Share
Urbanisation and population growth have led to an increasing demand for land resources worldwide. This global process of land transformation persists not only in highly dynamic urban areas in developing countries, but also in cities of the Western world. The exponential increase in impervious surfaces for establishing settlements and transport infrastructure results in negative consequences in many areas, such as increased surface water runoff and flood risk, decreasing groundwater recharge, and intensification of the urban heat island effect.
Thus, exact and area-wide estimation of impervious surfaces is of high value and must be repeated regularly to provide up-to-date information. In this regard, suitable remote sensing data offers a cost-effective solution for area-wide surveying and monitoring of impervious surfaces.
Traditional approaches for estimating imperviousness are mostly based on aerial surveys, which are conducted on demand at high cost. This data is commonly analysed by visual interpretation of the imagery, which is also very time- and labour-consuming. In contrast, modern space-borne earth observation systems provide data of different characteristics for mapping of the Earth’s surface. Commercial satellite remote sensing imagery with very high spatial resolution (VHR) of up to 30cm offers aerial-like capabilities for observation of the land surface. In addition, Europe’s Copernicus and NASA’s Landsat provides imagery with high spatial resolution of up to 10-15m at no cost. In addition, these data provide high temporal resolution (i.e. revisit times in the order of 5-16 days), which enables effective monitoring capabilities. State of the art machine learning techniques can be used on these data for semi-automatic image interpretation, to reduce temporal efforts for image analysis.
New techniques
Against this background, the Company for Remote Sensing and Environmental Research (SLU) has developed several products on impervious surfaces based on a range of different data sources with different spatial resolution. From the category of commercial satellite data with very high spatial resolution, the WorldView-2 satellite provides imagery with eight multi-spectral bands in the visible and near infrared region of the electromagnetic spectrum and a ground resolution of 50cm. Data from the European Sentinel-2 mission is available free of charge and offers 13 bands in the visible and infrared spectrum with up to 10m spatial resolution at a revisit time of five days. Finally, Landsat 8 provides freely available imagery with nine spectral bands at visible and infrared wavelengths up to 15m resolution on the ground.
Before applying the described techniques, ideally cloud-free images must be obtained from commercial satellite data providers or the data portals of NASA or Copernicus. If needed, suitable reference data can be collected by visual interpretation or from other data sources. In addition, all data must be properly georeferenced and satellite-based imagery must be atmospherically corrected and can be radiometrically enhanced.
In general, different surface materials are concentrated in a heterogeneous and highly complex manner in urban environments. In this context, very high resolution imagery with 50cm spatial resolution enables proper recognition and identification of individual objects. This allows the classification of pure pixels with regard to imperviousness, which is realised in an object-based image analysis (OBIA) approach. For this purpose, spectral, textural and contextual features are calculated based on a multiresolution segmentation in Trimble eCognition Developer software, although any other software with OBIA capabilities can also be used. These features are fed into a transferable knowledge-based classification tree to identify land cover with respect to different surface materials and subsequent binary classification of pervious and impervious surfaces.
For imagery with high spatial resolution (HR) of 10-15m, most objects in urban areas cannot be resolved and thus, pixels mostly contain mixed signatures of several objects on the ground. A solution to dissolve mixed signatures and to estimate the fraction of imperviousness is provided by machine-learning techniques such as support vector regression (SVR). This is a supervised learning method that allows non-linear modelling of a dependent variable – here, the degree of imperviousness of each pixel – using a set of independent variables (predictors). For the HR remote sensing imagery, spectral indices such as the normalised difference vegetation index (NDVI) as well as the original reflectance values are predictors for the estimation of imperviousness. For robust modelling of this relationship, about 300-400 pixels with reference information (fraction of imperviousness) are required. The SVR for impervious surface estimation is implemented as code in the free programming language R, but can be implemented in various machine learning software packages or any other programming language.
The derived products illustrate that the OBIA classification of the WorldView-2 imagery with 50cm spatial resolution offers a wealth of detail with regard to the characterization of impervious surfaces. In contrast, the estimations based on the HR imagery of Sentinel-2 and Landsat-8 still allow a general assessment of impervious surfaces at no data cost. These innovative products of imperviousness were validated against a traditional product based on a visual interpretation of an aerial survey provided by the city municipality of Munich. The VHR OBIA classification provides estimations of imperviousness with a root mean square error (RMSE) of 8.2 % compared to official data of imperviousness. The estimations based on HR data obtained RMSE values in the order of 10.7-14.7% and reveal that these remote sensing data sources still allow a general assessment with high accuracy. In general, the accuracy of impervious surface mapping is highly dependent on the spatial resolution of the input data – in other words, the accuracy generally increases with the spatial detail of the imagery.
In practice
These products can be used by users from different domains. For example, administrations of cities, municipalities, or countries can use this information for an objective documentation and monitoring of land consumption, to support land management and political decision-making.
In addition, these products support urban planning of public authorities, as well as private companies such as construction companies or planning offices. Common applications are rainwater treatment in terms of the proper dimensioning of the sewer network in general and flood risk management in case of heavy rain. Furthermore, the degree of imperviousness is an important driver of the urban heat island effect and thus, accurate spatial information provides assistance for the avoidance or the introduction of countermeasures of hotspots of urban heat.
To ensure high usability and flexibility of these information products, the impervious surface estimates are provided as raster data in standard formats that can be used with any GIS or other image processing platform.
The cost of the input data itself can be reduced dependent on the desired degree of spatial detail. HR imagery with 10-15m ground resolution can be acquired at no cost, while VHR imagery up to 30cm decreases expenses per square kilometre by a factor of two compared to traditional aerial surveys.
Another great potential for savings lies in the cost of data processing and thus labour. While semi-automated OBIA methods for the analysis of VHR data require some effort, machine learning techniques such as SVR can be applied almost fully automated if suitable reference information is available for training.
Tobias Leichtle is an image analyst and research associate at the Company for Remote Sensing and Environmental Research (SLU) ([email protected])
