OR/15/001 Products description

From Earthwise
Revision as of 10:38, 29 September 2015 by Dbk (Talk | contribs) (National Digital Elevation Model)

(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Jump to: navigation, search
Jordan, J J, Grebby, S, Dijkstra, T, Dashwood, C and Cigna, F. 2015. Risk information services for Disaster Risk Management (DRM) in the Caribbean. British Geological Survey Internal Report, OR/15/001.

Products

Service 1: Land use/land cover mapping

The objective of Service 1 is to generate land use/land cover maps and produce a vector layer of water features (e.g. lakes, ponds, rivers) for St. Lucia, Grenada, and St. Vincent & the Grenadines from high-resolution or very high-resolution optical satellite imagery. In keeping with the 'eoworld' framework, suitable imagery for this service was preferentially sought from European and Canadian sensors. Consequently, archived Pleiades imagery (acquired 2013–2014) and RapidEye imagery (acquired 2010–2014) with reasonable levels of cloud (and associated shadow) were obtained through the ESA Third Party Mission scheme. The basic characteristics of the obtained datasets are shown in Table 4.

Table 4 Basic characteristics of the utilised satellite data.
Dataset Wavebands Spatial resolution (m)
Pleiades Panchromatic (470–830nm)
Blue (430–550nm)
Green (500–620nm)
Red (590–710nm)
Near‐infrared (740–940nm

0.5

2

2

2

2

RapidEye Blue (440–510nm)
Green (520–590nm)
Red (630–685nm)
Red Edge (690–730nm)
Near‐infrared (760–850nm)

5

5

5

5

5

Existing land use/land cover maps from 2001 were available for the three AOIs following their production by the various partners of The Nature Conservancy’s Mesoamerica and Caribbean Region project (Helmer et al., 2007[1]; 2008[2]). These maps were produced at a spatial resolution of 30m through a combination of automated imagery classification of Landsat Enhanced Thematic Mapper Plus (ETM+) satellite imagery and manual delineation of IKONOS satellite imagery. The mapping was augmented with extensive field knowledge and observation. Accordingly, these maps provide excellent baseline data to build upon for the land use/land cover mapping under this service.

The Pleiades multispectral imagery was selected as the basis for the three new land use/land cover maps because its higher spatial resolution enables the most detail to be resolved. With a spatial resolution (pixel size) of 2m for the multispectral bands, mapping can theoretically be undertaken at a scale equivalent to 1:10 000. Given the areal extent of the AOIs, the spatial resolution and the resources available, it was decided that the most efficient mapping strategy was to utilise a largely automated approach with subsequent refinement and manual digitisation.

Before mapping, the land use/land cover classes for the three AOIs were first defined by considering those included on the existing 30m maps, which are based on the International Institute of Tropical Forestry classification scheme. Minor modifications to these classes were made where appropriate to reflect the anticipated discrimination capability of the satellite imagery (Table 5). For example, the 'low density urban' and 'Medium-high density urban' classes in the existing maps were modified to 'Roads and other built-up surfaces' and 'Buildings' given the ability to resolve these features in the high-resolution Pleiades imagery. Moreover, given their inherent similarities and the anticipated difficulty in discriminating between them, 'Pastures', 'Cultivated land' and 'Herbaceous agriculture' were merged to form a single class.

Table 5 Land use/land cover classes for the three AOIs.

St. Lucia

Grenada

St. Vincent & the Grenadines

Land use/land cover classes

Water

Water

Water

Wetland

Wetland

Mangrove

Mangrove

Mangrove

Buildings

Buildings

Buildings

Roads and other built‐up surfaces (e.g. concrete, asphalt)

Roads and other built‐up surfaces (e.g. concrete, asphalt)

Roads and other built‐up surfaces (e.g. concrete, asphalt)

Bare ground (e.g. sand, rock)

Bare ground (e.g. sand, rock)

Bare ground (e.g. sand, rock)

