NIMS - MSc Dissertations Geospatial Technologies (Erasmus-Mundus)
URI permanente para esta coleção:
Navegar
Entradas recentes
- Spatiotemporal Deep Learning for Detecting Illegal Deforestation Using Sentinel 1 and 2 Imagery and Context-Aware ModellingPublication . Waliba, Thomas Liyajabale; Gharun, Mana; Oliver, Sergio Trilles; Painho, Marco Octávio TrindadeDeforestation in the Amazon, primarily driven by cattle ranching, threatens biodiversity, climate stability, and Indigenous communities. Existing satellite monitoring systems like PRODES and DETER struggle to detect small-scale and illegal deforestation. This study develops a spatiotemporal deep learning framework that integrates multi-sensor satellite imagery and contextual information to improve deforestation detection and assess the legality of forest harvest activities. The model uses 3D U-Net architecture, capturing both spatial and temporal dependencies in Sentinel 1 and 2 data. A dual-head prediction system is employed, where the first head detects deforestation, and the second classifies its legality. The model incorporates context-aware illegality modelling using Feature-wise Linear Modulation (FiLM) to include GIS-based contextual data such as protected areas, indigenous territories and concessions. The composite loss function combines Binary Cross-Entropy (BCE) with logits and Tversky loss to improve sensitivity to small deforestation patches. Results show high performance in both deforestation detection and legality classification, with improved F1 scores when vegetation indices are used. A strong spatial association between deforestation and cattle ranching is observed, with 92% of deforestation occurring near cattle ranching zones in São Félix do Xingu, Brazil. The model's generalizability is confirmed through cross-regional testing. This study introduces a novel context-aware illegality modelling approach, advancing beyond traditional deforestation detection to inform policy frameworks like the EU Deforestation-Free Products Regulation (EUDR). Future work will enhance temporal resolution and integrate additional contextual data to broaden its applicability.
- AI-Based Flood Detection Using TerraMind and Sentinel DataPublication . Tariq, Ayesha; Knoth, Christian; Pla Bañón, Filiberto; Painho, Marco Octávio TrindadeFloods are among the most frequent and damaging natural hazards. Timely flood extent maps support emergency response, especially in high-risk regions such as South Sudan. Satellite flood mapping is often limited by data gaps, particularly the lack of cloud-free Sentinel-2 imagery during flood events. This thesis presents a deep learning framework that uses Sentinel-1 SAR and Sentinel-2 optical data when both are available and remain operational when one modality is missing. The framework combines a pretrained multimodal encoder (TerraMind) with a UPerNet decoder. Thinking in Modalities is used to enable inference with incomplete inputs. The model is fine-tuned on the Sen1Floods11 dataset and evaluated using Sentinel-1 only, Sentinel-2 only, and Sentinel-1 + Sentinel-2 configurations. Flood mapping is derived from pre- and post-event predictions and produces flood probability maps, binary flood masks, and uncertainty outputs. Benchmark results show strong segmentation performance on Sen1Floods11. In the South Sudan transfer case, the fused Sentinel-1 + Sentinel-2 configuration produces the most coherent flood patterns. Sentinel-1 only fails to produce reliable flood maps in this study area. It overestimates flooding and introduces scattered false detections. Sentinel-2 outputs are cleaner, but clouds and cloud shadows reduce the observable flood signal and lead to underestimation. Overall, multimodal fusion reduces single modality errors and supports more consistent flood extent mapping under varying data availability.
