LTAT magistritööd – Master's theses

Selle kollektsiooni püsiv URIhttps://hdl.handle.net/10062/30974

Sirvi

Nüüd näidatakse 1 - 20 1238

Väike- ja keskettevõtete infoturbe hindamine: kohandamise vajadused F4SLE raamistikule ja sellele tuginevale MASS tööriistale
(Tartu Ülikool, 2025) Neider-Kuusalu, Pille; Seeba, Mari, juhendaja; Oja, Tarmo, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
In today’s digital environment, ensuring cybersecurity has become critically important for businesses, yet small and medium-sized enterprises (SMEs) often face limited resources, low awareness, and a lack of suitable tools. Although cyber threats aect both large and small organizations, SMEs are particularly vulnerable due to the frequent absence of a systematic approach to security assessment. According to the Estonian Information System Authority, many smaller organizations lack the capacity to adequately evaluate their cybersecurity needs. The aim of this master’s thesis is to analyse to what extent the self-assessment tool MASS (Measurement Application for Self-assessing Security) meets the needs of SMEs, with a focus on usability, clarity, and practical applicability. While MASS is not based on any specic standard, it is analysed here in the context of the F4SLE (Framework for Security Level Evaluation) framework, which supports maturity-level assessments in cybersecurity. The study concentrates on the experiences, expectations, and practical challenges SMEs face when using the tool. The empirical part includes semi-structured interviews and expert evaluations to gather input on user experience and functional requirements. The analysis highlighted four main development needs: improving the questionnaire’s structure and usability, enabling role-based completion, adding a proling page, and enhancing linguistic clarity. This thesis contributes to a little-studied area in Estonia– the adaptation of cybersecurity self-assessment tools for SMEs– and presents concrete recommendations for improving the usability and applicability of the MASS tool.
Asset-Oriented Threat Analysis for Large Language Model Systems
(Tartu Ülikool, 2025) Karagjaur, Mihhail; Matulevičius, Raimundas, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Large language model (LLM) deployments continue to proliferate across enterprises without systematic guidance on risk analysis of the LLM-based systems. Addressing this gap, the present study designs and validates an asset-oriented threat model, tailored to LLM systems. The research follows a design-science research paradigm. The research method incorporates (1) a systematic literature review of 45 peer-reviewed and grey sources, which led to the definition of 13 parent attack classes, a total of 24 threat variants. (2) A design of a threat model, which formalized the LLM business and system assets, their security criteria, mapped threats, security requirements, and countermeasures. (3) Two validation procedures, comprising a feasibility analysis of the threat model’s applicability and an empirical test of a jailbreak attack. The feasibility analysis determined that the proposed threat model, mapped to the Mistral Small 3.1, achieved a completeness score of 0.93 out of 1.00. Thus, indicating all but one of the seven system assets were fully represented in the real-world system. To further substantiate the applicability of the threat model, a jailbreak attack (prompt-injection) was executed with 100 prompts from the JailbreakV-28K benchmark open dataset. Without an official safety measure enabled, 78% of applicable prompts resulted in harmful output. With the safety measure enabled, the rate of harmful output was reduced to 70%. Indicating partial but insufficient mitigation. The main artifact of the thesis is an asset-oriented threat model for LLMs. The artifact consists of the following components: 1. High-level UML class and state, and BPMN process diagrams, depicting an LLM system, and mapping elicited threats to the system’s assets. 2. An interactive web page, which allows practitioners to traverse the produced threat model and to acquire information about the elicited assets, threats, and proposed countermeasures. 3. Code of the interactive web page, empirical tests, and datasets, supporting local use of the threat model and reproducibility of the jailbreak empirical test. Findings conclude that the LLM system possesses a wide attack surface while adding unique vectors such as jailbreak and embedding inversion. The thesis provides security and AI engineers with a systematic approach to risk analysis and countermeasure selection. Although the threat model was validated on a single open-weight model, the baseline methodology is model-agnostic and extensible. Future studies could validate the threat model against a wide set of LLM systems and automate control recommendations in the scope of DevSecOps.
