Construction activities detected by Sentinel-1 and -2 from space

Kurzfassung

The construction sector has been booming recently, construction sites seem to arise everywhere around the urban or industrial agglomerations. This results from the excellent economic situation as well as the low interest rate policy of the recent years. The good order situation is expected to continue in the near future. Not even the pandemic situation in Europe could stop the construction boom. Furthermore, the amendment of the building regulations, with accelerated approval procedures, will presumably boost residential construction, according to the relevant ministry. This results in an increasing number of building permits and due to a missing automated connection between the building activities and the surveying of buildings by the Bavarian Agencies for Digitisation, High-Speed Internet and Surveying (ADBV), there is a high increase in the workload of the employees. In order to keep an overview and to create a relief of the employee’s new methods for the investigation of construction activities are explored, e.g., automated services that recognize new buildings from remote sensing data. Hence, time-consuming methods such as the visual comparison of the True Digital Orthophoto with the digital cadastral map could be replaced or at least much simplified when possible construction hotspots are already marked by the detection service. In this context, the following question arises: Do time series of the freely available Sentinel-1 and -2 images, processed to analysis ready data (ARD) in the extended Kennaugh framework, allow for the detection of newly built residential houses? We present the prototype of an automated service that continuously evaluates the acquisitions of Sentinel-1 and Sentinel-2 using a temporal Wavelet analysis in the extended Kennaugh framework of the MultiSAR System for the detection of newly built structures. It reports on an optimized post-classification data fusion approach to join the weekly Sentinel- 1 image and (due to cloud coverage) less than weekly Sentinel-2 image over Europe. Although the direct SAR-OPT image fusion was feasible in the extended Kennaugh framework to perform a pre-classification data fusion, we decided in favour of a weekly Sentinel-1 evaluation whose results are further stabilized by the evaluation of a Sentinel-2 acquisition whenever available. The whole process is designed to be implemented as asteady service that highlights construction activities over large areas. Thanks to the most sophisticated pre-processing to ARD, e.g. total backscattering intensity of Sentinel-1 or the newly developed LEAI for Sentinel-2, this approach needs neither training data nor computationally intensive machine learning algorithms. The validation of the individual time series of Sentinel-1 and -2 with ground truth data collected during a survey showed accuracies of up to 90% whereas the missed hits were residential houses smaller than the spatial resolution of the satellite images. In our study, Sentinel-1 and -2 satellite imagery of the European Copernicus program from begin of October 2019 to end of November 2020 are processed. The study area is located in the northeastern part of the Bavarian town Dingolfing where numerous residential buildings are currently under construction. The need for more living space arises from the near car factory of the ’Bayerische Motoren Werke AG’ (BMW). Thanks to the weather-independent imaging capability of the Synthetic Aperture Radar system of Sentinel-1, a time series of 67 images is used and for the multi-spectral system of Sentinel-2, due to the strong influence of the atmosphere and clouds, 46 images are considered. Before creating the time series, the optical data are transformed into Kennaugh-like elements based on the method of an orthogonal transform on hyper-complex bases and are decomposed into the individual Kennaugh elements in the same way as the polarimetric images processed by the MultiSAR System processor of DLR. Furthermore, a novel multi-spectral index, the LEAIndex (LEAI), is developed from the signatures of the Kennaugh-like elements. The peak in the temporal signature caused by the construction of a new building is extremely amplified using the LEAI in comparison to using the normalized Kennaugh-like elements or using the colour channels directly, which would be the standard way. Therefore, the LEAI is derived from the Kennaugh-like elements for the whole data cube. The LEAI time series of Sentinel-2 and the ones of the total intensity of Sentinel-1 are convolved with different scales of the Haar wavelet kernel. Applying the Haar wavelet, local changes in the time series are enhanced in a robust and reliable way so that the construction of a new building can be detected as a minimum in the temporal signature. Comparing different scales of the Haar wavelet enables the determination of the minimum number of time steps needed for a reliable detection of a new building. Afterwards, a simple threshold classification is applied to these Wavelet amplified data. The threshold is determined by a sophisticated pre-classification validation approach examining the completeness and correctness at the same time. The reference data can be seen as real ground truth data because it was provided directly by the house owners and collected in a local survey. An accuracy of up to 90 % is achieved with only the Sentinel-1 data set, combined with the Haar wavelet kernel on a scale that includes ten acquisitions. Using only Sentinel-2 data set shows an accuracy of up to 87 %, but due to the irregular intervals within the time series, it is necessary to extend the Haar wavelet to a scale covering twenty acquisitions. Especially by delimiting the data to the areas relevant to the construction of buildings, the developed method proves to be an excellent way to capture newly built residential houses from remote sensing data of the Copernicus mission. In summary, the LEAI for Sentinel-2 and the total intensity for Sentinel-1 combine all necessary information todetect the construction of a new building with an expected accuracy of more than 90%. In combination, the regular time series of Sentinel-1 providing structural information is stabilized by the time series of Sentinel-2 providing spectral information. A certain time gap concerning the detection by Sentinel-1 and Sentinel-2 was observed and could be explained by the collected ground truth data. Sentinel-1 reports on the erection of the first wall whereas Sentinel-2 highlights the foundation of the base plate. This is an important fact that justifies the decision for a post-classification data fusion in retrospect. Based on these two multi-temporal data stacks and the known time gap of the respective peak in the temporal signature an automated process can be implemented. The original idea was to transfer the approach to the whole of Bavaria in order to support the ÄDBV. In general, even the extension to complete Germany or neighbouring countries would be feasible. From the methodological perspective, the time series analysis tested and validated on the ARD cube should be transferred to the original Level 1 data for sensitivity reasons. In this way, even the finest reasonable ground sampling of Sentinel-1 Interferometric Wide Swath Mode with about 5 by 20 metres could be investigated in order to refine the detection of smaller houses. Furthermore, the smoothing effect of the Haar wavelet is increased. From a user’s point of view, the output format should be reworked, i.e., one must evaluate whether vector or raster format is preferred or which semantic resolution is wished, e.g. only binary (yes/no) or given as probability. In the optimal case, the detected building is written to a database for further handling in the competent authority. For now, we can resume that the automated detection of construction activities on residential buildings from satellites of the Copernicus mission is feasible with a really high accuracy.