ABSTRACT. Recent advancements in computer vision have significantly improved slum detection using very-high resolution satellite imagery. However, current algorithms are limited to identifying slums based on single-temporal labels and lack the capability to perform multi-temporal analysis. Here we present a supervised learning model trained on multi-temporal labels, specifically designed to maintain consistent performance across temporal variations. We evaluate our model against baseline approaches trained on singe-temporal labels. Case studies in two cities, Caracas and Karachi, demonstrate that incorporating additional temporal satellite imagery during training produces more consistent and reliable results for multi temporal analysis. Our study suggests new ways to leverage temporal data in slum detection to effectively monitor urban poverty and track dynamics of informal settlements over time.

ABSTRACT. Urban deprivation mapping is critical for addressing inequalities and achieving Sustainable Development Goal (SDG) 11.1.1, which focuses on ensuring access to adequate housing and services in urban areas. This study introduces a geospatial framework to operationalize previously conceptualized urban Domains of Deprivation related to unplanned urbanization, limited infrastructure, and limited services within city segments at the city-scale. Leveraging open, global datasets, including Google’s V3 building footprints and 2.5D building heights, the model assigns deprivation scores (ranging from 0 to 6) based on binary thresholds derived from median values. Validation against reference slum boundaries provided by the IDEAMAPS network achieved an F1- score of 0.45 for high-deprivation areas. The results highlight the spatial distribution of deprivation across Nairobi and demonstrate the reliability of dense building indicators for identifying informal settlements. The framework demonstrates computational efficiency, enabling citywide analysis using accessible resources, and highlights its potential to inform urban planning and targeted interventions through scalable geospatial methodologies aligned with SDG 11.1.1.

ABSTRACT. This paper investigates the identification of slums in Medellín, Colombia, conducted by two independent research teams. By comparing two similar methodologies, based on Earth observation (EO) data and utilizing manual visual image interpretation (MVII), this study highlights overlaps and differences in the identified areas. The findings reveal significant congruence in core informal area zones but notable divergences in peripheral and less densely developed regions. These results underscore the necessity for transparent conceptual definitions of the target class 'slum', for transformation of conceptual approaches in mapping approaches, for the explanation of effects of ambiguities in both, and thus for the development of improved mapping protocols.

ABSTRACT. Updatable and scalable maps of urban deprivation are needed to plan, upgrade, and monitor dynamic neighborhood-level changes within developing world cities, especially in Sub-Saharan Africa. Earth Observation data provides a promising solution for consistent, accurate high-resolution maps globally. However, most studies use very high spatial resolution images, which often cover only small areas and are cost prohibitive. Additionally, most of the previous work has only focused on distinguishing only between slums and formal areas. Our current work has started to look into moving beyond, slum-non slum dichotomy to look at the degree of deprivation within a city. Therefore, this work focuses on determining if we can expand our slum, non-slum, scalable machine learning approaches to represent this degree of deprivation. This is our first attempt to do this, and model results indicate moderate accuracy, but tend to over predict slum areas, especially in the areas where our on the ground data are captured. Future work should focus on understanding modeling uncertainty, expanding on the ground data locations, and the model inputs to determine our ability to scale up the degree of informality data.

ABSTRACT. Today, there is a consensus that informal settlements continue to expand in both area and population. These settlements are often associated with large cities, such as Dharavi in Mumbai and Kibera in Nairobi. However, secondary cities, rather than megacities, are the epicenters of population growth and urban expansion. The lack of reliable and consistent data, both globally and nationally, has hindered accurate measurement of whether informal settlements are also growing rapidly in secondary cities. This research project analyzes the growth rates of informal settlement areas and populations in Argentina at the national scale from 2016 to 2023, using RENABAP's quantitative and qualitative survey data combined with Earth Observation-based data. We found that the entire sample of informal settlements in Argentina continues to grow at an annual rate of 1.79%. More importantly, informal areas are more prominent and expanding at a larger scale in the less populated cities of the country. Our study confirms, at a national scale, that secondary and smaller cities have a higher number of informal settlements and faster growth rates than primary and larger cities. The scale of this difference should be alarming, prompting a reevaluation of how this compares to global trends in informal settlements. This is fundamental for allocating resources and shaping urban policy globally, especially to meet the needs of informal dwellers.