Semi‐deciduous forest

Quarry

Quarry

Seasonal Evergreen forest

Semi‐ or Drought Deciduous, coastal Evergreen and mixed forest or shrubland

Semi‐deciduous forest

Evergreen forest

Lowland forest (e.g. Evergreen and seasonal Evergreen)

Drought Deciduous open woodland

Drought Deciduous, coastal Evergreen and mixed forest or shrubland

Evergreen forest

Evergreen and seasonal Evergreen forest

Elfin and Sierra Palm tall cloud forest

Elfin and Sierra Palm tall cloud forest

Deciduous, coastal Evergreen and mixed forest or shrubland

Montane non‐forested vegetation (e.g. high‐ altitude pastures)

Woody agriculture (e.g. cacao, coconut, banana)

Elfin and Sierra Palm tall cloud forest

Blue Mahoe plantation

Pastures, cultivated land and herbaceous agriculture

Nutmeg and mixed woody agriculture (e.g. cacao, coconut, banana)

Woody agriculture (e.g. cacao, coconut, banana)

Golf course

Pastures, cultivated land and herbaceous agriculture

Pasture, cultivated land and herbaceous agriculture

Golf course

Golf course

The basis for each land use/land cover map was the result of supervised per-pixel classification of the imagery according to the land use/land cover classes outlined in Table 5. Due to their inherent similarities, it was anticipated that some of the classes — in particular some forest types, and 'Bare ground', 'Roads and other built-up surfaces', 'Pastures, cultivated land and herbaceous agriculture' — could be particularly difficult to discriminate using the limited spectral information contained in only the blue, green, red and near-infrared bands of the Pleiades imagery. Therefore, textural information was also incorporated in the form of the Grey-Level Co-occurrence Matrix (GLCM) parameters of entropy, dissimilarity, second moment and homogeneity (Haralick et al., 1973[3] and Herold et al., 2003[4]). These parameters were derived from the Pleiades green band in the ENVI 4.8 software package (Research Systems, Inc.) for a 3 × 3 pixel (i.e. 6m × 6m) window and a co-occurrence window shift of 4 pixels (i.e. 8m) in both the x- and y-direction. These 4 textural bands were merged with the 4 Pleiades multispectral bands to create an 8-band spectral-textural dataset for each AOI for input to the classification. Classification of the datasets was performed using a supervised neural network (NN) classification algorithm. The NN used in this case was a Multi-Layered Perceptron NN with a back-propagation learning algorithm for supervised learning (Richards and Jia, 2006[5]). Using a three-layered NN (i.e., input, output and one hidden layer), land use/land cover classifications were performed in ENVI 4.8 with the default training parameters. Each classification was supervised with the aid of a set of training pixels that were carefully selected to represent the spectral and textural characteristics of each of the land use/land cover classes. These training pixels were identified in Pleiades imagery by using the existing maps as a guide.

The classified images did not initially provide full areal coverage of the AOIs owing to the obscuring effects of cloud cover and associated shadowing in the imagery. Accordingly, areas with missing land use/land cover information in the classified image were first patched using the results of RapidEye image classifications achieved using the same approach as outline above. Any remaining unclassified areas were then patched using the existing land use/land cover information (Figure 4). The St. Vincent & the Grenadines classification was patched entirely using the existing land use/land cover maps due to the poor quality of the RapidEye imagery.

Next, the preliminary land use/land cover maps were augmented using a combination of rule- based refinement and manual delineation. Rule-based refinement comprised reclassifying the classes of some pixels according to a set of rules or criteria. For instance, in all 3 cases rule-based refinement was used to reclassify forested pixels as 'Elfin and Sierra Palm tall cloud forest' if they occurred above a specific elevation. Similarly, for St. Lucia, rule-based refinement was used to distinguish between 'Lowland forest' and 'Evergreen forest' according to the elevation at which the transition occurs. Land use/land cover classes that were difficult to discriminate using a combination image classification and rule-based refinement were manually delineated. This involved visually identifying and mapping (at 1:10 000-scale) classes such as 'Mangrove' and 'golf course' from the Pleiades imagery, which was pan-sharpened to a spatial resolution of 0.5m using the panchromatic band.

