The normalised Sentinel-1 Global Backscatter Model, mapping Earth’s land surface with C-band microwaves

With S1GBM’s characteristics as a global, PLIA-normalised, high-resolution C-band backscatter dataset, a direct validation experiment is not feasible since we lack matching reference backscatter data collected during airborne or ground based radar campaigns. Other existing global mosaics were generated based on different time-spans, polarisations¹⁷, frequencies¹⁸, or do not share the novel feature of the PLIA-normalisation²⁰.

On these grounds, we prefer to assess the characteristics of the S1GBM layers with respect to different land cover types on a global scale, and to incorporate the gained knowledge into an easy-to-use classification algorithm for permanent water bodies (PWB). This simple mapping experiment acts as an example and should on the one hand demonstrate the integrity and quality of the S1GBM mosaics (and document its limitations), and on the other hand, stimulate more advanced applications and ingestion-models by the remote sensing- and the wider user -communities. Our validation of the obtained PWB-map compares—over a representative and diverse set of eight world regions (see Fig. 1b)—the S1GBM mosaic with a reference water body map, as well as with true-colour imagery from the Sentinel-2 optical sensor. This arrangement should also portray the shape and texture of the S1GBM mosaic and help the audience with the interpretation of the SAR imagery, which as stated at the outset, allows a unique view on the Earth’s surface.

In the following, 1) we examine in detail the appearance and spatial features of the S1GBM VV- and VH-mosaics over the region of Bordeaux, also investigating the effect of the PLIA-normalisation. Then, 2) we derive the characteristic C-band backscatter signature for global land classes. Finally, 3) we perform the PWB-experiment in eight world regions a) to evaluate the dataset’s integrity, b) to demonstrate its spatial information and exemplify its use, and c) to comment on the S1GBM’s assets and caveats.

Detail example Bordeaux

Figure 2 gives an example of the land cover signal in the S1GBM VH and VV mosaics over Bordeaux, France. Comparing it with the recent PROBA-V-based Land Cover dataset of the Copernicus Global Land Service (CGLS LC100⁵²), several surface features are apparent in the mosaics, including urban areas with varying density in both VV- and VH-channels. In the VH mosaic, a clear discrimination of forest areas (cf. with LC100’s broadleaf in brighter green, needle leaf in darker green) against crops (brighter yellow) and vineyards (darker yellow) is apparent. The cross-polarised VH-backscatter is more sensitive to vegetation-density, -structure, and -status, as multiple scattering between branches and volume scattering increases the share of backscattered microwaves with changed polarisation. Most prominent, in both VH and VV, is the very large contrast between land surfaces and open waters with significant lower backscatter signatures. This is the basis for our PWB-mapping experiment discussed in detail in the subsequent section.

We would also like to draw the attention to the spatial detail carried by the S1GBM mosaics, with various features at deca- and hectometric scale shown for example in Fig. 2. For instance, one can see bridges, highways, railways, and airports in the Bordeaux metropolitan area in the south-west corner of the here displayed T1-tile (100 km extent). Also, in the west, from north to south, the shorelines of the Gironde estuary and its downstream rivers are clearly mapped, resolving small islands and narrow straits. Agricultural plots and forest sections may be differentiated especially in the VH mosaic, e.g. with particular structures in the north-west corner. For further exploration, users may visit the open web-based S1GBM viewer⁵¹ offering a pan-and-zoom exploration of the full S1GBM VV- and VH-mosaics.

Figure 2b,d allows the comparison of the S1GBM VV backscatter mosaic (which underwent the PLIA-normalisation) against the mean of non-normalised Sentinel-1 VV backscatter from the same observation period (not part of the dataset publication; just for comparison). As discussed above, radar backscatter is strongly dependent to PLIA, and hence Sentinel-1 SAR images are subject to the observation geometry defined by the mission’s relative orbit configuration and the overlapping pattern (cf. global map in Fig. 1b). One can clearly see this impact in Fig. 2d, where data from all local orbits are averaged in their native orbit geometry (i.e. mean of σ⁰ (θ_ro, t), resulting to characteristic linear artefacts of backscatter discontinuities along the limits of the (repeating) orbit footprints. The mini-map of the Bordeaux-T1-tile in Fig. 2d plots the number of input Sentinel-1 scenes, also reflecting the heterogeneous coverage pattern induced by the different number of overlapping relative orbits (from 2 to 4 in this area), each with a different local PLIA-range, generally. Notably, the triangular zone covered by only 2 orbits (yellow, 194 scenes) is a zone that features a PLIA-spread that is not large enough to reliably estimate the local PLIA-slope β. This zone is part of the pixel domain where we applied the static slope value of −0.13 dB/° to the S1GBM mosaic, with a resulting backscatter image that is free from orbit-related artefacts (Fig. 2b). We note that the sections covered by 3 or 4 orbits in this example are normalised with the regular regression slope, letting us conclude that our approach yields a smooth mosaicking impression in areas of mixed coverage density.