ABSTRACT. In informal settlements, accurate population data are often lacking, outdated, or incomplete, yet such information is crucial for assessing the population at risk in hazard-prone areas, such as parts of Medellín, Colombia. This study presents a methodology to estimate the population at risk from landslides in informal areas by combining deep learning-based building detection on aerial remote sensing images with a population estimation. A Mask R-CNN instance segmentation model is trained to detect individual buildings in orthophotos of informal settlements from 2019, using official cadastral data from 2018. Population estimates are made based on the size of the detected buildings. When applied to informal settlements within landslide-prone regions, our methodology identifies 10,180 more buildings and 28,575 more people compared to official cadastral data, revealing an underestimation of people at risk in these areas. This approach demonstrates the value of combining instance segmentation with population estimation to enhance risk assessment in data-limited settings.

ABSTRACT. Slums and informal settlements are among the most common forms of urban development in Africa. These areas often lack essential infrastructure and services, face socioeconomic disparities, and are increasingly vulnerable to climate risks such as floods. The Space4All project aims to promote sustainable ur- ban development by integrating advanced geospatial technologies with locally grounded approaches to analyze the intersection of livability and flood exposure in pilot cities in Ghana, Kenya, and Mozambique. To develop a scalable approach, we use open data such as Sentinel-2 satellite imagery. The methodology combines state-of-the-art AI models, Earth Observation data with Citizen-generated data collected via a custom app. The results highlight deprivation hotspots, revealing the intersection of spatial inequalities and flood exposure. Flood exposure is assessed through a comprehensive approach that integrates local knowledge gathered from workshops in informal settlements with flood models and historical rainfall data. This research provides actionable insights for urban planners, policymakers, and NGOs to prioritize targeted interventions and investments, fostering resilience through community-driven approaches.

ABSTRACT. After the Russian Federation troops entered Ukrainian territory on February 24, 2022, significant damage occurred throughout the country. In areas experiencing continuous conflict, much of the agricultural land has become unsuitable for farming. We characterize the development of agricultural cultivation areas based on multiple spectral indices using Sentinel-2 imagery to reflect wartime effects for two test sites in Eastern Ukraine between 2019-2023. More specifically, we establish season-trend analysis, automatically detect structural ruptures as associated to the conflict, and test for significant differences of the time series data w.r.t. the pre-invasion and post-invasion periods for the growing seasons. Our findings highlight the relevance of identifying agricultural cultivation areas and analyzing them through time series profiles, particularly with respect to the Normalized Difference Vegetation Index and Chlorophyll Index, to detect patterns of war-related impacts. Season-trend analysis-based decomposition of the time series reveals that the trend component for the CI-based indices is the most informative, showing a sharp decline during the onset of the war. Additionally, structural breaks observed in the CI-green index occur in very close temporal proximity to the invasion, suggesting that this chlorophyll index could serve as a valuable predictor for the start of the conflict. Finally, changes between the pre- and post-invasion periods are also highly statistically significant.

ABSTRACT. Increased heat stress on urban areas due to rising heat waves is threatening people’s lives and well-being. Temperature is a crucial parameter in climate monitoring and advances in remote sensing technology have expanded opportunities for monitoring surface temperature from space. With numerous thermal satellite missions anticipated in the coming years, there is an urgent need to improve methods for retrieving surface temperatures in cities. While Earth Observation (EO) satellites excel for Land Surface Temperature (LST) retrieval, urban surface properties and geometry, along with the trade-off between temporal and spatial resolution, complicate urban surface temperature (UST) retrieval. This study explored combined use of EO data for monitoring detailed, accurate, and frequent UST. EO4UTEMP is a Living Planet fellowship project exploiting satellite data for the assessment of urban surface temperatures. A UST retrieval algorithm is developed, including emissivity corrections that account for the third dimension of the urban surface. Coupled with a thermal imagery downscaling approach, the UST retrieval algorithm is increasing the UST observation frequency. The methodology was assessed against in-situ UST measurements in Heraklion, Greece, achieving a Mean Absolute Error of 3.6 K during the day. The methodology is transferable to cities worldwide, showcasing new technologies and promoting EO data use in urban meteorology.

