MUAS 2018 > Session details
Paper 120 - Session title: Methods of Urban Mapping (continued)
11:30 Sentinel-1 Enhanced Resolution Image by Deep Learning with TerraSAR-X
Ao, Dongyang (1,2); Dumitru, Corneliu Octavian (1); Datcu, Mihai (1) 1: DLR, Germany; 2: Beijing Institute of Technology, China
To improve the quality of SAR images, we proposed to train a deep neural network with TerraSAR-X. This is done by using a Dialectical Generative Adversarial Network (Dialectical GAN) to generate high-quality SAR images. This method is based on the analysis of hierarchical SAR information and the “Dialectical” structure of GAN frameworks. As a demonstration, a typical example will be shown where a low-resolution SAR image (e.g., a Sentinel-1 image) with large ground coverage is translated into a high-resolution SAR image (e.g., a TerraSAR-X image). Three traditional algorithms are compared and a new algorithm is proposed based on a network framework by combining conditional WGAN-GP (Wasserstein Generative Adversarial Network - Gradient Penalty) loss functions and spatial gamma matrices under the rule of dialectics. Experimental results show that the SAR image translation works very well when we visually compare the results of our proposed method with the selected traditional methods.
Translation of Sentinel-1 data to TerraSAR-X image resolution has attracted great interest within the remote sensing community. First, the high resolution of TerraSAR-X generates SAR images rich in information that allow new innovative applications. Second, the wide area coverage of Sentinel-1 images reduces the need for multiple acquisitions, and decreases the demand for high-cost data. Third, it is much easier for researchers to access Sentinel-1 images than TerraSAR-X images because the Sentinel-1 images are freely available, while the TerraSAR-X images are usually commercial.
For validation, we used images of urban areas, so we can apply a spatial matrix to extract geometrical arrangement information. Our method learns an adaptive loss function based on the image pairs at hand, and is regularized by the prescribed image style, which makes it applicable to the task of SAR image translation. Based on the advantages of using a GAN, we have achieved very good results with detailed visual effects demonstrating that our method is better than the existing traditional methods being compared in our presentation.
Paper 124 - Session title: Methods of Urban Mapping (continued)
12:10 Out and Up: Monitoring Urban Growth in the Global South
Gevaert, Caroline Margaux (1,2); Anderson, Edward (2) 1: University of Twente, Netherlands, The; 2: World Bank Group
Rapid urbanization is challenging urban planners in many developing countries. Earth Observation has the potential to monitor this growth and support urban planners in developing suitable response mechanisms. Cloud-free imagery acquired through Earth Observation can be visually interpreted to identify changes. Machine learning algorithms such as deep learning can process these images to automatically identify buildings to characterize growth patterns. To do this, adequate training samples are needed, which can be obtained for example through participatory mapping initiatives. The third level of interest for the urban planner is the Floor Space Index (FSI), essentially stating the building footprint times the number of floors per plot. The FSI can be used by urban planners to support population estimates, building valuation estimations, and disaster risk management exposure data.
So given recent developments in Earth Observation such as (1) larger satellite constellations providing more frequent imagery and (2) the proliferation of Unmanned Aerial Vehicles providing more detail, how close are we in practice to providing the urban planner in an African city with: cloud-free imagery, building footprints, and FSI as a service? Specific challenges of the Dar es Salaam case study include frequent cloud coverage and the typology of the urban area. Most of the population lives in informal settlements, whose typical settlement characteristics (small buildings, narrow footpaths, irregular construction materials) make them challenging to classify in satellite imagery. This submission compares the results of a recent study funded by the World Bank and executed by Planet Labs which focusses on these three levels to the information which can be generated from UAV imagery.
Cloud-free, 3.7 m spatial resolution Planetscope mosaics were obtained using the Planetscope satellite constellation for the entire city in 2017. A large-scale participatory mapping project (Ramani Huria) produced the OpenStreetMap building labels which served as training samples. A standard U-NET architecture was trained using 70% of the training labels, obtaining an F1 score of 77%. Many false positives appeared in the informal areas, where the spatial resolution was likely not clear enough to distinguish individual buildings. A smaller study area used a stereo pair 0.8 m Skysat imagery to predict the building height and FSI of the downtown area.