- The Spatial Orientation Test in Virtual Reality: Comparing Allocentric Imagination and Embodied Egocentric Perspective-TakingPublication . Pant, Bibek; Schwering, Angela; Kruka, Jakub; Painho, Marco Octávio TrindadeSpatial perspective-taking is conventionally assessed using paper- or computer-based Spatial Orientation Tests. However, these two-dimensional modalities rely heavily on allocentric input and screen-based responses, which may fail to capture the intended embodied, egocentric demands of real-world navigation. With the increasing adoption of virtual reality in cognitive research, there is an opportunity to design a virtual reality– based Spatial Orientation Test based on embodied egocentric experience to capture true perspective-taking ability. This thesis investigated the concurrent validity of a novel virtual reality–based Spatial Orientation Test by comparing it with a traditional computer-based Spatial Orientation Test. In addition, the study examined whether performance discrepancies between modalities could be predicted by individual spatial thinking styles (allocentric vs. egocentric) or technological factors such as prior virtual reality experience and comfort. Twenty-nine participants completed both tasks in a counterbalanced within-subjects design. Results revealed a strong positive correlation (r = .73) between modalities, indicating that both tasks measure a common underlying spatial ability. Contrary to hypotheses, neither self-reported spatial thinking style nor prior virtual reality experience predicted performance differences or modality-specific advantages. These findings suggest that the virtual reality–based Spatial Orientation Test is a valid and accessible measure of spatial orientation. Crucially, however, the results also highlight the importance of task design, particularly given evidence that participants may adopt alternative solution strategies rather than the intended (egocentric) approach. These insights inform the design and interpretation of spatial orientation tests and support the principled development of future virtual reality–based Spatial Orientation Tests.
- Detection of Artisanal Small-Scale Mining(ASM) Using Segmentation-Based Deep Learning and Sentinel 2 Imagery In Madre de Dios, PeruPublication . Ofobi, Augustine Aborah; Gharun, Mana; Painho, Marco Octávio Trindade; Pla Bañón, FilibertoArtisanal and small-scale mining (ASM) is a major driver of environmental degradation in tropical regions, contributing to deforestation, land degradation, and water pollution. However, mapping ASM activities using satellite imagery remains challenging due to their small spatial extent, spectral variability, and similarity to other forms of land disturbance. This study aims to investigate how transferable ASM models can be from one mining site to another and focuses on two major ASM hotspots: Madre de Dios, Peru (source region), and the Kayapó Indigenous Territory, Brazil (target region), with analysis conducted using multi-temporal Sentinel-2 imagery acquired between January to July of 2024. The study presents a deep learning–based approach for detecting ASM sites using multi-temporal Sentinel-2 imagery and forest loss information derived from the Hansen Global Forest Change dataset. A U-Net semantic segmentation model with a ResNet50 encoder was employed to capture both fine-scale spatial details and high-level contextual features associated with mining activities. The model was trained using labeled data generated from areas of forest disturbance and evaluated using standard segmentation metrics, including Intersection over Union(IoU), Dice coefficient, precision, recall, and F1-score. Within the source region (Madrede Dios), the best-performing model achieved an F1-score of 0.933 and an IoU of 0.880,with overall accuracy ranging between 0.96 and 0.97 across feature configurations. To assess model generalization, both zero-shot transfer and fine-tuning experiments were conducted across geographically distinct regions characterized by differences in vegetation structure, soil properties, and mining morphology. Direct zero-shot transfer to Kayapó resulted in a reduced F1-score of 0.763 and an IoU of 0.648, reflecting substantial performance degradation due to domain shift. However, transfer learning through limited fine-tuning improved performance to an F1-score of 0.850 and an IoU of0.749, increasing spatial overlap with reference ASM areas from 51.66% to 80.06%. Spatial overlap analysis further demonstrated the model’s ability to capture mining-related land cover changes beyond conventional forest loss mapping, highlighting its potential as a complementary tool for environmental monitoring. Despite limitations associated with the use of global forest loss data as proxy ground truth, the study demonstrates that deep learning combined with freely available satellite imagery offers a scalable framework for ASM detection. The findings contribute to the development of transferable remote sensing models for monitoring environmentally destructive activities in data-scarce regions.