Backscatter signature analysis

Delving into above concept that SAR backscatter characteristics in the S1GBM are determined by land cover, we analysed the backscatter signature for the global land surface for each major land cover class (LCC). We globally aggregated data from the normalised S1GBM VV and VH mosaics per LCC and formed the backscatter distribution within each LCC, allowing the discrimination of typical SAR backscatter signatures for a specific land cover class.

Land cover definitions

As land cover dataset, we selected the above-mentioned PROBA-V-based CGLS LC100 for its full global coverage and the (for global datasets) relatively high spatial resolution with a pixel spacing of 100 m. To allow a fast pixel-by-pixel comparison, we resampled the CGLS LC100 to the Equi7Grid at 10 m using nearest-neighbour-downsampling. After a first inspection of backscatter signatures, we grouped the 23 LCC of the LC100 to 13 major LCC, accounting for the similarity between certain classes: Respective open and closed forest classes were aggregated to evergreen needle leaf forest, evergreen broad leaf forest, deciduous needle leaf forest, and deciduous broad leaf forest, and herbaceous wetland was grouped with herbaceous vegetation. Table 2 lists the main statistics per land cover and the group aggregations.

C-band backscatter signatures

The C-band backscatter signatures of our major 13 LCC are plotted for VV- and VH-polarisation as distribution-density-“heatlines” in the upper part of Fig. 3, illustrating the global average backscatter levels of each surface class, and the variance within. Forest and water-body classes have a very narrow distribution, whereas snow and ice and bare vegetation have a greater spatial backscatter variability. Snow and ice packs often have a heterogeneous structure from its complex genesis involving melting and freezing phases, leading to a mixture of surface- and volume-scattering when observed by radar. Likewise, the LCC bare vegetation comprises very different surfaces dominated by rocky, sandy, or mountainous surfaces, each governed by a distinct backscatter behaviour and hence create the wide spread within this LCC.

The LCC-heatlines in Fig. 3a,b are approximately ordered by the mean backscatter value. On top, one can find the two water LCCs with a very low backscatter level that is caused by mirror-like-reflection away from the sensor, followed by bare and herbaceous vegetation LCCs that are dominated by dry conditions and hence are generally weak scatterer. The LCCs moss & lichen, shrubs, and agriculture feature medium backscatter and variation thereof. Higher backscatter levels are observed over the forest LCCs, where volume and multiple scattering become more dominant, as well as over the LCC urban & built up, where corner reflections acting as echo cause the strongest radar backscatter.

When comparing VV and VH polarisation, the biggest difference is in the overall level of backscatter, with about 7 dB between both polarisations across all LCCs. The order of LCCs as a function of mean backscatter is mostly the same for VV and VH, except for the water and ice classes. Interestingly, the open sea class shows a steeper drop from VV to VH, whereas shrubs show a comparatively small drop. We found that the strongest changes in the backscatter distributions are apparent in the non-forest vegetation classes, e.g. for bare vegetation and agriculture, supporting our initial assumptions on the sensitivity of Sentinel-1 VH backscatter to complex vegetation dynamics and crop varieties.

Permanent water body mapping

Following up to what we have already seen along the rivers in Fig. 2, water bodies (represented by the LCCs open sea and permanent water bodies) show a most distinctive backscatter signature in relation to other land cover classes (cf. 3a-b). Effectively, water surfaces show in radar images a strong contrast with land surfaces. The reason for this are the different microwave scattering mechanism over water- and land-surfaces and the side-looking geometry of SAR systems. A specular reflection of the radar pulses by the water surfaces leads to backscatter intensities received by the sensor that are much lower than for most other land cover types. With the S1GBM VV- and VH-mosaics at hand, we exploited this discriminative feature of water bodies and employed a simple permanent water body mapping method. Unlike the backscatter mosaics of the S1GBM, the obtained PWB map can be validated directly, as we have available matching global water body maps as a reference. Moreover, the experiment should demonstrate the ease of realising a land cover mapping application in short time, exploiting the novel S1GBM data and its high-resolution radar imagery of the Earth’s land surfaces.