Secondly, a similar study was conducted using UAV imagery obtained in 2017 over the same area. ESA’s Urban Thematic Exploration Platform (U-TEP) processed the imagery into an orthomosaic, Digital Surface Model, and Digital Terrain Model. The higher resolution of these images can facilitate the extraction of the information from informal areas.
Comparing the results of the two FSI estimates provides novel evidence regarding the trade-offs between high-cadence satellite constellations (covering a larger area with fewer operational costs) to UAV imagery (higher spatial resolution, higher agility in image acquisition, but also higher operational costs). Comparing the similarities and locations of differences of both results indicates the importance of spatial resolution and vertical accuracy for urban monitoring applications in developing nations, providing valuable input for the next generation of R&D activities.
Paper 145 - Session title: Methods of Urban Mapping (continued)
11:10 Evaluating Sentinel-2 imagery for mapping human settlements
Schug, Franz; van der Linden, Sebastian; Okujeni, Akpona; Hostert, Patrick Humboldt-Universität zu Berlin, Germany
Mapping urban areas with remote sensing is particularly challenging due to the spatial spectral heterogeneity of human settlements and their complex dynamics related to land cover change. Sentinel-2 offers improved spatial, temporal and spectral resolution compared to similar globally operating optical systems and is, therefore, a highly interesting data source for urban monitoring. In this regard, efficient workflows are desired to make best use of Sentinel’s rapidly growing archive. Moreover, the contribution of Sentinel-2, as a globally available source, to the existing state-of-the-art in urban mapping must be better understood. Especially its complementarity to existing urban mapping products is of interest.
In this study, we used machine learning regression to produce an impervious surface fraction layer for Berlin, Germany, and surrounding areas from Sentinel-2 data. For model training, we used synthetic mixtures generated from an image spectral library representing pure surface cover types. We apply support vector regression (SVR) in order to efficiently and accurately produce surface cover fractions. Based on these fractions we performed a systematic quality assessment of our product in comparison to existing continental and global scale settlement maps, such as
- the decameter resolution Global Urban Footprint (GUF) based on TerraSAR-X data,
- the Global Human Settlement Layer (GHS) based on Landsat imagery and other open source data,
- the European Settlement Map (ESM) with a focus on built-up vs non-built-up areas based on SPOT imagery and the EEA Urban Atlas,
- and the Copernicus Imperviousness (CopImp) product.
Our library-based approach is independent from the tedious spatial definition of training areas and appears promising for the generalized mapping of built-up areas across multiple cities. For validation, we defined 13 regions of interest in neighbourhoods with different characteristics and performed a pixel-wise comparison with cadastral reference data. The accuracy of the SVR product is largely similar to that of the other evaluated datasets with regard to above-ground human infrastructure. As also reported in Klotz et al. (Rem. Sens. Of Env., 178, 2016), we find that GUF and GHS are particularly useful to map dense settlements, but systematically overestimate urban coverage in low-density built-up areas. Whereas accuracies for dense urban areas are similar among the datasets, GUF and GHS feature lower user’s and producer’s accuracies than SVR and ESM in some neighbourhoods. SVR is comparable to CopImp in terms of overall and area-specific RMSE values.
The challenge of conceptually delineating urban space is important. GUF, GHS, ESM and CopImp offer accurate information with regard to their respective objective. Beyond the classic mapping of densely built-up urban areas, machine learning regression might be particularly beneficial in areas with low built-up density and thus complementary to existing products. Using free Sentinel-2 data, SVR is potentially able to globally map settlements based on dense time series that also allow monitoring dynamic urban developments. Additionally, its ability to provide gradual land cover information is useful to better understand urban patterns. We therefore conclude that there is great potential of mapping a diversity of human settlements with Sentinel-2 when employing an SVR-based approach.