- Designing and Evaluating an Interactive Dashboard for Communicating Efficiency–Experiential Trade-Offs in Cycling Route-PlanningPublication . Neri, Margaux Elijah Palma; Kray, Christian; Oktay, Simge Özdal; Granell-Canut, CarlosCycling route planning involves navigating trade-offs between efficiency-oriented factors, such as travel time and distance, and experiential considerations, including safety, environment, and infrastructure quality. Although these factors are incorporated in many routing systems, there is often limited support for communicating them and allowing users to compare and balance such trade-offs for a more meaningful interpretation of route alternatives. This thesis examines how interaction and visualization design choices, specifically preference-adjustment mechanisms and coordinated visual components, support the communication of efficiency–experiential trade-offs in cycling route-planning dashboards. A design-oriented empirical study was conducted using an interactive dashboard that presents multiple route alternatives with efficiency and experiential metrics. Discrete and continuous modes of preference-adjustment mechanisms, enabling users to specify multi-criteria routing priorities, were evaluated. The dashboard was then assessed through a controlled user study focusing on users’ reasoning about trade-offs, perceived cognitive effort, usability, expressiveness, and trust across different commuting contexts. The findings indicate that differences between preference-adjustment mechanisms had limited impact on objective task performance but influenced how users engaged with and interpreted trade-offs. The two mechanisms differed in how cognitively manageable they were perceived to be, with more structured interactions and immediate visual feedback supporting clearer reasoning during exploratory route planning. Furthermore, coordinated map–metric visualizations were found essential in enabling comparison between route alternatives and in understanding the consequences of preference adjustments. No interaction mechanism emerged superior across all evaluated dimensions; instead, the results highlight a trade-off between efficiency-oriented and deliberation-oriented interaction styles. The thesis concludes by deriving design implications for interactive cycling route-planning dashboards and similar geospatial decision-support systems that aim to support transparent and interpretable multi-criteria decision-making.
- Data-Driven Anomaly Detection in Energy Consumption Using Deep Learning and Adaptive Threshold CalibrationPublication . Harris, George Nana; Risse, Benjamin; Kamps, Oliver; Henriques, Roberto André PereiraCampus energy consumption monitoring at scale is challenging because university buildings are heterogeneous and loads vary with weather, occupancy schedules, and seasonal transitions. Heating was selected as the focus because it is a major and strongly weather-driven contributor to campus demand. In addition, operational datasets rarely contain labelled fault events, limiting supervised evaluation of anomaly detection methods. This thesis proposed a forecast-driven anomaly screening framework for campus heating-energy monitoring using 15-minute smart-meter data from ten buildings at the University of Münster. A joint multi-building GRU model, conditioned on building identity, learns normal heating demand dynamics from cumulative heat meter readings transformed into 15-minute incremental consumption (kWh) and provides one-step-ahead forecasts as an expected baseline. Candidate anomalies are defined as persistent deviations between measured consumption and this baseline and are detected using residual scoring with adaptive EWMA thresholds to account for non-stationary residual behaviour. To enable calibration and benchmarking without ground truth, event-based synthetic anomaly injection is applied to derive per-building detector parameters and evaluate sensitivity across fault-like patterns. The GRU achieved R² = 0.86 and WAPE = 0.24 on the test split. In the injected-event evaluation, the detector achieved high precision; recall was highest for sustained positive deviations (over-delivery and bias), moderate for drift, and lowest for under-delivery and freeze-like patterns that resemble legitimate low-load operation. Finally, the framework was applied on an independent deployment-style monitoring window (October–December 2025), where detected event candidates were aggregated per building and visualized in a prototype campus spatial dashboard to support cross-building comparison and operational prioritization.