Customer-Centric Redesign of the Mortgage Application Process in Baltic Banks
(Tartu Ülikool, 2025) Beisembayev, Merey; Milani, Fredrik Payman, juhendaja; Halas, Yana, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Placing customers firmly at the center of the business underpins the main idea behind customer centricity. Previous studies have effectively indicated the benefits of integrating customer centricity into business processes, such as gaining a competitive advantage and increased customer satisfaction. However, the mortgage process remains understudied from a customer perspective, particularly in the context of the digital environment. This thesis investigates how Baltic banks can redesign their digital mortgage application process to be more customer-centric, focusing on the experiences and pain points of private mortgage clients in Estonia, Latvia, and Lithuania. For this purpose, the research adopted a qualitative approach to analyze collected data from competitor analysis of Baltic banks' mortgage webpages and semi-structured interviews with borrowers. The competitor analysis resulted in 40 content elements and features, identifying customer service gaps in how banks communicate mortgage information to prospective borrowers before application. Furthermore, through thematic analysis of the semi-structured interviews, we extracted 6 themes and 22 subthemes, characterizing customers' unmet needs, challenges, and overall experience dealing with the mortgage process. Finally, the empirical findings were synthesized to propose 12 actionable customer-centric recommendations for improving the customer-centricity of the mortgage process. This thesis provides a deeper insight into the mortgage experience of Baltic banks' borrowers and contributes to the implementation of customer-centric business redesign in the banking industry.
Framework for Privacy-Preserving Synthesis of Textual Data
(Tartu Ülikool, 2025) Stomakhin, Fedor; Laur, Sven, juhendaja; Kamm, Liina, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
To safeguard patient privacy, sharing medical record data for research must adhere to various privacy regulations. To facilitate data sharing, various data protection techniques have been proposed, such as pseudonymization, anonymization and the use of synthetic data. The aim of synthetic data generation is, based on an original dataset, to produce a new dataset in a way that preserves the statistical relationships within the original data while not exposing any identifying or sensitive information about the data subjects therein. Synthetically generated data can still be insufficient from the point of view of privacy-preservation. For this purpose, approaches rooted in differential privacy (DP) have been proposed. DP typically relies on worst-case assumptions about attackers' knowledge, potentially leading to overly conservative measures. Applying DP principles to free-form text, such as medical epicrises, is complicated by their high dimensionality and complexity, as the same information can be conveyed in many different ways. In this work, motivated by the challenges of sharing textual health data, we propose and apply a general framework for evaluating privacy risks in text generated by large language models (LLMs). Considering a journalist attack model, we adapt differential privacy principles, quantifying privacy loss (ε, δ) based on the outputs of specific attack functions rather than relying on worst-case assumptions of DP. We demonstrate the framework by establishing baseline privacy characteristics via direct n-gram sampling analysis on both medical and social media texts and by exploring membership inference signals using surprisal analysis on LLMs fine-tuned with social media texts. While assessing synthetic data from standard LLMs highlighted methodological challenges, the framework provides a methodology for evaluating the privacy properties of text generation models and their outputs, informing decisions on sharing such data for research purposes.
Vision-based Localization on City Scale Using Open Street Map
(Tartu Ülikool, 2025) Sokk, Helena; Muhammad, Naveed, juhendaja; Zabolotnii, Dmytro, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Autonomous vehicles, like other robots, need to localize themselves in order to navigate. While global navigation satellite systems (GNSS) such as GPS can provide such vehicles with localization information, the GNSS information might not always be available. Since localization is one of the crucial components in self-driving vehicles, it is important to develop robust techniques to accomplish it. One such localization technique for vehicles to localize is using particle filters, given a map of the environment. The goal of this thesis was to implement a robust localization framework that integrates the particle filter with vision-based street sign detection and Open Street Map, without any reliance on GNSS. The proposed framework was able to localize the vehicle within a radius of 10 meters of its ground truth location, showing promising results. The implemented framework provides a good starting point for any future improvements and experiments in the problem of GNSS-free localization in autonomous vehicles.
Õpetajamaterjal programmeerimise õpetamiseks II ja III kooliastmes
(Tartu Ülikool, 2025) Keps, Kaisa Liina; Suviste, Reelika, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Learning programming is essential for developing students' problem-solving and frame thinking skills, but teachers are often reluctant to teach programming because of a lack of materials. This master’s thesis aimed to create materials for teachers to contribute to the teaching of programming and to reduce the workload of teachers. The materials were aimed for the students of the second and third school level and developed using the ADDIE model, which was used to analyze teaching and learning methods for programming, to design the lesson structure and to create the necessary materials. The creation of the materials was supported by different online environments and a link to the teacher materials was included in the web application progema.ee. Based on feedback from teachers and students, improvements were made to the materials, the quality of the materials was assessed, and limitations and opportunities for further development were identified.
