Automated production of LMS Water and Wetness layers
The Copernicus Land Monitoring Service (LMS) HR Water and Wetness layers production work-flow was designed by GEOMATRIX UAB to ensure smooth transition from the former GMES Initial Operations (GIO-Land) production (carried out by GEOMATRIX UAB as well) and introduce new innovative solutions for the future LMS HR layers production based on the Copernicus Sentinel 1/2 products. Copernicus LMS HR Water and Wetness layers are produced following the proposed high level work-flow description and requirements of the EEA Technical specification.
A simplified schema of integrated Water & Wetness production work-flow is shown on the top of the left-side block. This work-flow is used for processing all MS and SAR image stacks separately. Colorus indicate consecutive processing stages: black – pre-processing of satellite imagery and ancillary data; yellow – classification of separate images; green – accumulation of classified masks from a production stack; blue – computing of probability indexes, brown – statistical filtering of the results; magenta – final deliverables.
Time series of cloud-free image mosaics are processed into standard biophysical variables, then masked, filtered and classified into separate binary Water and Wetness masks with Water pixels included in Wetness masks which are summed together into combined (integer) Sum of Water Masks (SWAM), Sum of Wetness Masks (SWEM), Sum of Dry Land Masks (SDLM) and Sum of Image Masks (SIM) layers. Water & Wetness Probability Index (WWPI) is computed for each image stack separately with low-frequency classification "noise" filtered out and aggregated into a specified resolution Water and Wetness Presence Index (WWPI) for final delivery (see the 4th line image on the left-side block of examples). Filtering (SWAM / SIM) and (SWEM / SIM) ratio values into high (0.75-1.0) and low (0.25-0.75) frequency ranges produces Permanent Water Layer (PWAL), Temporary Water Layer (TWAL), Permanent Wetness Layer (PWEL) and Temporary Wetness Layer (TWEL) and eliminates Water and Wetness classification errors (ratio values below 0.1 – or 10% of the images). Image on the bottom line of the left-side block of examples shows an extract of the pilot testing area covered with combined PWAL, PWEL, TWAL and TWEL layers.
Pixel resolution and alignment of the final products – no matter what were the original resolutions of WWPI and WAWEL products from different image stacks – are defined automatically by the settings of data processing framework used for the aggregation process, so there there is no need for a separate work-flow step of re-sampling into a given resolution resolution. This approach allows easy aggregation of different types of imagery – from RapidEye (5 m resolution) to Landsat (30 m) or medium resolution Sentinel-1 GRD EW – into a single product. To present capabilities of different Water and Wetness extraction methods, on the left side of 2 and 3 rows we loaded Sentinel-1 GRD IW image (10 m resolution) taken on 2016/03/19 over the pilot area in southern France, preprocessed into 8-bit VV and VH SAR polarisation bands, while on the right side - Sentinel-2 MSI image (10 m resolution) taken on 2016/03/19 - both masked with EU-Hydro coast-line and classified into Water (blue) and Wetness (magenta) layers.
Water and Wetland masks are extracted following slightly different algorithms developed for different types of satellite data. Currently we have developed algorithms for extraction Water and Wetness masks from IRS-6, RapidEye, SPOT 5/6/7, Sentinel-1 and Sentinel-2 images. Algorithms of process automation allow flexible manipulation of processing steps or algorithms to be used for a particular task. Examples on the 5th line of the left-side block present combined PWAL, PWEL, TWAL and TWEL layers extracted from IRS-6 MSI (left picture) and Sentinel-1 SAR (right picture) satellite images.
Classified Water & Wetlands Layer produced from Sentinel-1 SAR images shows excellent capabilities of the radar sensor to find open water surfaces (no matter how shallow or turbid/contaminated that water would be), but it is not able to distinguish permanent wetlands with dense and tall macrophyte vegetation because of a strong back-scatter signal mixing with many other vegetation types within oermanent grassland areas. So the most practical solution would be combining classified layers produced from optical and radar imagery and complement their thematic coverage - just as shown in the image on the bottom row of the examples block.
The final steps of the process include automated semantic analysis and correction phase is based on routine filtering of WWPI, PWAL, PWEL, TWAL and TWEL layers with various ancillary datasets (EU-DEM, EU-Hydro, GIO-Land products, etc.). The final quality control is done manually by revising 100% of the products coverage by qualified photo-interpretors holding the experience from GIO-Land Lot 6 production. Systematic errors detected by manual revision are corrected by updating the automated processing software and re-processing the datasets.
The most important extension to the former GIO-Land production work-flow is a concept of running parallel production work-flows on stacks of different satellite images. This enables optimised classification algorithms for different image types and balance impacts of those different time series within the final products – no matter how different were numbers of images (SIM values) in those stacks. Also some work-flow modifications are introduced in the initial stage to enable automated discovery, retrieval and package integrity checks.
Classification of of Wetness and Water layer from SAR images is based on ability of smooth surfaces to bounce (Water) or disperse (Wetness) the incoming SAR signal, causing low back-scatter registered by the receiver. Low values in both VV and VH polarisation bands give a strong indication of high surface wetness in areas without grass vegetation, while low VH signal along with high VV signal indicates water or high wetness surfaces under the grass vegetation cover.
To achieve equal distribution of SAR and MSI products “weights” in the final statistical analysis, we compute WWPI and WAWEL layers separately for SAR and MSI products and then calculate average values of non-0 values while producing the final integrated products. This solution will provide a quick work-around to balance the inputs from SAR and MSI sensors in the final products, however it causes a serious risk of propagating MSI products classification errors into the final product in case if MSI stack is small and classification errors have systematic patterns.
All the processing steps, re-processing requests and parameters used for the processing will be recorded in the on-line Log-book, accessed only by the production team. Records of the production log-book may contain any kind of textual information and any digital attachments (pictures, documents, etc.), the Log-book database is searchable by any combination of custom key-words to make the information easily to aggregate and access. Production log-book will be used not only for recording of production and QC steps, but also as a knowledge-base archives for the future projects.