Methods Sentinel 2

The snow cover monitoring is important to understanding the climate system and the terrestrial water cycle. In mountain regions it is is also useful for operational agencies in charge of the water resources management and the winter tourism. New Sentinel-2 data are expected to bridge the gap in observation frequency between mid to low resolution snow products (typically derived from MODIS) and high spatial resolution snow products (typically derived from Landsat). Here we present an ongoing effort to generate a snow product at 20 m resolution every 5 days in mountain regions using Landsat-8 and Sentinel-2 data. The goal is to distribute in 2016 the first snow products at least for the French Alps and Pyrenees using Sentinel-2A and Landsat-8. The snow detection is performed using flat-surface reflectances images generated by the MACCS level 2A processor (Hagolle et al., 2015). The algorithm is based on the Normalized Difference Snow Index (NDSI) and a digital elevation model. A first pass of snow detection is performed based on thresholds in the NDSI and in the red band in order to compute a snow-line elevation at the scale of the input tile (110 km by 110 km). Then a second pass is performed with less strict thresholds for the pixels located above this elevation. The original cloud mask provided with level 2A data at 200 m resolution is also revisited to maximize the snow/no snow detection.  The algorithm was benchmarked using SPOT 4 (Take 5) and Landsat-8 images over the High-Atlas in Morocco, the Pyrenees, and the French Alps. The snow cover product is supported by the French Space Agency (CNES) and developed in the frame of the Theia Land Data Center which is a French national inter-agency organization designed to foster the use of images issued from the space observation of land surfaces. Theia is delivering quality controlled products (annual satellite coverage at national scale, surface reflectance time series) and also services through Scientific Expertise Centers (SEC) which are groups of laboratories conducting research work and developing  innovative methods .

Reference

Hagolle, O.; Huc, M.; Villa Pascual, D.; Dedieu, G. A Multi-Temporal and Multi-Spectral Method to Estimate Aerosol Optical Thickness over Land, for the Atmospheric Correction of FormoSat-2, LandSat, VENμS and Sentinel-2 Images. Remote Sens. 2015,7, 2668-2691.

Figure Example of a snow and cloud mask from a SPOT-4 (Take 5) image taken in the High-Atlas mountains on 27-03-2013. The snow mask polygon is in magenta, the cloud and cloud shadow mask in green.

The key and necessary condition for exploiting Copernicus and other contributing and/or open data missions is to operationally and accurately manage to register all datasets in a common georeference system. Although, most acquired 2D and 3D geospatial data are a priori linked to a certain global, world geodetic system, when one seeks for accurate spatial positioning, local reference ellipsoids must be employed. This registration/ ortho-rectification process is not trivial requiring, usually, detailed Digital Elevation Models (DEMs), labor-intensive and time-consuming procedures, especially, for very and ultra high resolution imaging data or data which are not linked a priori with a reference system. Moreover, despite the numerous research efforts and developed algorithms there are, still, important challenges regarding the automated and accurate registration of very large images, multivariate and multimodal data. In particular, current research efforts on image and data registration are focusing on addressing problems of different sensors, different viewpoints and different time/date acquisitions at large spatial scales (Karantzalos et al., 2014; Seth Price, 2015).

To this end, in this paper we have designed, developed and validated a fully automated registration framework for the registration of Sentinel-2 data to various other imaging and radar datasets. The proposed approach applies recent robust formulations incorporating Markov random fields and powerful discrete optimization algorithms (Karantzalos et al., 2014). In a similar way, we have formulated the deformable registration of Sentinel-2 data as a minimal cost graph problem, where nodes correspond to the deformation grid, a node's connectivity corresponds to regularization constraints, and labels correspond to 2D deformations. We have, also, experimented with a descriptor-based framework (Seth Price, 2015) towards quantitative evaluating the performance of both implemented frameworks. The descriptor-based approach includes several processing steps like feature extraction, feature matching, estimation of the transformation model, geometric transformation and interpolation. The first steps are still challenging and in particular, regarding the feature extraction procedure feature descriptors are employed. Once these descriptor provide several invariant features then enough inliers to recover the registration transformation can be calculated.