ABSTRACT. The demand for detailed building roof data has led to the development of automated extraction methods to overcome the inefficiencies of traditional approaches. The complexity of urban residential building geometries presents significant challenges in achieving high accuracy. Re:PolyWorld, combining point detection and graph neural networks, offers a promising solution for reconstructing high-Level of Detail (LoD) building roof vector data. This study aims to enhance Re:PolyWorld's performance on complex urban residential structures by introducing attention-based backbones, dataset augmentation, and additional area segmentation loss. Despite dataset limitations, our experiments showed improvements in Ampoint position and line distance accuracy (12.33 pixels and 4.57 pixels) by integrating attention in the backbone, with a notable increase in building reconstruction score of 85.48% by implementing combination of attention and additional area segmentation loss. These results demonstrate the potential of advanced neural network architectures in addressing the challenges of complex urban residential geometries.

ABSTRACT. Remote Sensing Visual Question Answering (RSVQA) is a task aiming at automatic answering questions related to overhead imagery. Many studies have been conducted in recent years, focusing on the methods and the data. However, a recurrent problem is the lack of generalization abilities and robustness to questions with similar semantics but different wording. This work focuses on the data part, specifically the questions. Our objective is to make RSVQA models more robust to various changes in questions, more generalizable (e.g. to unseen phrasing, synonyms) and less susceptible to bias in the data. To this end, we propose to leverage the abilities of Large Language Models (LLMs) in the field of natural language processing, to enrich a RSVQA dataset by generating new questions with the same meaning and semantics. To showcase the effectiveness of this process we compare and confront a baseline, relying on back translation, and the proposed LLM-based approach on an urban dataset (RSVQA-HR). Our experimental study, with quantitative evaluation performances, highlights that models trained with the proposed data augmentation scheme are indeed more robust to unseen questions.

ABSTRACT. Wildfires are a growing global concern, causing significant damage to urban infrastructure each year. This study presents a novel approach for building damage assessment using generative artificial intelligence, focusing on the transferability of high-resolution satellite imagery models to lower-resolution datasets. Our diffusion-based model is trained on the xView2 Wildfire Building Damage Benchmark, a dataset specifically designed for wildfire-induced building damage detection. The model is further evaluated on real-world wildfire incidents in Lahaina, Hawaii, and Athens, Greece, demonstrating its effectiveness in damage localization across varying spatial resolutions. With competitive performance on benchmark datasets and practical utility in real-world scenarios, this work highlights the potential of generative AI for geospatial disaster assessment and urban resilience.

ABSTRACT. The increasing demand for smart urban systems and advanced mobility solutions highlights the critical need for precise, multimodal, and up-to-date Geographic Information System (GIS) data. While vehicular networks are well-mapped, pedestrian infrastructure often remains underrepresented, posing challenges for autonomous navigation and urban analysis. To address this issue, we present Street2GIS, an end-to-end open-source, multimodal framework that integrates multiview and cross-cue data fusion for automating the generation of GIS shapefiles for pedestrian infrastructure from street-view imagery and geo-coordinate data. Street2GIS employs hierarchical transformation networks to combine depth maps, semantic maps, and elevation data for geospatial feature extraction and the creation of georeferenced shapefiles. The results accurately identify roads, sidewalks, buildings, and vegetation, ensuring precise spatial alignment of urban elements within a 70-meter radius of each location. Evaluated using ground truth data from the Estonian Topographic Database, the framework demonstrated robust performance in Tartu, Estonia, achieving an F1-score exceeding 90% and alignment and similarity metrics of 91% and 72.95%, respectively. The proposed framework efficiently produces detailed urban maps and provides timely, high-resolution pedestrian infrastructure data suitable for applications such as autonomous navigation, urban planning, and environmental monitoring.