Based on above insights from the Sentinel-1 backscatter signature analysis, our first step was to spatially merge all water- and all land-LCCs and build the combined backscatter signatures for VV and VH (Fig. 3c,d). The water distribution (all water classes; bright blue) is plotted for both polarisations next to the non-water distribution (all land classes, bright brown), already demonstrating an acceptable feature separation. However, as one can see in the heatlines above, water has still some significant overlap with some land LCCs, e.g. with bare vegetation, herbaceous vegetation, and moss & lichen. Naturally, this translates to a considerable overlap in the merged distributions below, especially in the VH case and for moss & lichen. We concluded that for these LCCs no robust separability against water bodies is given in the S1GBM data and excluded the three classes from further PWB-mapping. Also, we dropped the LCC open sea in further processing as we limit the PWB experiment to inland surfaces (that are also covered by the reference dataset). The backscatter distributions of the PWB LCC and the selected land LCCs are shown in dark blue and brown (permanent water bodies and selected land classes in Fig. 3c,d), with a noticeably improved separability, especially in VH polarisation.

As a next step, evoking the theory of Bayesian inference with equal priors for binary classification, we obtained a statistically optimal global threshold for VV and VH, each. In this respect, we identified two thresholds, −15.0 dB for VV and −22.9 dB for VH polarisation, which we applied in a third step as an upper-bound backscatter-value on the complete S1GBM mosaics to map the global PWBs. Note again that the LCCs bare vegetation, herbaceous vegetation, moss & lichen, and open sea are not included in the PWB-mapping and are masked in all later results.

Although the VV and VH mosaics are redundant to some degree, the consideration of both channels is most advantageous for the PWB-mapping. First, the classification based on Bayesian inference is more robust when resulting from two discriminations. Second, while the VH mosaic offers a better separability between water and non-water (having less overlap in the distributions and hence less false positive and negative classifications), and the heatline of the PWB-LCC is better defined in VH, the VV mosaic offers in general a higher spatial detail due to its stronger backscatter signal and hence more favourable signal-to-noise ratio.

By applying the obtained thresholds to the normalised S1GBM mosaics as simple classification rules

and through joining them with logical “AND”, we were able to produce a global PWB map in less than two hours, using 70 parallel cores on the VSC-3 supercomputer.

Evaluation of S1GBM mosaics and PWB map

To evaluate our S1GBM permanent water body (PWB) map, we chose as a reference dataset the Global Surface Water (GWS²³) from the European Commission’s Joint Research Centre (JRC-EC). The GSW offers globally at a 30 m native sampling different variables on water bodies, e.g. annual seasonality, occurrence, recurrence, or maximum extent, and is based on 36 years of Landsat data in its newest version (GSW1_2). Although the annual seasonality for 2015 or 2016 was not accessible from version GSW1_2 at the time of writing this manuscript, we found the Seasonality 2015 dataset of the GSW1_0 version suitable as a reference. Pixels valued with seasonality “12” (i.e. all months) are labelled permanent water and constitute our reference PWB map, which we warped by means of bilinear resampling to the Equi7Grid at a 10 m pixel spacing.

The evaluation presented in this paper was carried out on a representative and diverse set of eight world regions (see locations in Fig. 1b). For each region, classification results were assessed by a pixel-by-pixel comparison between the PWB map from S1GBM and from the GSW reference. Having such binary maps (water vs. non-water) it was straightforward to generate an “accuracy layer” representing the four elements of the commonly used confusion matrix, i.e. true positives, false positives, false negatives, and true negatives, to discuss the skill of the S1GBM to map PWBs. Areas belonging to the four excluded LCCs were masked in the result plots. Furthermore, to give some visual guidance in the evaluation regions, we acquired from the Copernicus Sentinel-2 Global Mosaic (S2GM) service the RGB-composite for the year 2019⁵³ (the mosaic for 2015 was available only over Europe).

In the following, we present results for four large-scale regions (500 km × 500 km) in Fig. 4, and for four small-scale regions (120 km × 120 km) in Fig. 5. For each region, the S1GBM VV mosaic is displayed on the left panel (space-saving/omitting the VH mosaic, which contributes likewise to the PWB mapping), the accuracy maps showing the performance against the GSW reference in the centre panel, and the Sentinel-2 RGB-composite to aid visual interpretation on the right panel. The accuracy maps are annotated with the respective User’s Accuracy (UA) and Producer’s Accuracy (PA), as the percentage of the agreement between the two PWB-maps.