Figure 4 Data used to patch the land use/land cover maps for (A) St. Lucia, (B) Grenada and (C) St. Vincent and the Grenadines.

The final land use/land cover maps are raster images with a spatial resolution of 2m, and thus provide an order of magnitude increase in the amount of detail they contain in comparison to the existing maps (Figure 5). Each 2m pixel in the raster is attributed with its associated land use/land cover class.

Figure 5 Comparison of the level of detail provided in the new 2m land use/land cover maps (middle) and the existing 30m maps (right, © Nature Conservancy Mesomerica and Caribbean Region). Pleiades imagery is shown on the left (includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A., France, all rights reserved).

In addition to land use/land cover, vector layers (shapefiles) documenting water features (e.g. lakes, ponds, rivers, streams) and the basic road network were also created for all 3 AOIs. These features were manually digitised at 1:10 000-scale from the pan-sharpened Pleiades imagery. Where available, existing vectors layer provided by various sources (e.g. local Physical Planning Offices, OpenStreetMap) were utilised as baseline information during mapping. These layers were edited to add, remove and shift features, or re-digitised at a finer scale, as necessary. Cloud cover, associated shadowing and dense vegetation made it difficult to delineate linear features such as rivers, streams and roads in the imagery (Figure 6). With such features generally only resolvable at the resolution of the Pleiades imagery, complete coverage of the AOIs was not possible.

Figure 6 Delineation of rivers, streams and roads from pan-sharpened Pleiades imagery.

The land use/land cover raster and vector datasets for each country were also compiled into map format. An example (for St Lucia) is illustrated in Figure 7. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.

Figure 7 Land use/land cover map of St Lucia.

Service 2: Hazard mapping to support landslide risk assessment

Landslide inventory mapping

RapidEye and Pleiades images were available for interpretation and landslide inventory establishment. The RapidEye images are available at resolution of 5m while the Pleiades imagery was pan-sharpened to a resolution of just 0.5m. The project imposed scale limitations pre-determined that the landslide inventory should be established at a scale of 1:20 000 with key areas (no more than 50%) at 1:10 000 for St Lucia, and at 1:20 000-scale for Grenada. This scale limitation affects the mappable minimum size of landslides. Based on the experience of mapping landslide polygons in St Lucia using the Pleiades high resolution images as guidance the following limitations apply. At a scale of 1:20,000 the minimum mappable size is approximately 30 by 30m or 900m2. At 1:10 000 this enhances to mappable polygons of about 15m side lengths (225m2). At a scale of 1:5 000 it is possible to map elements with effective minimum dimensions of about 10m (100m2), depending on the terrain. In many areas landslides were observed that were smaller than the mappable size and these were captured as landslide clusters. Visibility of a feature is also dependent upon terrain slope. On very steep slopes the plan dimensions of an event may be much larger than the minimum dimensions discussed above, but the intersection with a near-perpendicular view may become so small that detection is not feasible. Dimensions of the polygons are taken from a ‘flat’ map and are not adjusted for slope. These simplifications affect the cumulative frequency-area distribution (see Validation results).

Landslide activity can result in the disturbance of vegetative cover and exposure of soils at the surface. Many of the landslides in the inventory were triggered by Hurricane Tomas (30/31 October 2010; Pmax ~ 400–600mm). This hurricane was of an intensity comparable to a 1:180 year event, but as it was preceded by drought conditions it is estimated that the combined likelihood ‘drought/rain’ exceeds 1:1000 years (ECLAC 2011[6]). As a consequence, the resultant disturbance of the landscape was much more severe than could be expected on the basis of the severity of the hurricane alone. ‘Landslide’ is a generic descriptor for slope movements including rotational slide, planar slide, debris flow, mud flow, debris avalanche. Generally these take place in deeply weathered materials, where for dry soil conditions a rapid infiltration can lead to a sudden loss of strength, the initiation of slope deformation and a rapid transition from sliding to flow.