Exploring Social Bias in Language Models through the Lens of Cinema
(Tartu Ülikool, 2025) Rikanson, Liisa; Sabir, Ahmed Abdulmajeed A, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Language models have revolutionized natural language processing, becoming an integral part of many applications. However, these models often exhibit societal biases embedded in their training data, raising concerns about their fairness and ethical deployment. Measuring these biases usually requires creating datasets with time-consuming human annotation, which is costly and hard to expand. To address this challenge, we propose a data curation framework and CineBias, a novel dataset of 1,012 stereotypical sentence pairs covering seven bias categories, extracted from Hollywood movie subtitles with minimal human intervention. We evaluate the language models BERT, RoBERTa, and ModernBERT using the CrowS-Pairs Score (CPS) on CineBias, and find bias levels comparable to established benchmarks (e.g., BERT 61.2% CPS). This shows that CineBias provides a scalable way to measure bias. We also demonstrate its applicability to low-resource languages with an Estonian case study.
Measuring the Impact of Developer Experience Improvements on Engagement in Pipedrive
(Tartu Ülikool, 2025) Laats, Henri; Pfahl, Dietmar Alfred Paul Kurt, juhendaja; Dorokhov, Mykhailo, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Pipedrive is a customer relationship management (CRM) software company where a significant portion of the workforce consists of developers. Understanding how changes in developer experience (DX) affect overall employee engagement is key to allocating resources effectively. In this context, employee engagement refers to how motivated and enthusiastic employees are about their work and their organization, reflecting their commitment to stay and go the extra mile. Engaged employees care about performing well and contributing to the company's success, which can lead to growth and increased revenue. Pipedrive tracks overall employee sentiment through company-wide engagement surveys and collects additional feedback from developers through DX surveys. However, the relationship between improvements in DX and overall employee engagement has not been analyzed, leaving it unclear whether a better developer experience also boosts employee engagement. The aim of this thesis is to analyze data from previous surveys to understand the relationships between areas in developer experience and overall employee engagement. The thesis will also explore which aspects of DX most influence engagement scores. Various statistical methods such as correlation, regression and trend analysis are used to identify these key relationships. The findings provide insight into whether improving developer experience can influence engagement positively and describe how to identify areas for potential DX improvements that would have the biggest impact on engagement.
Alignment and Safety Challenges in a Superintelligent AI Landscape
(Tartu Ülikool, 2025) Jõemaa, Evelin; Eden, Grace, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Artificial intelligence (AI) technologies are developing at a rapid pace, and with that, the discussions on AI safety have gained importance. This thesis aims to explore the concerns that members of online AI alignment communities have in regards to AI safety and alignment and compare them to the opinions of the key figures in the AI safety and alignment field. Five interviews were conducted via online conferencing platforms with participants recruited from online AI alignment communities. The interviews were on the topics of AI safety, AI alignment, and ethical concerns of AI development. Additionally, five interviews with key figures in the AI safety and alignment field were analysed from interviews available on YouTube for comparison with participant interviews. Interviews were transcribed, and thematic analysis was conducted to identify key themes. The findings show a significant concern that the participants have regarding current safety measures. The concerns were often related to the rapid speed of the advancements in AI technology and the shortcomings they saw with the current safety measures in being able to handle the developments. Participants saw the unintended consequences of AI development being a bigger risk in the future, but also highlighted the already present risks of current AI models, such as the ability for people to create deep-fakes, etc. The community forum participants highlighted the need for more collaboration between private companies and governments to have better measures put in place internationally for developing safe AI models.
Mastering the Unseen: Approaches to Hard-to-Detect Viral Cytopathic Effect
(Tartu Ülikool, 2025) Makarov, Aleksandr; Fishman, Dmytro, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Viral infections pose persistent global health challenges, making rapid, accurate assessment of viral activity crucial for research and diagnostics. Cytopathic Effect (CPE), morphological changes in host cells upon viral infection, serves as a critical visual indicator of viral load, yet its manual microscopy based assessment is laborious and subjective. Furthermore, simple automated classification often fails to quantify infection severity and struggles with "hard-to-detect" cases. This work presents a comprehensive survey and performance evaluation of various computer vision techniques, ranging from image classification to weakly and strongly supervised segmentation with both classical and deep learning-based models, for the automated detection and localisation of CPE induced by xenotrophic murine leukemia virus (x-MuLV). Our analysis demonstrates that supervised segmentation techniques provide a significantly more robust pathway for viral load quantification than explainability-based classification methods, particularly when analysing images displaying subtle cellular alterations with low viral load. This automated methodology offers an efficient, objective, and scalable alternative to manual inspection, facilitating high-throughput analysis and deeper insights into infection dynamics. Following extensive data preparation, this work systematically compared existing computer vision methodologies, thereby identifying and validating best-performing approaches for consistent and quantitative Cytopathic Effect characterisation, which offers a powerful tool to accelerate drug discovery, advance fundamental viral research, and improve automated virological assays.