Large-scale examinations

Figure 4a–c shows the southern part of Finland, an area accommodating a multitude of small and large post-glacial lakes. Those are clearly visible in dark colours representing low backscatter values in the S1GBM mosaic, while the other parts of the country (which is dominated by vast forests) shows rather uniform medium backscatter. The optical RGB-composite from Sentinel-2 does not feature the same accentuation of the lakes, troubled by remainders of cloud coverage in the yearly mosaic. The PWB accuracy map shows perfect agreement between S1GBM and GSW, with an UA and PA of 100% each. We identified two reasons for the excellent performance: First, the C-band backscatter signatures of the predominant land covers in Finland, such as forests, cities, agriculture, are well distinguishable against water bodies and hence allow an almost sterile PWB-mapping. Second, northern Europe is well covered by the Sentinel-1 mission and the S1GBM has been built with a high data density, letting us expect the best mosaic quality.

Moving to the region of the Lake Superior Basin in Canada and USA presented in Fig. 4d–f, we encounter a very similar, cold-temperate environment, but with a substantial higher share in spacious inland water bodies. Also here, the accuracy map shows a perfect agreement between S1GBM and GSW, which, in our interpretation, is clearly because of good feature separability in the SAR image. Particularly remarkable is that North America is much less covered by Sentinel-1 than Europe and that the imperfect modelling of the PLIA-dependency over water surfaces (as apparent e.g. in the east section of Lake Superior) does not impair the S1GBM PWB-mapping. Generally, imperfect PLIA-normalisation of SAR images is prominent over water bodies, whose specular reflection regime is characterised by a very strong PLIA-gradient (i.e. the slope β). However, we note that also the Sentinel-2 mosaic has striping artefacts bound to orbit footprints, and additionally suffers from cloud cover. The latter is a common problem in optical observation of higher latitudes, but is without effect in SAR imagery.

Figure 4g–i depicts the situation for a section of the Albertine Rift Valley in eastern Africa with its lake system. Reflecting to a great deal the region’s diverse flora, which is displayed in many green and brown tones in the RGB-composite, the S1GBM VV mosaic shows a much more heterogeneous pattern than in the above examples. The forested sections in the west show distinct higher backscatter values than the savanna sections in the east, and also other geomorphological features correspond well with the radar and optical mosaic. Concerning the PWB-mapping, we see again perfect agreement, but with one large exception: the eastern end of Lake Albert is entirely labelled in red as false land, suggesting that these water areas are missed in the S1GBM PWB map (what can be confirmed after a quick check with common thematic maps). In this area we see the impact of the relatively poor input data density of about only 50 Sentinel-1 scenes (cf. Figure 1b), and apparently, we overlooked the impact of a few images with outlying backscatter levels during the manual quality curation. Moreover, the three Sentinel-1 relative orbits covering this area create almost identical viewing angles and yield a very small PLIA-range, troubling our backscatter normalisation. As a result, striping artefacts appear not only over water bodies (cf. Canada example) but also over land (in north-west part Fig. 4g), while, however, the Sentinel-2 mosaic is likewise affected by striping issues (cf. Figure 4i), for other reasons, though.

The last row in Fig. 4(j–m) is centred at Bangladesh and displays the confluence of the Ganges and Brahmaputra streams, which are joined downstream by the Meghna river and ultimately discharge into the Bay of Bengal. Also in this region, the geomorphological features perceivable in the RGB-composite are reflected well by strong textural patterns in the S1GBM mosaic, promoting its broader use in land cover applications (note also the zoom-in plotted in Fig. 5j–m). The PWB-mapping results are inconclusive, as rivers of all sizes are correctly mapped, but many pixels are labelled in yellow as false waters. We consider this disagreement between S1GBM and GSW to be most likely a result of the different temporal resolutions of the two datasets, as the S1GBM is a two-year data aggregation reduced to single layers, whereas the GSW allows monthly snapshots of water bodies. For example, the Hoar ecosystem—which appears as yellow bulb in the north-east of Fig. 4k—is a large monsoon-fed lagoon system that is labelled by the GSW with seasonality-values ranging from 9 to 12 months. In the S1GBM mosaics, which are built using temporally averaged backscatter, these areas are obviously dominated by the high occurrence of water surfaces and act therefore as “most-of-the-time water bodies”. Some more vindication comes from the Sentinel-2 yearly mosaic, which also draws the Hoar area with a water texture. We conclude on this matter that seasonal water bodies are not properly modelled by our simple approach with Eq. 3, and it would need additional inputs from variance measures like the backscatter standard deviation.