To map a particular landform as a landslide requires a landslide scar and/or landslide deposits to be visible on the satellite image. Mapping of landslide events is in the first instance on the basis of simple spectral/colour signatures. In the case of the relatively low resolution RapidEye images this is not very reliable and results therefore in rather low confidence mapping. The better resolution offered by the Pleiades image enables much more detailed interpretation leading to much higher confidence in the final mapped product. As this exercise involved the establishment of a multi-temporal landslide inventory, the detailed Pleiades image could therefore be used to enhance the overall confidence of the final product.

As the differences between exposed soils and vegetated surface are quite distinct the use of automatic classification of ‘bare earth’ sites was tested in a part of St Lucia to aid the landslide identification process, using the 2011 RapidEye image as a pilot study. However, it was found that this approach leads to a large over-estimation of the areas affected by landsliding. Many cultivated fields are included in this automatic classification. The additional effort involved in fine-tuning the classification outweighed the benefits for image interpretation and therefore it was not pursued for other images.

General practice of mapping landslides is to investigate at a more detailed scale and then upscale to the desired level of detail. This enhances the confidence that the features are mapped correctly. The practical approach to this project therefore involved mapping the whole Island at a scale of 1:10 000 or at an even more detailed scale where features were uncertain. Outlines were established on the basis of representation at 1:10 000-scale. As a consequence of this practice it was possible outline landslide events in the size range smaller than 1000 m2 (approximately 100 events) and this has resulted in a more ‘complete’ landslide inventory. However, with its high resolution of 0.5m the Pleiades images offer interpretation of the landscape at much greater detail and there are therefore opportunities to enhance the capture of landslide polygons, both in detail of feature outlines and in number of small events (covering less than about 100 m2). The Pleiades images also offer detailed investigations of individual sites and events. In combination with other data (land use, topography, etc.) it is possible to generate highly detailed geomorphological maps that not only show the spatial extent of an event, but can be attributed with information on the likely nature of deformation and, in combination with other images, a timeline of event progression. This level of detailed interpretation falls outside the scope of work for this project but is discussed in a case study in Section 3 of this report to highlight the tremendous additional value of these new, high resolution products.

Landslide inventory establishment — St Lucia

RapidEye images from 2010, 2011, 2012 2013 and 2014 and one Pleiades image from 2014 were available for interpretation and landslide inventory establishment for St Lucia (Table 6; Figure 8, Figure 9, Figure 10, Figure 11). The satellite images are affected by variable cloud cover, particularly the 2010 and 2014 RapidEye for St Lucia. The SE quadrant of the 2013 RapidEye image for St Lucia was of particularly poor quality and this affected determination of small landslide events.

