National-scale imperviousness mapping and detection of urban land changes

Abstract

With high land take for urbanization a serious concern, the European Union and Member States have set quantitative targets for limiting land take in Europe. Consistent and reliable information on land take is crucial for measuring progress in achieving these targets. In the big data era, Earth observation data offers opportunities to regularly and consistently monitor land use change dynamics on a national scale. However, often-used methods such as image classification do not solve the mixed pixels problem common to urban areas with high heterogeneity. Moreover, the urban land use changes detected by these methods are often affected by random noise caused by non-relevant changes in atmospheric conditions, sun angle, vegetation phenology, etc. Therefore, the resulting products are often judged as being of insufficient quality for operational application.In this study, we present an approach to measuring annual land take using imperviousness mapping and change detection. Based on the Google Earth cloud computing platform, we first applied a data mining approach on annual Sentinel-2 image scenes to extract temporal and spectral information at Germany wide scale. Then spectral unmixing algorithm was used to estimate sub-pixel fractions of imperviousness. We end by using imperviousness changes to detect newly built surfaces. The change detection products were validated using high-resolution historical aerial photos and images. The results achieved an accuracy of 87% for detecting new built-up and 98% for detecting changes. The difference between the two numbers is caused by the areas with changes on the surface but not for urban use.We found that the integration of temporal information improves the performance of spectral unmixing than using spectral information alone, because it can better distinguish imperviousness from other land covers with similar spectral signatures but different temporal characteristics. We also found that, due to the sub-pixel analysis approach, spectral-unmixing-derived imperviousness is sensitive to urban features even with a size of just a few pixels. Moreover, the imperviousness layer represents urban land with a continuous index, meaning that random noise, represented as a low-range change of fractions, can be filtered out. Our study demonstrated that remote sensing products can be tailored to fulfil the needs of specific applications supporting sustainable land-use policy. This opens up more possibilities for the diverse and growing user communities of urban land change products.