Small-scale examinations

Figure 5 depicts the small-scale example regions with respect to the PWB-mapping experiment. The first row in a-c) zooms to the Swiss lakes in central Europe and both, the radar and the optical mosaic, feature a high level of heterogeneity and detail, with many individual forests, cities, valleys, rivers, alpine lakes, and with the airport north of Zurich resolvable (in the centre-left of the box). The results from the PWB-mapping are very good with high UA- and PA-values, but with two anomalies: First, the southern arm of Lake Lucerne (in the south-west) shows some red segments of false land along the mountain flanks reaching into the lake. After inspection of the S1GBM mosaics we can state that this is clearly an artefact from the terrain modelling with the rather coarse, 90 m-sampled VFP SRTM Digital Elevation Model (DEM) during the Sentinel-1 preprocessing. At the time of the project, we selected the VFP DEM³⁵ for its complete global coverage and its manually-checked quality, and accepted the coarse resolution (with respect to the 10 m-sampled Sentinel-1 SAR data). The second small anomaly can be found in the Alps in the south of the image, with the west-end of the Klöntalersee labelled in yellow as false water. The S1GBM is artefact-free at this location, and after checking the GSW’s seasonality, we hypothesise that ice covers this mountain lake during winters and leads to the different interpretation.

Figure 5d–f presents the area around the confluence of the Amazon and Tapajós streams in central Pará in Brazil. Here, the rivers ramify into a multitude of lagoons and channels at various sizes, forming a complex system of water bodies. Fortunately, while the Sentinel-2’s RGB mosaic appears impure and rugged from contamination with the frequent cloud coverage in the central tropics, the Sentinel-1 mosaic offers a clear image that fully resolves the capillary structure of the water bodies and its shorelines. We consider this a remarkable feature, also recognising the very low input data density of the S1GBM mosaics in this area (cf. Figure 1b). Concerning the PWB-mapping, we obtained a good agreement with the GSW’s reference, labelling most PWBs correctly and misclassifying only small sections of the lagoons and river-arms. The false-water deviations are bound again to the seasonality of those segments that are most of the time under water, much alike to the situation in Bangladesh discussed around Fig. 4j–m. The red-labelled areas highlight water bodies which are mapped by the GSW but not by the S1GBM, and are of particular interest, as they exemplify that water surfaces seen by optical sensors are not necessarily identical to those seen by radars⁵⁴. Swamp-like structures and waters with out-growing vegetation show a completely different SAR signature and hence might be distinguishable from open waters within a SAR image.

The third small-scale example is the Great Salt Lake in Utah, USA, as displayed in Fig. 5g–i. The S1GBM offers many details of Salt Lake City’s structures in the south-east, and of the mining facilities at the eastern shorelines of the lake, as also visible in the RGB-composite. Obviously, the radar image does not account for the difference in salinity between the north- and south-section of the Great Salt Lake that is visible in the optical image. However, our S1GBM PWB method maps correctly—contrary to the GSW reference—the east-west rail causeway splitting the lake, which one can see as a red line in the accuracy map in Fig. 5h. With its pronounced semi-arid climate, this region shows a different behaviour than above examples. The dry conditions and the sparse vegetation with its weak scattering trouble seriously the S1GBM PWB-mapping, with many false water pixel all around the area. Here, we see the weak performance of the simple threshold approach with Eq. 3 in regions with a general low backscatter from land, and hence small contrast to water bodies.

Figure 5j–m zooms into the Sundarbans at the southern shorelines of Bangladesh, with its multifaceted surface and its complex river-deltas. Both, the true-colour image from Sentinel-2 and the VV-mosaic from Sentinel-1 produce a feature-rich image and highlight the mangrove forest in the southern section with strong green colour or high backscatter, respectively. Adjacent to the north, the rice and bean agriculture draws large contrast patterns in the satellite images. For the PWB-mapping, a similar result as from the larger view on this region (cf. Figure 4k) is obtained, with all rivers and channels correctly classified, but with a substantial overestimation of permanent water bodies in areas of high water seasonality. To what extent rice fields and its managed inundations play a role here is left unanswered by the data, though, as managed rice fields typically show significant jumps in seasonal backscatter time series.

Source: Ecology - nature.com