Conference Agenda

Urban areas are central to a range of sustainability challenges, ranging from land take and the loss of natural habitat to the exposure to climate change impacts. As a result, a large number of models have been presented that project future urban land change under different scenario conditions. The sustainability of future developments depends strongly on the characteristics of urban land, including for example the density, the building material, and the presence of slums. Yet, most of urban change models include only one type of urban land, corresponding with areas recognized as built-up land in satellite imagery. This severely limits their relevance for understanding urban development in general and for sustainability assessments specifically. Here, we present an urban growth model that includes multiple classes along the rural-urban gradient and that is applied to model urban development in five countries in Southeast Asia. In this model, change is driven by future population developments, but these can lead to both expansion and densification, depending on the scenario. In addition, we characterize the different types of urban land in terms of their building materials and demonstrate the relevance of this characterization for assessing future risk for river flooding. Interestingly, the expansion scenario leads to a lower expected annual damage in Lao PDR and Cambodia, while the densification scenario leads to a lower expected annual damage in Myanmar, Thailand, and Viet Nam. Finally we point at the opportunities that recent developments in remote sensing provide for better characterization of urban areas, and thus also for more relevant urban models.

In the context of GDA-APP (Global Development Assistance - Analytics and Processing Platform) project, a segmentation model has been developed in order to identify urban areas using Sentinel-2 composite image (NRGB). The objective is to create a fully-automatic module to delineate urban area for instance for change detection, informal settlements detection, damage infrastructure assessment, etc…

The modeling pipeline follows several training steps. The first step is corresponding to a pre-training for segmentation using BigEarthNet dataset ~500.000 patches of Sentinel-2 as input and its respective ESA Worldcover building layer as reference: in this way the model will learn how to interpret Sentinel-2 imagery with the objective of detecting urban area. The second step is to improve detection to building scale by using pairs of image of Sentinel-2 composite and Microsoft building layer corresponding mask with ~30.000 patches. The third step improves delimitation capabilities by using a similar dataset but by enhancing spatial resolution from 10m to 3m using a super resolution model. All model versions are made in order to be responsive to any type of urban landscape and performing the segmentation at a global scale; indeed selected images are located over various areas distributed worldwide to make it as much general as possible. Moreover the complexity of the task requires several training steps to go deeper in precision, in terms of spatial resolution and so in building delimitation. The overall architecture corresponds to a UNet composed of 34 layers with ResNet blocks, and optimized with the use of loss function such as focal loss, architecture and functions well known in machine learning for segmentation application. Current model is trained with 10m pairs of images (cf step 2), its performance is estimated through accuracy metric, giving an overall score of more than 95% however the real interest is find in the accuracy of buildings detection, which is approximately to 65%. By following each steps described above, pre-training and final learning with 3m spatial resolution imagery are promising for improve results.

Monitoring Sustainable Development Goal (SDG) indicator 11.3.1 on land use efficiency is crucial for sustainable urban development. This indicator, which assesses the ratio of land consumption rate to population growth rate (LCRPGR), helps urban planners and policymakers optimize land use and curb unsustainable urban sprawl. Efficient land use supports several interlinked SDGs by promoting denser urban living, reducing cities' ecological footprints, enhancing livability, and preserving natural resources. Thus, monitoring this indicator is vital for ensuring sustainable growth and improving quality of life in urban environments.

Earth observation (EO) time series offer a powerful tool for monitoring urban land use efficiency, providing key insights into the dynamics of urban expansion and development. This presentation will first introduce a state-of-the-art method for urban mapping at global scale, leveraging open EO data such as Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 MultiSpectral Imager (MSI) data and deep learning. Then it will showcase the SDG11.3.1 LCRPGR results in 500 cities during 1985 to 2020 using existing built-up datasets derived from Sentinel-1/2 and the extensive archive of Landsat long time series. The findings demonstrate that while population datasets from various sources minimally affect LCRPGR calculations, different built-up datasets lead to significant variances. Moreover, challenges arise when employing improved built-up products derived from Sentinel-2 data together with the built-up products derived from Landsat data, impacting the LCRPGR calculations. Consequently, this has led to many cities showing a negative LCRPGR during 2015-2020, inaccurately suggesting a decline in density. These results highlight the urgent need to enhance the accuracy and consistency of built-up data, given its significant role in determining the values of SDG 11.3.1 Land Use Efficiency.