Kindlustusettevõtte kõnede automaatne transkriptsioon ja sentimendi analüüs
(Tartu Ülikool, 2025) Lehtsalu, Kevin; Sirts, Kairit, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Automatic call transcription and analysis is a rapidly developing field within natural language processing, enabling organizations to extract valuable information from unstructured audio data. This master’s thesis explores how such solutions could be applied in the insurance domain, where recorded customer calls contain important insights into client needs, service quality, and internal processes. Although tools for automatic processing exist, they have not been systematically implemented in the organization under study - the content of calls has so far been assessed manually. In the first part of the thesis, various automatic transcription models (Whisper, Kaldi, and Wav2Vec 2.0) are tested to determine which performs best for processing insurance related calls in Estonian. The models are evaluated in terms of transcription accuracy and technical applicability, taking into account the specific challenges of low resource languages, such as morphological complexity and limited training data. The second part focuses on sentiment analysis based on the transcribed texts. Both lexicon based and machine learning based methods are compared to assess their ability to detect customers emotional stance or satisfaction. Such information is valuable for improving customer experience and gathering meaningful feedback. Based on the results, the thesis provides recommendations for selecting the most suitable transcription model and assesses under which conditions automatic sentiment detection may offer added value. As a next step, the organization could consider developing a prototype based on automatic analysis to support content-based processing of call recordings and improve both service quality monitoring and data management.
Improving the Sprint Review and Delivery Process by Enhancing the Monitoring System
(Tartu Ülikool, 2025) Soosalu, Karl; Pfahl, Dietmar Alfred Paul Kurt, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis focuses on the development and implementation of a tool aimed at automating sprint summaries for the Data Services and Reporting (DSR) department at Playtech Estonia. The primary goal of the tool is to provide real-time insights into sprint performance by analyzing story point estimation accuracy, time logged versus estimated hours, and task completion rates. By integrating with Jira, the tool automates data collection and analysis, eliminating the need for manual data extractions and enabling continuous feedback loops. The development process involved using Java, Spring Boot, and Vue.js, and testing was conducted to ensure functional correctness and usability in real-world conditions. The tool features a dynamic and user-friendly dashboard that supports better visualization of key metrics. The results demonstrated the tool's ability to track dynamic changes in sprint data and offer valuable insights contributing to improved sprint planning and retrospective analysis.
ERP-süsteemide juurutamise edutegurid väikse ja keskmise suurusega ettevõtetele
(Tartu Ülikool, 2025) Kivisoo, Elari; Ševtšenko, Eduard, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This master’s thesis identified the critical success factors (CSFs) for ERP system implementa-tion and developed a practical implementation framework designed to support small and me-dium-sized enterprises (SMEs) in systematically planning and successfully executing ERP projects, while considering their limited resources and specific needs. The theoretical part of the study was based on a conceptual literature review, during which 72 success factors identified in prior research were synthesised into 11 thematically distinct cate-gories. The relative importance of each factor was assessed based on its frequency in the liter-ature, and the results were used to construct a structured framework supporting effective ERP project management. The applicability of the framework was evaluated through a practical case study, which ana-lysed the success of a completed ERP implementation project and the related success factors. The case study established a qualitative link between the fulfilment of various CSFs and the project outcomes, highlighting their differing levels of influence. As a result, an practical framework was developed to provide methodological support for SMEs in the preparation and implementation of ERP systems. The author combined theoreti-cal insights with practical experience to develop a solution that can be applied in real-world ERP implementation contexts. Future research could involve testing the framework in various industries and companies to assess its broader applicability.