Figure 8 Satellite images used for landslide event identification for St Lucia. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved and material ©2014 BlackBridge, all rights reserved.
Figure 9 Capture of a small scene in St Lucia (Chateau Belair, image width about 400 m) illustrating the variations in quality of the images (from left to right) RapidEye 2010, 2011, 2012, 2013 and Pleiades 2014. Streams (blue lines) and a rural path (black lines) are added to aid cross-referencing. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved and material ©2014 BlackBridge, all rights reserved.
Table 6 Summary of landslides identified for each year of the multi-temporal image stack from 2010–2014 in St Lucia.
Year, image, date Landslides Notes
2010 RapidEye 18/8/2010 27 active, 2 inactive The image captures the state of the Island before Hurricane Tomas.
Some 40% of the land surface was affected by cloud cover (and associated shadow and could‐fringe effects). The small number of landslides that were captured where the land surface is visible is likely the result of much vegetation regrowth masking previous events. Many inactive sites, susceptible to re‐activation, are therefore not included.
Absence of older images (closer in time to major disturbing events such as hurricanes and troughs) limits opportunities of extend the size of this initial dataset.
2011 RapidEye 03/1/2011 1025 active, 3 inactive This image captures all events generated by Hurricane Tomas (October 30–31, 2010).
Thick cloud covers approximately 2% of the Island. For a further 5% the view is obscured by cloud fringe effects. Approximately 10% of the land surface is affected by cloud shadows, but this did not obstruct interpretation significantly.
2012 RapidEye 29/9/2012 489 active, 304 inactive Only 14 new events were identified (these polygons were not recognised as active in the 2011 inventory). An additional 30 events are identified as active, but initiation of slope instability is uncertain; twenty events were identified in areas where cloud cover was encountered in the images of 2010 and 2011; ten events did not exist in the 2010 inventory and were obscured by clouds in the 2011 inventory.
Some 16% of the land surface is not visible due to clouds with a further 2% obscured as a consequence of cloud shadow effects.
2013 RapidEye 14/2013 198(238)* active, 173 inactive *SE quadrant of the Island is not included in first value; the second value represents a larger total where persistence of landslide activity is plausible (i.e. all these polygons are active in the inventories of 2011, 2012 and 2014).
Approximately 35% of the land surface is not visible due to cloud cover, with a further 17% obscured as a consequence of shadows and poor image quality.
2014 Pleiades 25/2/2014 459 active, 311 inactive Some 129 landslides were new events not included in the 2011 inventory.
Approximately 5% of the land surface is not visible due to cloud cover, with a further 25% slightly affected by a thin clouds and shadows that only slightly affect image interpretation.
All images 1233 Total number of polygons in the multi‐temporal inventory (including 2 polygons mapped as inactive throughout the period 2010‐2014). 50% of the landslides were mapped as active in one year only. 27% were observed in two years, 16% in three years and 6% in four years. Less than 0.5% was observed in all five years of the period 2010‐14.
Figure 10 Pie diagrams of the number of polygons classified as active, inactive, not a landslide and those where identification was not possible due to cloud cover or, in 2013, due to poor image quality in the SE section of the Island.
Figure 11 Landslide inventory maps of St Lucia showing the distribution of active landslides (NB identification is affected by cloud cover and a poor quality SE quadrant for 2013; see Table 3).

Field checking involved a six-day field investigation in the first week of October 2014. The identification of routes and areas to concentrate on was facilitated by the ECLAC (2011)[6] report. This report provided excellent guidance in terms of the impact of Hurricane Tomas in 2010.

While in the field, the team was able to access all available digital information on a GPS-enabled laptop. The routes taken are shown in the figure below. The aim was to see as many potential landslide sites as possible in this short period of time with a focus on the following landslides sites (Figure 12):

  • The area in the vicinity of the Roseau Dam
  • The valleys and hillslopes of Fond St Jacques and Migny
  • Road-side landslides in the Colombette
  • The Micoud/Thomazo/Barre de L’Isle road
  • Landslide clusters in the Ti Rocher, Trois Pitons
  • Landslide events in Marc
  • Rockfalls off the Pitons
  • Detailed interpretation of a landslide at Chateau Belair
Figure 12 Map of St Lucia with landslide events in purple outline, identified from RapidEye (acquired in 2010–2013) and Pleiades (2014) satellite imagery. Field checking routes taken are shown in brown. Green dots represent land-use classification field check spots.

At the Roseau Dam many landslides were observed from the 2011 image and the road provided good access to the interior. Along the way it was possible to field check many sites mapped as landslides, with those generated in 2010 close to the dam still clearly recognisable (Figure 13). The road was in many places affected by recent landslides, including on the stretch from the Roseau Dam to L’Anse la Raye (Figure 2).

The Fond St Jacques area was heavily affected by flowslides triggered by Hurricane Tomas. Many events originated in deforested, cultivated fields in the upper slopes (ECLAC 2011[6]). The Migny road was severely affected and remains out of service. This area was used to evaluate the potential of a ‘bare earth’ classification for landslide identification (Figure 14).

