Data Analytics

Today, we are faced with a number of demanding scientific and technical challenges that - if resolved successfully - will lead to advanced and highly demanded applications of Earth observation data. The given challenges arose in many fields of remote sensing and data interpretation and have to be tackled by advanced methods and paradigms. Prominent examples range from new satellite sensor concepts and computational sensing via efficient distributed data processing on ground up to automated data interpretation and knowledge extraction for individual end users. We will give a survey of innovative sensor concepts, new approaches of how to combine remote sensing data with geomatics, future prospects for cloud-based data handling and services, and data interpretation by (deep) learning and interactive visualization. These approaches will be complemented by assessments of typical new user-oriented applications that are expected to result from each new paradigm. In particular, we will address the prospects of computational imaging, the advantages offered by exploiting compressed data or of compact descriptors based on selected metadata, the potential of extracting high-level semantic information by combining sensor data with already existing knowledge contained in publicly accessible databases, and data trend investigations based on cloud-computing concepts together with advanced algorithms for the analysis of image time series. Finally, these approaches will be compared with totally new ideas resulting from the introduction and application of quantum computing. In order to assess the feasibility and the application potential of each proposed new approach, we will provide typical application scenarios based on the parameters of the current series of European Sentinel satellites and its stakeholder community. This does not only include the description of algorithms but will also consider the design of user interfaces and analysis tools such as the provision of standardized data analysis platforms. This results in a global scenario of future activities for the remote sensing community at large.

In many cities the population is exposed to elevated levels of air pollution. Often the local air quality is not well known due to the sparseness of the official monitoring network, or unrealistic assumptions being made in urban air quality models (such as vehicle emission factors). However, new sources of alternative air quality data rapidly become available. From space by satellite instruments like TROPOMI (providing air quality information on a 7 x 3.5 km2 resolution), and on the ground by new sensor technologies allowing for low-cost in-situ measurements. Numerous research groups, companies, and citizens are already experimenting with these low-cost sensors. The objective of the RETINA project at KNMI is to develop operational services able to produce high spatio-temporal resolution maps of urban air pollution. This is not straightforward due to the localized nature of pollutants such as nitrogen dioxide (NO2). With a new data assimilation approach we combine the heterogeneous measurements with atmospheric dispersion models, making optimal use of all available information. In a preliminary study, we assessed the in-field performance of low-cost NO2 sensors in a citizen science campaign in Amsterdam. We found that the current generation of sensors can provide useful data (with an error around 7 μg/m3), but only after extensive calibration efforts. The second step was to develop a prototype system for the city of Eindhoven, where an alternative (mid-cost) air quality network is operational since 2015. We implemented an atmospheric dispersion model, driven by emission proxies from open data sources (e.g. road network, traffic intensity, population density). With regression techniques we find the emission factors for the best overall model performance, while we use Kalman filtering to adjust the model results locally. The developed techniques in the RETINA project are sufficiently versatile to be applied to other pollutants (such as particulate matter) and to data from other sensor networks in other cities. The RETINA system will first be used to better understand the satellite observations of air pollution within an urban area. In a later stage, the satellite observations will be used as an additional data stream for assimilation in the system, and thus provide better air quality information for cities which lack a representative official monitoring network.

Sentinel-2 satellites, with its global coverage and short revisit time, acquire a dataset ideal for observation of land surface changes. However, automatic or semi-automatic solutions are very susceptible to atmospheric fluctuations and become erratic due to large amount of "false positives" – algorithms marking cloudy data as changes. Even though Sentinel-2 data are available for almost two years, there still does not exist a single scene (non multi-temporal) cloud detection mechanism with high accuracy, suitable for on-the-fly processing. Trying to find a globally sound solution, we have tested several existing approaches, starting with the default cloud masks, which are part of L1C products, Sen2Cor classification, widely used Fmask and other single scene and multi-temporal algorithms. In order to obtain our own algorithm, fast enough to be used for on-the-fly cloud detection, we have employed a machine learning process, based on classical Bayesian probability estimation approach. The learning process used the cloud masks of each of the cloud detection algorithm as a separate training dataset to obtain a set of parameters for our algorithm. Clouds detected with our algorithm were validated against manually classified data, originating from a repository of manually labeled Sentinel-2A spectra (Hollstein et al.) and our own in-house cloud database, for each trained set of parameters. Finally, we have integrated the subsequent procedures within our Sentinel Hub services to be able to identify and visualize cloudy areas on-the-fly. The benefits of having an accurate and compute-optimized cloud detection are enormous - starting with filtering scenes based on localized cloud coverage, creating dynamic cloudless mosaics, having good change detection results, etc. We will present our findings on the cloud detection algorithms, compare results from different training datasets and show the classifications on Sentinel Hub.