- HERI3D: A Comparative Analysis of Traditional and Deep Learning-Based 3D Reconstruction Techniques Using UAV Imagery for Cultural HeritagePublication . Guo, Ting-Jia; Risse, Benjamin; Catricheo, Constanza Andrea Molina; Oliver, Sergio TrillesThis thesis presents a comprehensive comparative analysis of traditional and deep learning-based 3D reconstruction techniques using Unmanned Aerial Vehicle imagery for cultural heritage documentation. The research evaluates four distinct reconstruction paradigms: a traditional Structure-from-Motion and Multi-View Stereo pipeline implemented in COLMAP, neural implicit surface reconstruction using Neuralangelo, radiance field representation via 3D Gaussian Splatting, and a feed-forward geometry-grounded Transformer model known as VGGT. The study uses datasets from three architecturally diverse castle sites in North Rhine-Westphalia, Germany: Schloss Münster, Burg Lüdinghausen, and Schloss Raesfeld. To ensure a fair comparison, a unified evaluation framework was established. This framework incorporates standardized image preprocessing, point cloud refinement, and geometric registration against airborne LiDAR reference data. The performance of each method was assessed through visual qualitative analysis and quantitative evaluation metrics, including Root Mean Square error for accuracy, Cloud-to-Cloud distance for completeness, and local geometric feature descriptors. The results demonstrate that traditional photogrammetry implemented in COLMAP remains the most reliable method for geometric accuracy. Among the learning-based approaches, VGGT with a moderate image count of 24 images consistently achieved the highest completeness and a balanced trade-off between accuracy and geometric stability across all sites. While 3D Gaussian Splatting provides superior visual continuity and color consistency, increasing the number of training iterations primarily refines surface appearance. Neuralangelo maintained global shape continuity but tended to smooth or underrepresent fine-scale architectural details. The findings highlight that reconstruction performance is strongly mediated by sites-pecific factors, including architectural complexity, UAV flight constraints, and the availability of reference data. This study contributes a reproducible comparative framework that serves as a structured reference for future digital heritage preservation efforts. The results emphasize that no single method dominates across all evaluation criteria and that method selection should align with specific documentation objectives.
- A Spatial Statistical Analysis of Women’s Lands Tenure in the Urabá Region, ColombiaPublication . García Polanco, Linda Katherine; Pebesma, Edzer; Mateu Mahiques, Jorge; Tang, Vicente de AzevedoThis research concerns women’s land tenure in Urabá Region in Colombia supported by spatial econometric modeling. It is motivated by the persistent gender inequalities in land tenure and the lack of comprehensive gender-disaggregated cadastral data, which limits the effective determination of land titling areas, empowerment of women and effective planning policies to address women’s needs. To deeper the understanding of the factors influencing women’s land tenure, this study examines a combination of socioeconomic and environmental variables. Women’s land tenure relies on socioeconomic conditions and soil productivity. Consequently, this research considers the following variables: Unsatisfied Basic Needs Index, Victimization Risk Index, Family Violence Against Women, Multidimensional Poverty, Land Suitability, Suitability for Banana Cultivation and Gini Index. Additionally, the armed conflict will be part of the analysis considering the historical predominance of illegal armed groups in Urabá, an external factor that has influence on land tenure. Building on these variables, the methodology used was based on geospatial information and on the social context of the region. The methods involve secondary sources primarily from agencies in Colombia to build the dataset and the implementation of the spatial models, SAR (Spatial Autoregressive Model), SEM (Spatial Error Model), SDM (Spatial Durbin Model) and CAR (Conditional Autoregressive Model). The literature review addressed topics starting with the land tenure and the gender variable, supported in geospatial technologies and statistics. Applying this methodological framework, the research reveals that lands, where the spatial modeling was implemented, reflect the critical land tenure of women with 1.5% of the total area of Urabá region. This study shows that interpreting the statistical model requires a transdisciplinary perspective, as an exclusively mathematical interpretation may result in inaccurate conclusions of the territory's reality and the population's needs. Building on these insights, the study ultimately contributes a novel transdisciplinary framework that combines spatial econometrics and geographic analysis to detect and analyze critical areas of women’s land tenure in Urabá, offering insights for policy and planning.
