale on a 15 x 10cm patch.

The final design is a product of intensive discussions with the book author, which allowed me to realize her vision and enhance it with my cartographic skill set.

A map showing three different ways of modifying map geometry for increased readability. Cities and their public transportation networks are most dense in the city centres, which presents a challenge for creating maps which are readable in the central portions, while still showing the whole network. My maps enlarge the city center through unipolar anamorphosis based on square root and natural logarithmic functions.

This is a student project (Thematic Cartography) from the first year of my bachelor's. The project outputs were a large paper map (A1 - 59x84 cm), a text description and a short Python script for ArcPy.

I created an interactive map to display around 400 recycling containers in Prague, which at the time had sensors for estimating how full they are. Sensor readings get collected every few hours and published online as a JSON. Since creating the original map, sensors were installed into more than 6000 other recycling containers in the city and the map still displays all of them.

Using jQuery and Leaflet, I wrote code to access the API and visualise sensor readings. The map was created for a course "Distribution of Spatial Data", which I took during my master's.