ABSTRACT. Buildings account for a substantial portion of global energy consumption, underscoring the importance of accurately assessing their energy efficiency. Traditional approaches for evaluating building energy efficiency typically rely on extensive tabular data gathered through labor-intensive on-site surveys, impeding their practicality for large-scale applications. Remote sensing data has emerged as a promising alternative due to its extensive spatial coverage and growing availability. However, existing approaches that utilize single-modality remote sensing data are limited in their capacity of predicting building energy efficiency. In this study, we propose a novel multi-modal, multi-task neural network to model building energy efficiency using remote sensing sources. Our approach builds upon three key pillars: the integration of diverse remote sensing data, the implicit inclusion of tabular attributes during training, and the application of foundation models for feature extraction. Extensive experiments on data from the city of Peterborough demonstrate the effectiveness of our method, achieving an F1 score of 68.14% for energy-efficiency classification. These results highlight our approach’s potential to enhance building energy modeling, providing a scalable and efficient alternative to traditional energy assessment techniques.

ABSTRACT. Monitoring land disturbance and land reclamation associated with oil/gas operations is crucial in gaining a better understanding of cumulative impacts of mining activities on the environment, assessing the collective effects of oil/gas production on habitats, and helping formulate effective strategies for ecological restoration and environmental sustainability. The use of satellite remote sensing imagery and deep learning techniques offers a desirable solution for identifying the extent of oil/gas well sites and detecting the land surface changes on the disturbed sites. In this study, we created an oil/gas wells dataset in Alberta, Canada based on high-resolution satellite imagery from RapidEye-2/3, WorldView-2/3, and SPOT-6. Our dataset consists of two subsets: one is for well site identification, which contains 12,095 paired satellite images and mask images of both well site and road labels with a resolution of 2 m and a size of 512×512 pixels; the other is for well site change detection, which is comprised of 328 pairs of bi-temporal images, land cover maps, and binary and semantic change maps with a resolution of 1.5 m and a size of 512×512 pixels. We tested commonly used deep learning networks for object detection and change detection on our dataset. The results can be used as a baseline for future deep learning architectures that focus on well site identification and change detection.

ABSTRACT. Accurately identifying slums is critical for effective urban planning and improving the lives of slum dwellers. This study presents a comprehensive evaluation of four deep learning architectures—U-Net, U-Net++, LinkNet, and the Segment Anything Model (SAM)—for semantic segmentation of slums using both medium-resolution Sentinel-2 and high-resolution Google Earth Pro imagery. We investigate the impact of five loss functions (Dice, Focal, BCE, Jaccard, and Tversky) on segmentation performance. Our results demonstrate U-Net's consistent superiority across both datasets, highlighting its effectiveness in capturing slum characteristics. LinkNet emerges as a competitive alternative, particularly in high-resolution imagery, showcasing its efficiency and suitability for precise boundary delineation. The study underscores the crucial role of spatial detail, with all models exhibiting improved performance on the high-resolution dataset. While SAM shows promise, its performance lags, emphasizing the need to further adapt foundation models to specialized tasks. This research provides valuable insights into model and loss function selection for slum detection, contributing to the development of more effective and robust slum mapping solutions.

ABSTRACT. On the wake of the global urbanization the world is experiencing, aspects of urban morphology become increasingly relevant to monitor as they are both directly impacting city dwellers and informative of our societal choices. With the rise of the quality and the quantity of remote sensing derived data pertaining to urban morphology, new paths of investigations can be formulated, exploring the morphological aspects of the global urbanization. In this study, we investigated whether disparities exist in the spatial distribution of intra-urban forms in different world regions. Using a global dataset of intra-urban patterns (I-UPTs) derived from remote sensing data, we computed the Richness and Simpson Diversity Index of I-UPTs of different regions. With this, we found differences in the diversities of the different regions of up to a factor 4.3 for the Richness and up to 9.6 for the Simpson Diversity Index. Stepping on this, we conclude that the history of global urbanization led to a substantially uneven distribution of the intra-urban morphological diversity across the different regions of the globe.