- Design and Usability Evaluation of an Adaptive Navigation User Interface under Privacy-Oriented Location ObfuscationPublication . Belay, Brktawit Haftu; Kray, Christian; Oktay, Simge Özdal; Huerta Guijarro, JoaquínLocation-based services and mobile navigation apps typically assume access to highly precise, continuously updated user location. This thesis investigates how navigation user interfaces can remain usable when location data are intentionally obfuscated to protect privacy. Conventional turn-by-turn applications assume continuous, precise GPS access, which exposes sensitive information about users’ homes, daily routines, and visits to sensitive locations. To explore a UI-level alternative, a mobile navigation prototype was designed featuring two parallel modes: a Precise mode resembling mainstream navigation apps and an Approximate / Block-Level mode. The design of these modes was grounded in theoretical requirements derived from literature on location privacy, location obfuscation/uncertainty, degraded GPS, and landmark-based wayfinding. A controlled within-subjects study (N = 30) compared both modes using a simulated, screen-based decision-point task with four turn decisions per mode. While Precise mode used a standard blue dot and metric distances, the Approximate mode visualized only a coarse position halo, simplified geometry, and landmark-based, qualitative directions presented as a deliberate privacy setting. Results showed higher turn-decision correctness in the Approximate mode within this prototype and scoring setup (94.2% vs. 77.5%). Importantly, this difference is partly explained by intentional block-level design choices that treated multiple turns as acceptable at specific intersections, reducing rigidity compared to the Precise mode condition. Subjective ratings were broadly comparable between modes; Approximate mode was rated slightly clearer on average, while differences in clarity and System Usability Scale (SUS) scores showed strong presentation-order effects. In line with H4, participants perceived Precise mode as more exact and Approximate mode as substantially more privacy-preserving. Everyday preferences remained situational, with participants valuing precise guidance for complex tasks and expressing high interest in an Approximate mode for privacy-sensitive trips. Overall, the findings demonstrate that a carefully designed, block-level navigation UI can support effective decision making at navigation-relevant decision points while offering meaningful privacy benefits. This thesis contributes a concrete Approximate-mode UI design, empirical evidence on the privacy-usability trade-off, and implications for future privacy-preserving navigation interfaces.
- Multi-sensor INLA-SPDE Bayesian Flood Mapping: A United Nations Case StudyPublication . Aguilar, Francisco Javier Lozada; Mateu Mahiques, Jorge; Risse, Benjamin; Castelli, MauroThis thesis addresses the critical challenge of flood detection in conflict-affected regions characterized by severe limitations in ground truth data, with a specific focus on South Sudan in collaboration with the United Nations Global Service Centre. Within the broader context of the Anthropocene and the intensification of climate-driven hydrological extremes, conventional pixel-based classification approaches often fail to capture the spatial continuity of flooding processes and remain highly sensitive to noise and uncertainty inherent in satellite observations. To address the dual challenges of small data regimes and multi-source sensor fusion, this research proposes a Bayesian spatial modeling framework based on Integrated Nested Laplace Approximation combined with Stochastic Partial Differential Equations (INLA-SPDE). By conceptualizing flood occurrence as a latent spatial process rather than a collection of independent discrete events, the model coherently integrates heterogeneous Earth Observation data (including Sentinel-1 SAR backscatter and Sentinel-2 optical spectral indices) together with physically interpretable static topographic covariates. A harmonized dataset covering ten Areas of Interest (AOIs) in South Sudan was specifically constructed for this study, comprising 130 flood rasters used for both training and validation . The results demonstrate that the proposed INLA-SPDE framework produces robust probabilistic flood susceptibility maps with outstanding predictive performance (AUC = 0.9949). This work constitutes, to the best of our knowledge, the first implementation of a binomial INLA-SPDE model for satellite-based natural hazard detection, advancing the field from deterministic flood mapping toward principled, evidence-based probabilistic inference in support of humanitarian decision-making.