For the validation of the developed fully automated registration framework a dataset containing Sentinel-1, Sentinel-2, Landsat-8, MODIS, PROBA and high resolution image mosaics from google earth over the same geographical region was created. Moreover, we have experimented, also, with spectral bands of the same sensor with different spatial resolution e.g., register a Landsat-8 multispectral band (30m) to Sentinel-2 (10m) and register the Landsat-8 panchromatic band (15m) to Sentinel-2 (10m) bands. Experiments were covering cases with data of lower spatial resolution (e.g., MODIS, PROBA) and higher spatial resolution (e.g., google earth mosaics). We have, also, experimented with different similarity metrics like normalized cross correlation or mutual information during the block matching procedure.

The performed comprehensive qualitative and quantitative evaluation indicated that the graph-based deformable registration can more efficiently exploit the spectral variation of multitemporal, multi-sensor satellite data. Moreover, imaging data can be registered more accurately and faster than radar data. This was in accordance with the literature. However, the developed fully automated framework managed, in terms of spatial accuracy, to recover the geometry of optical and radar data with relative low displacement errors. In particular, during all our experimetns the lower the resolution (e.g., MODIS, PROBA) the lower the registration accuracy. The very promising experimental results indicate the high potentials of the developed approach.

Karantzalos K., Sotiras A., Paragios N., 2014, Efficient and Automated Multi-Modal Satellite Data Registration through MRFs and Linear Programming, IEEE Computer Vision and Pattern Recognition Workshops

Seth Price, 2015. Rectifying the Planet, Free and Open Source for Geospatial Conference FOSS4G North America 2015

The recently launched Sentinel 2 satellite with its sensor Multi Spectral Imager (MSI) offers unprecedented possibilities for automated and much more precise glacier mapping at global scales than other freely available sensors (e.g. the commonly used Landsat). With its 10 m spatial resolution in the VNIR bands (20 m in SWIR) and the large swath width of 290 km it will be possible to get an image of all glaciers in a country such as Switzerland or Austria in a single day. The higher repeat frequency also offers better chances for a cloud free acquisition in a comparably short end-of-summer time-window. As the spectral ranges of the VNIR and SWIR bands of MSI and the Operational Land Imager (OLI) on Landsat 8 are very similar, glaciers (clean ice and snow) can be mapped automatically with previously applied methods such as the TM3/5 (OLI4/6, MSI 4/11) band ratio.

In this study a Sentinel 2 precursor dataset (from the commissioning phase) acquired on August 29, 2015 over the Swiss Alps is used to map glaciers with the band ratio method mentioned above and compared to outlines derived from a Landsat 8 OLI scene of the same region acquired only 2 days later (on 31.8. 2015). Additional to the classic red/SWIR band ratio, we have also used the 15 m resolution panchromatic band of OLI instead of the red band to map glaciers. Before the high-resolution band ratios were calculated, the SWIR bands of OLI and MSI were resampled with a bilinear interpolation to a two-times higher spatial resolution. The ratios were applied to the raw digital numbers of all bands without any further correction and threshold values were manually selected.

First results show that the red / SWIR ratio for MSI required an additional threshold in the blue band to improve mapping of snow and ice in shadow while this was not required for OLI. In general, all outlines overlap within the geometric accuracy of the orthorectification, but the 30 m outlines from the OLI red/SWIR ratio were generally found outside of the two others and the 15 m outlines from the OLI panchromatic bands were also outside the 10 m outlines from MSI. The larger extents with the coarser pixels indicate that mixed pixels were generally included, also in shadow regions. This is also obvious from the larger amount of excluded medial moraines with MSI, requiring a higher workload for manual editing. On the other hand, at 10 m spatial resolution the debris-covered parts are much better visible and a higher accuracy and consistency for the manual mapping can be achieved. Despite the foreseen changes in the sensor calibration and geometry, the above results will likely not change much.