ABSTRACT. Landslides are hazardous phenomena that pose considerable threats to the environment and urban infrastructure. This study focuses on both detecting landslides from remote sensing imagery and evaluating the impact of landslides on urban infrastructure. For landslide detection, we implement a machine learning technique, i.e., Support Vector Machines (SVM) and deep-learning U-Net, in the Pomarico region of Italy. Sentinel-2 satellite data from 2019 were utilized for the analysis, with ground control points collected from the study area to validate the techniques’ results. The findings demonstrated that the SVM and U-Net achieved overall accuracies of 87.00% and 91.00% in detecting landslides, respectively. Additionally, the study revealed that landslides displaced electrical poles across the region, caused significant damage to several buildings, and disrupted the existing road network. In summary, these results provide valuable insights for urban planners and administrators, helping to enhance the effectiveness of methods used to simulate landslides and assess their impact on urban infrastructure.

ABSTRACT. Urban Geography is an increasingly data-rich research field where a wide range of research questions can be addressed by the joint analysis of remote-sensing and other data sources like social media data. A valuable step in data-driven explorative studies is the division of the study area into regions that are meaningful with respect to the data and the research question. Administrative regions are unlikely to fulfill this requirement. As an alternative for situations with datasets of different types, we algorithmically created clusters based on hierarchical clustering approaches with geographical constraints. Thereby, space can be divided into regions that are contiguous and accommodate requirements for within-region homogeneity. This approach uses well-established methods and is flexible and easily extendable to accommodate even heterogeneous data types. We demonstrate this in a case study from North African cities where we jointly use remotely sensed built-up age and text embeddings generated from big social media data.

ABSTRACT. Cities are called upon to become more sustainable. One urban planning concept to improve local accessibility that has been implemented by a number of cities around the world is the x-minute city. Yet, there is no clear consensus on which physical aspects of the urban fabric influence the x-minuteness of an area or city. With this in mind, this study shows that exploring one aspect, urban structure types, has a crucial impact on x-minuteness by comparing different areas in Berlin, Germany. The results show that the x-minuteness of an area is significantly dependent on the urban structural type. Furthermore, we show that different urban structural types not only affect the x-minuteness but are also directly related to the spatial distribution of amenities.

ABSTRACT. Urban form and urban function are closely related – and so are daily activities, interests, and concerns of urban citizens. While remote sensing data nowadays allow for the documentation of changes of the urban form on a global scale, and data on urban functions is readily available from multiple geodata sources, data on activities, interests, and concerns of urban dwellers are less accessible. This study relies on social media data in combination with earth observation-based mapping products to explore the relationship between social media (i.e., Twitter, now known as X) posts and urban morphology. Choosing the megacity Tokyo, Japan, as an example, our results revealed significant correlations between topics discussed on Twitter and urban morphologies. In high-rise areas the dominant topics were related to tourism, culture, and commerce, while in low- and mid-rise areas the dominant topics were more likely connected to daily life activities. These findings demonstrate a complex interaction between topics discussed on social media, urban morphology, and urban functions, highlighting the potential of social media data as a valuable resource for capturing the citizens’ interests or topics discussed.

ABSTRACT. The African continent is undergoing rapid urbanization, marked by the unprecedented expansion rates of its cities. In Morocco, economic growth is driving infrastructure development, paving the way for escalating urbanization, while simultaneously presenting challenges for sustainability and the wellbeing of its population. This study investigates urban heat structure, specifically the urban heat island (UHI), across three Moroccan cities—Nador, Fes, and Ouarzazate. Using modeled land surface temperatures obtained from the simple biosphere model (SiB2), this research analyzes how urban heat structure varies with the cities’ climate zones and morphological characteristics. The findings should guide urban planners and decision makers to make informed policies that address urban climate concerns.