In-Depth Analysis of Miscalibration In Binary Classification
(Tartu Ülikool, 2025) Aavola, Heili; Allikivi, Mari-Liis, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Reliable probability estimates from binary classifiers are crucial for decisionmaking. While standard evaluation metrics provide an overall assessment of calibration quality, a deeper examination of miscalibration patterns can offer further insights into how calibration methods perform. This thesis presents an in-depth analysis of miscalibration patterns for five post-hoc calibration methods: Isotonic Calibration, Logistic Calibration, Beta Calibration, Histogram Binning, and Simplified Venn-Abers. Using a synthetic data framework with five diverse, known true calibration maps, we performed 100 simulation runs for each method-map combination. A suite of five specialized characterization plots was employed to visualize and understand nuanced error profiles, including accuracy, bias, variance, and directional tendencies in misestimation. The results reveal distinct behavioral characteristics and trade-offs. Parametric methods (Logistic, Beta) exhibited high stability but incurred significant systematic bias when their functional assumptions did not match the true probability landscape. Non-parametric methods (Isotonic, SVA) demonstrated superior adaptability and lower average error but with step-like outputs and slightly higher variance in complex regions. Histogram Binning showed considerable artifacts tied to its fixed-bin structure. The characterization plots successfully highlighted consistent directional biases and other nuanced error patterns not evident from aggregate metrics. This granular understanding reveals the precise behavior of different calibration methods, offering a more nuanced basis for selecting approaches tailored to specific application needs and risk sensitivities, particularly in complex or risk-sensitive contexts, moving beyond single performance scores.
An Approach for Designing Responsible Privacy Heuristics
(Tartu Ülikool, 2025) Pontes Da Costa Reis, Beatriz; Gharib, Mohamad, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Privacy compliance is a major concern for legal entities handling Personal Information (PI), as noncompliance leads to substantial fines. Regulations require these entities to implement privacy protection mechanisms (privacy solutions) and inform data subjects (DSs) about PI processing. However, DSs often struggle to understand relevant information and effectively use these mechanisms, leaving their privacy vulnerable. Privacy heuristics (PHs) offer a potential solution by assisting users in making informed decisions. Yet, their design is complex, prone to bias, and, if done irresponsibly, can lead to unethical or manipulative outcomes. This thesis addresses these challenges by developing an approach that provides design principles for guiding and evaluating Responsible Privacy Heuristics (RPHs) in privacy-aware systems. Following the Design Science Research methodology, we formulated the principles to satisfy six meta-requirements derived from ethical principles: Integrity, Non-manipulation, Beneficence and Non-maleficence, Autonomy and Control, Context-aware and Accessible, and Regulatory Compliance. Each principle is paired with acceptance criteria that practitioners can use to verify correct application. The clarity and applicability of the resulting eleven design principles, as well as the validity of their acceptance criteria, were evaluated by privacy domain experts. We demonstrate the applicability of the approach through a practical example, following the steps of the methodological process. The resulting design was validated via a moderated A/B test with 12 end-users. Participants were asked to complete demographic questions, read a scenario, interact with their assigned design version, and then respond to a post-task questionnaire that assessed perceived usability, perceived informed decision, perceived autonomy, and perceived consequences awareness. In addition, we evaluated informed decision and decision-awareness to measure the new privacy solution’s effectiveness. The results show that the RPH version matched the standard PH version in usability, while being slightly more effective in preventing the selection of privacy-invasive options and enabling informed decision-making, without compromising user autonomy.
Estonian Simultaneous Speech-to-Text Machine Translation
(Tartu Ülikool, 2025) Lepson, Henrik; Fišel, Mark, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Simultaneous machine translation is a task, where the translation system is expected to start translating before having access to the entire input sequence. This makes it a challenging and error-prone task. This thesis explored the feasibility of using pre-trained open models for simultaneous speech-to-text translation on the Estonian-English, Estonian-Russian, English-Estonian and Russian-Estonian directions. Two types of systems were evaluated: cascaded and end-to-end. The cascaded system relied on Whisper large-v3-turbo and NLLB-200 distilled 1.3B. The end-to-end system was based on Seamless M4Tv2 large. In addition, both systems used Voice Activity Detection (VAD) and LocalAgreement. The systems were compared with and without fine-tuning. For fine-tuning, a synthetic dataset with more than 4 million samples was created from various publicly available datasets. The dataset contained a 1:1 mix of full and partial sequences. The evaluation results showed that both systems are strongest on the Estonian-English direction followed by English-Estonian. Estonian-English direction can be translated without additional fine-tuning. Both systems struggled on the Estonian-Russian and Russian-Estonian directions. The translation quality and latency improved for both directions after fine-tuning.