Paper 161 - Session title: Methods of Urban Mapping (continued)
12:30 A New Approach to Property Valuation Using Remote Sensing Data and Machine Learning in Kigali, Rwanda
Bower, Jonathan (1); Braun, Andreas (2); Brimble, Paul (3); Bachofer, Felix (4); McSharry, Patrick (5) 1: The International Growth Center; 2: The University of Tübingen, Germany; 3: Ministry of Finance and Economic Planning, Rwanda; 4: German Aerospace Center; 5: Carnegie Mellon University - Africa, Rwanda Campus
This study trials a property valuation methodology for Rwanda's capital, Kigali by applying machine learning techniques to parcel transactions data from 2013 to 2015 and remote sensing data from 2009 and 2015 on building footprints. The building information (building footprint and building type) is derived from aerial images (2009) and a Pleiades scene (2015). Additionally, spatial parameters describing the spatial interrelations of properties within the city are derived from a multi-source dataset and include distance-based values (distance to CBD, schools, public transport, etc.), accessibility, location parameters (topographic position) and neighbourhood statistics (green area, urban structure type, density values, etc.). This approach can help to understand the key determinants of land and building values and be used to create a database with accurate estimates of property values for each parcel in Kigali. This database will be used to provide the local tax administration with an important tool for collecting property tax revenues as the main method of property valuation in Rwanda is through taxpayer self-assessments. In order to support the tax administration in determining the validity of these self-assessments, the estimates from this database can be used to trigger certified counter-valuations if necessary with the aim of raising domestic tax revenues. Therefore, this study highlights how remote sensing methodologies can be effectively used to complement traditional valuation techniques through an application in Rwanda.
Paper 175 - Session title: Methods of Urban Mapping (continued)
11:50 Predicting Slum Dwellers’ deprivations from Space
Soukup, Tomas (1); Kolomaznik, Jan (1); Borja-Vega, Christian (2); Mimmi, Luisa M. (2); Patel, Amit (3) 1: GISAT s.r.o., Czech Republic; 2: World Bank Group, USA; 3: University of Massachusetts Boston, USA
One of the most pressing development challenges we face in today’s world of rapid urbanization is how to respond to the unmet demand for basic infrastructure services, like adequate housing, clean water, and sanitation. Half of the world’s population lives in cities and close to 1 billion live in slums. Megacities in developing countries are growing faster than ever, mostly in an unplanned way. Dhaka, the capital of Bangladesh, is particularly challenged due to congestion, poor infrastructure and regular flooding during heavy rainfall.
When looking at approaches to impact of ways to upgrade slums, a challenge is often a lack of adequate data, particularly a lack of spatial data. Standard mapping rarely keeps up to cover such rapid growth of unplanned settlements. Conducting household surveys are useful, but they are expensive and thus not frequently updated. This is a problem when trying to monitor dynamic and elusive slum, whose very definition is controversial and context-specific.
So what is the solution? It turns out that using Earth observation (EO) data – may help to address this critical data gap. It is already being used in many other applications, from recognizing cross-border territories in disputes, tracking wildlife, locating hazards and disasters, to identifying migration patterns.
Over the past two decades, the temporal and spatial resolution of earth observation imagery has increased dramatically. As an example, tools like Google Maps, now became ubiquitous in our daily lives. Along with improved imagery, advanced algorithms are being developed which can be used to detect and describe slums, answering critical questions:
Where are slums/informal settlements located?
How do their appearances change over time?
Do they present spatial characteristics incompatible with basic infrastructure supply (living space, water and sanitation, roads and safety from natural hazards) supporting insight into service deprivation patterns?
With an eye towards using innovative data analytics for development impact in slums, World Bank’ Water Practice (GWAGP) launched a pilot in Dhaka under the ‘Water Supply and Sanitation in Rapid Urbanization’ umbrella about a year ago. In collaboration with the WASH Poverty Diagnostic team in Bangladesh, the remote sensing services company GISAT, a member of an EO4SD Urban consortium working for the European Space Agency (ESA), and researchers from the University of Massachusetts Boston multidisciplinary work has been carried out on this topic. Results from Phase 1 of the study acknowledged by the bank stakeholders will be presented, explaining how the information from in-depth local household surveys can be integrated with geoinformation captured remotely for Dhaka slums and used in service deprivation modelling.