Sentinel-2 is a multispectral, high-resolution, optical imaging mission, developed by the European Space Agency (ESA) in the frame of the Copernicus program of the European Commission. It is based on a constellation of 2 satellites: Sentinel-2A launched on 23 June 2015 and Sentinel-2B scheduled for the end of 2016. To ensure the highest quality of service for the Sentinel-2 mission, ESA set up the Sentinel-2 Mission Performance Centre (MPC) supported by several Expert Support Laboratories (ESL).

This presentation reports on the activities performed during MPC commissioning (9 months since S2A launch) by the ESL L2A team (Telespazio France and DLR) in charge of the Calibration of the Level-2A processor, Geophysical validation of all the Level-2A products and maintaining the Level-2A processing baseline and product definition;

Sen2Cor Level 2A processing is applied to Top-Of-Atmosphere (TOA) Level 1C ortho-image reflectance products. Sen2Cor L2A outputs are a scene classification image, aerosol and water vapour maps and the ortho-image Bottom-Of-Atmosphere (BOA) corrected reflectance product. Scene classification is generated together with Quality Indicators for cloud and snow probabilities at 60 m resolution.

This presentation is oriented towards the Sen2Cor L2A processor and L2A products users and will therefore concentrate on:

The description of L2A products contents and format:
(Surface reflectance product, Aerosol Optical Thickness map, Water vapour map at 60m, 20m and 10m resolution
Scene Classification: 12 classes at 60m and 20m resolution
Quality indicators: Cloud probability and Snow probability.)

The description of the Sen2Cor processor operations
(e.g. SNAP, DEM usage, calibration files)

Up-to-date cal/val results for the cloud screening and scene classification as well as for the atmospheric correction.

Note: The details of the methods used for the Validation of Sentinel-2A products are requested to be presented in a parallel poster presentation.

Sentinel-2A is part of the Copernicus program from the European Commission which guarantees the European access to monitoring information for environment and security applications. Its main objectives are: agriculture and forests monitoring, tracking of land-cover change, coastal and inland waters survey, and risk and disaster mapping. For that, Sentinel-2A is equipped with a multispectral imager (MSI) which includes visible, near infrared and shortwave infrared optical sensors. The MSI is a push-broom instrument comprising a three-mirror anastigmatic telescope with a pupil diameter of about 150 mm.

The successful launch in 23 June 2015 was followed by three and a half months of commissioning activities that included the calibration and validation of several geometric and radiometric parameters. The commission activities were performed by teams from the European Space Agency (ESA), at ESOC, ESTEC, and ESRIN, as well as, by the Centre National d’Etudes Spatiales  (CNES) supporting the image quality activities. DLR and Tesat were involved in the commissioning of the optical communication system.

In order to maximise the energy absorbed by the solar array, a drive mechanism ensure direct pointing of the solar cells towards the Sun during along the Sun illuminated part of orbit. It is known that any on-board moving parts can potentially induce movement of the whole spacecraft due the conservation of angular momentum and, consequently, perturb its nominal pointing. Due to the very large swath as well as the high resolution of the images (several bands with 10 meter pixel size) acquired by the MSI instrument, even small movements  of the spacecraft can be detected. On the other hand, these movements can be compensated when processing the image data using a geometric model and very accurate attitude information. The sensor suite of Sentinel-2A includes a set of three star-trackers and fiber-optic gyros that are key to compute its attitude.

Due to a timing difference between the MSI and the attitude provided by the on-board Gyro-Stellar Estimator (GSE), Sentinel-2A angular oscillations were initially not fully compensated resulting in a residual inter-band misalignment. This paper describes the prototyping of an on-ground attitude estimator at ESTEC to enhance the inter-bands and multi-temporal image co-registration. This estimation algorithm is based on the on-board GSE and resorts to star-trackers and gyro measurements to compute high accuracy attitude information that is used to compensate the platform oscillatory movement. The verification and implementation activities are also addressed. The very accurate inter-band co-registration results obtained from images compensated using the new on-ground restituted attitude validate its good accuracy as well as its integration in the processing chain.