The Sustainable Development Goals (SDGs) are a universal agenda to address the world’s most pressing challenges. Robust monitoring mechanisms and timely, accurate and comprehensive data are essential in guiding policies and decisions for successful implementation of the SDGs. Yet current ways of monitoring progress towards the SDGs such as through household surveys cannot address the SDG data gaps and needs. Along with EO data, citizen science offers a solution to complement traditional data sources.

The complementarity of citizen science and EO approaches for SDG monitoring has also been acknowledged in the literature. For example, the authors of this contribution carried out a systematic review of all SDG indicators and citizen science initiatives, demonstrating that citizen science data are already contributing and could contribute to the monitoring of 33 per cent of the SDG indicators. As part of this review, they also identified overlap between contributions from citizen science and EO for SDG monitoring. One specific citizen science tool integrating citizen science and EO approaches that could complement and enhance SDG monitoring is Picture Pile. Picture Pile is a web-based and mobile application for ingesting imagery from satellites, orthophotos, unmanned aerial vehicles or geotagged photographs that can then be rapidly classified by volunteers. Picture Pile has the potential to contribute to the monitoring of fifteen SDG indicators covering areas, such as deforestation, post disaster damage assessment and identification of slums, among others, which can provide reference data for the training and validation of products derived from remote sensing.

This talk presents the potential offered by Picture Pile and other citizen science tools and initiatives focusing on urban applications to complement and enhance official statistics to monitor several SDGs and targets including SDG 11 Sustainable Cities and Communities. Recommendations will also be provided for how to enable partnerships and collaborations across data communities and ecosystems in order to mainstream citizen science and EO data for SDG monitoring and reporting of urban issues.

The rapid urbanization combined with insufficient low-cost housing supply in the Global South results in the proliferation of informal settlements. Our study addresses the pressing need for accurate information by User and Data-centric Artificial Intelligence (AI)-based methods for mapping informal areas. In collaboration with local communities and several (inter)national stakeholders, we co-designed an AI-workflow based on free-cost Earth Observation (EO) and geospatial data to map informal areas in eight cities across the globe. Together with our Early Adopter and Stakeholders, we identified local data questions and validation needs. For example, in Buenos Aires, the main question is to identify unmapped settlements (compared to the national registry). Presently, we are evaluating the performance of the Multi-Branch Convolutional Neural Network (MB-CNN) model across our pilot cities. The model takes Multispectral and SAR data in combination with additional geospatial features (e.g., building density) as input and delivers the segmentation map. This is done in preparation for increasing the reference data quality via our user portal. The initial model results (for informal areas) show for the city of Nairobi an F1 score of 0.79 and for Medellin an F1 score of 0.75. Mumbai, Salvador and Buenos Aires also performed relatively well, with F1 scores of 0.70, 0.61 and 0.57. In contrast, the model performs below 0.5 in Jakarta, Lagos and Mexico City (more complex cities in terms of morphology and reference data). The results indicate that the fusion of datasets provides higher accuracy than using only S-2. This can be attributed to the additional contextual information, which aids in more accurate identification and mapping. Cities with large and densely built-up settlements (e.g., Medellin, Mumbai, Nairobi) significantly improved when building morphology was added. Next, we will be improving reference data by crowdsourcing via our portal.

Informal settlements are the primary residential areas for the urban majority. Despite their significance, a lack of accurate data and often outdated records pose significant challenges for public entities in meeting residents' needs. Consequently, our understanding of the spatial distribution, evolving patterns, and growth over time of these settlements remains limited. However, emerging remote technologies offer new opportunities for mapping and collecting high-quality data. This presentation focuses on the utilization of urban data and AI/ML-based mapping technologies to gain deeper insights into the complex realities of informal settlements. In this context, we will present UNITAC's current research and introduce our work on the Building & Establishment Automated Mapper (BEAM), specialized software for AI/ML-based detection of building footprints in informal settlements

The BEAM tool, developed by UNITAC in partnership with eThekwini Municipality in South Africa, leverages AI to map rooftops in high-resolution aerial photography. It meets a crucial need identified by the municipality for automated mapping processes, crucial in a city where a quarter of residents reside in informal settlements, and census data updates occur only every 10 years. This results in a lack of accurate data about its population, particularly in informal settlements, complicating infrastructure planning and human settlement upgrades. Additionally, the BEAM tool is being upscaled for use with satellite imagery (currently it works on high-resolution aerial photography) in eight Central American cities. Consequently, the tool underscores its innovative support for city governments in the global South, leveraging machine learning to assist these governments in addressing key sustainable development goal targets and indicators, as well as tackling the urgent challenges of urbanization.