Cell Cycle Phase Classification from Microscopy Images
(Tartu Ülikool, 2025) Zeynalli, Ali; Fishman, Dmytro, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Accurate classification of cell cycle phases is essential for cancer research and drug discovery. While fluorescence microscopy provides high-contrast, biomarker-specific signals that support precise classification, it relies on staining protocols that limit scalability and compromise cell viability. In contrast, bright-field microscopy offers a label-free, cost-effective alternative but poses challenges due to its lower contrast. This study compares five computational strategies for cell cycle phase classification using fluorescence and bright-field microscopy: traditional feature-based classification, segmentation-based classification, mask-guided classification via segmentation, nuclei patch classification, and nuclei patch classification via segmentation. Results show that fluorescence images support near-perfect classification performance across all methods. For bright-field images, the highest balanced accuracy of 0.770 was achieved using a nuclei patch classification approach with a ResNet-50 backbone, followed closely by mask-guided classification. These findings highlight the potential of deep learning models for accurate cell cycle classification in bright-field microscopy, advancing the potential for scalable applications in biomedical research.
Enhancing Mowing Event Detection by Mitigating Semi-Transparent Cloud Anomalies in Optical Satellite Image Time Series
(Tartu Ülikool, 2025) Tamkivi, Karl Hendrik; Komisarenko, Viacheslav, juhendaja; Shtym, Tetiana, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Cloud contamination in optical satellite imagery poses a major challenge in remote sensing, particularly for applications that rely on high-quality temporal data and where infrequent satellite revisits make each observation valuable. Traditional pixel-based cloud detection methods often struggle with semi-transparent clouds, which can be difficult to distinguish from atmospheric effects or land surface variations. This thesis introduces a time series-based approach for detecting semi-transparent cloud contamination in Sentinel-2 optical time series for Danish grasslands. A supervised anomaly detection model was trained to estimate cloud anomaly probabilities, which were then integrated into an existing mowing event detection framework through loss function modifications, custom network layers, or post-processing techniques. The results demonstrate that incorporating cloud anomaly probabilities improved model reliability by reducing false positives caused by cloud contamination. The findings highlight the potential of uncertainty-aware learning for enhancing event detection and other remote sensing applications affected by optical data contamination.
CircularCheck: A Tool for Detecting Circular Reporting
(Tartu Ülikool, 2025) Kaljuste, Kasper; Kangur, Uku, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
In the modern information landscape, the speed with which news is spread has reached unprecedented levels. This poses significant challenges in ensuring the accuracy and independence of information. Circular reporting is a situation where a piece of information appears to come from multiple independent sources, but in reality comes from only one source. Such practices can be intentional or accidental and contribute to the spread of false information by creating an illusion of corroboration. While circular reporting has been studied in intelligence and scientific literature, its detection in journalism, particularly in a small media ecosystem like Estonia, has received little attention. This thesis addresses the problem of detecting circular reporting in Estonian online news media. We present a system that detects circular reporting by building reference hierarchies and comparing article content across ERR, Delfi, and Postimees. Here we show that using a combination of link-based and text-based methods, it is possible to flag suspicious reference patterns for manual validation. The results show that 47 positive cases were detected by link analysis and 4 by text similarity. Self-referencing structures were the most reliable. These results reveal that although circular reporting is not widespread, it does occur and can be identified with relatively simple heuristics. The system does not attempt to verify the truthfulness of the information but instead focuses on tracing the propagation of references. This allows researchers and journalists to better assess the credibility and independence of sources. In a broader context, the results offer a framework that can be adapted to other media ecosystems and help improve media transparency.
Edge-Case Handling via Message-Based V2X for Enhanced Vehicle Autonomy
(Tartu Ülikool, 2025) Siniväli, Lisanne; Muhammad, Naveed , juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Autonomous vehicles (AVs) are becoming more common in public traffic, but they still struggle with rare, unpredictable situations (also known as edge cases) that current onboard perception systems often fail to handle. Existing AV systems mainly rely on cameras, LiDAR, and radar to understand their environment, but these sensors are limited by range, field of view, and environmental conditions. Infrastructure-based Vehicle-to-Everything (V2X) communication has been proposed as a solution to address these issues, but many approaches are complex and inefficient. This thesis investigates how AV safety in edge-case scenarios can be improved using a lightweight, event-driven V2X communication layer. The proposed system is based on simplified Decentralized Environmental Notification Messages (DENMs) that are triggered only by critical events. Compared to the usual onboard-only setups, this approach extends the detection range and gives the vehicle more time to react, especially in situations where perception fails or results in a delayed reaction. And since it only sends messages when needed, it avoids network overload while still increasing safety. The results suggest that you do not need a complicated or high-bandwidth system to make AVs safer in tough situations. With the right infrastructure support, even a small addition like this can act as a reliable safety layer and help AVs handle edge cases more confidently.

Sirvi

Viimati lisatud