With the advent of Sentinel 1 and 2 satellites together with the Landsat constellation, dense satellite image time series with a high spatial resolution (up to 10m) are now available. Methods for analysing full and dense time series, which were previously only applicable to medium and coarse spatial resolution time series, are becoming applicable on satellite image time series that provide high spatial details. This offers a great opportunity to explore the full potential of time series analysis for ecosystem monitoring. However, this opportunity comes with challenges and requires new methods that can efficiently handle dense satellite image time series of multiple satellite sensors (e.g. Landsat, Sentinel-1 and 2), and are able to perform temporal analysis while accounting for a spatial context. This would enable the monitoring of land surface dynamics, disturbances, and extremes at unprecedented detail. We present an overview of open-source software, with emphasis on R packages, that have been developed for satellite data-based change monitoring (e.g. deforestation and regrowth monitoring), inter- sensor calibration and fusion, and land cover monitoring. Based on this overview, we highlight current challenges and needs in the current context of big Earth Observation data analysis with dense high spatial resolution satellite image time series available globally. With support from the R Consortium, a new R package called stars will try to address a number of these challenges. Potential solutions enabling more transparent, open and reproducible earth observation research are formulated.

To capture the complexity of urban environments and the dynamicity of the processes transforming them (e.g. urban sprawl, rural-to-urban transformation, land-use change), data scientists rely on the accessibility of open geospatial data. Ideally, these should be reliable, of suitable spatial resolution, up-to-date and interoperable. Local authority data repositories and institutional geoportals are increasingly becoming the most preferred free and open data-sharing resources. Depending on the type of urban application, the data mining exercise can also include unconstructed data collated from the web and social media. However, these data frequently require manipulation and transformation to be ready for use. In this arena, free-access EO data from satellite missions (e.g. Sentinel-1A/B and 2A/B) and EO-derived mapping products (e.g. the Copernicus Land Monitoring Service) are key resources for change detection and time series analysis in urban applications, of such renowned value that the challenge for scientists and practitioners is how to generate metrics and indicators to address user-oriented (or better, user-defined) questions and/or real-world issues. In light of the recognised need for up-to-date and relevant examples of EO data use in R&D activities (Byfield et al., n.d.) and the challenging topics identified by the EO Open Science Community (Snik, 2015), this paper aims to share research experiences of urban remote sensing applications based on analytics integrating both free and non-free EO and geospatial data. In particular the paper showcases approaches of data analytics developed for: (1) geospatial analysis of shallow geohazards in dense urban environments, and (2) generation of mapping products showing constraints and opportunities for sustainable use of land. Drawing from case studies in Italy, value and limitations of open data from Copernicus services (e.g. Urban Atlas), web data mining and city data portals are discussed with regard to post-processing chains of ground deformation estimates retrieved from multi-interferogram processing of Synthetic Aperture Radar (InSAR) Big Data. Open data are therefore exploited as informative layers to reduce the redundancy of hundreds of thousands of InSAR observations, understand their cause-effect relationships and narrow down to the most relevant areas of concerns that the potential stakeholder should focus on. Potential of data analytics combining high resolution geological data, Landsat and Sentinels time series, and free land cover/land use data are instead explored with regard to the provision of land indicators for the Sustainable Development Goal 11 ‘Make cities and human settlements inclusive, safe, resilient and sustainable’. More specifically, the integration with proprietary geological data is used to test the transformability of EO and EO-derived open data into a mapping tool that can inform decisions on strategic planning of green/blue infrastructure in cities. References Byfield, Val, Kapur, Ravi, Del Frate, Fabio, Mathieu, Pierre-Philippe, Higgins Mark (n.d.) EO Open Science 2.0 - Training a new generation of data scientists. White paper, http://esaconferencebureau.com/docs/default-source/15c12_library/training-a-new-generation-of-data-scientists.pdf?sfvrsn=0 Snik, Frans (2015) Summary of the "jam session" during the EOscience2.0 workshop in Frascati 12-14 Oct 2015. http://esaconferencebureau.com/docs/default-source/15c12_library/summary-of-the-jam-session-.pdf?sfvrsn=0