The Colombette landslide was initiated in the upper parts of the flanks of Mount Tabac, north of Soufriere. The deeply weathered pyroclastic bedrock and lightly cemented ash soils rapidly disintegrated to form a debris slide stripping the lower slopes of vegetation, soil and roadway structure (ECLAC 2011[6]). Satellite images clearly show the outlines of the landslide in 2011 through to 2013. However, the 2014 image, albeit providing greater detail, is partly obscured by clouds. If earlier images had not been available, it is unlikely that this landslide would be detected on the basis of the 2014 image alone (Figure 15).

The Micoud/Thomazo/Barre de L’Isle road traverses the steep terrain of the centre of the Island and forms an essential transport link between Vieux Fort and Castries. Along this road many landslides are known to occur and these are subject to substantial stabilisation works (Figure 16).

A debris flow in the Ti Rocher, Trois Pitons area was identified on the satellite image of 2011. The translational slide/debris flow has a length of some 300m. The highest point is at approximately 230m above sea level and its runout drops by more than 110m. It originated in weathered bedrock comprising andesite, basalt and some agglomerates. Local soils belong to the Bocage Stony clay. On the satellite image of 2012 a substantial part of the lower part of the event was overgrown, making identification very difficult. It shows that, unless captured close to the event occurrence, recognition of landslides is very difficult in an environment where recolonization of affected slopes by vegetation occurs in a very short period of time (Figure 17).

The area around Marc was identified in the ECLAC (2011)[6] report as being particularly affected by landslides. Many of these were small translational or rotational events in deeply weathered bedrock and lightly cemented, mainly granular soils. The landslides occurred on slopes steeper than approximately 25 degrees and rapidly disintegrated to form flows. The events seriously affected communities where the houses were constructed on the hill-slopes (Figure 18). Identification of individual events is difficult in this area because of the patchwork of colours from housing, infrastructure and variations in vegetation on steep slopes and the relatively poor resolution of the 2011 RapidEye image (5m). In order to map these very small events with some degree of confidence at a 1:10 000 scale the higher resolution of the 2014 Pleiades image is required. However, by the time this image was taken, many of these smaller events were re-vegetated and their signatures difficult to establish. It is not impossible to map these small events using the images available, but it requires interrogation of the data at scales that are much more detailed than stipulated for this exercise.

The Pitons form arguably the most charismatic images representing St Lucia. These steep rock slopes are generating rock falls and several trails were mapped following Hurricane Tomas. Since then the interpretation was downgraded to ‘inactive’. However, during the field visit a loud rockfall was heard and the scars of recent events were observed. Local narratives report regular rockfalls from the Petit Piton. It is evident that this area remains one of continued activity and could benefit from careful observation and monitoring (Figure 19 and Figure 20).

The Chateau Belair site (approximate location 719400/1527030) has been used to evaluate opportunities that exist for detailed interpretation of satellite images. Comparison of the area of interest at three different scales (1:20k, 1:10k, 1:5k and 1:1k) shows how polygons drawn around landslide signatures range from very course outlines around possible multiple events (this affects the size frequency distribution by over-emphasising larger events) to very detailed metre-scale outlines of surface features. At scales of 1:5k to 1:1k it is possible to create detailed outlines of areas where evidently planar slides disintegrate into flows and where small slides rapidly transfer into debris flows down steep gradients. This is particularly facilitated by the high-resolution of the Pleiades image (see Figure 21). Comparison of the interpretation performed using this image with the stack of RapidEye images of previous years enables determination of the time at which small landslide scars are initiated.

The satellite image interpretation initially leads to identification of surface features, but further investigation using a digital elevation model shows that landslide activity at this site is affected by a topography determined by a much larger ancient (and potentially relict) rotational landslide. Combining all information enables the establishment of a detailed geomorphological sketch map that can provide useful information on the changes in activity of deformation at a remote site (Figure 22 and Figure 23) and considerable detail of morphological features of individual events (Figure 24). These interpretations require substantial field verification and the Chateau Belair site was therefore visited in October 2014. Field observations corroborated the satellite based interpretation, and this provided further confidence in the approach taken.

Access to remote sites can be difficult and time consuming. In the case of Chateau Belair, access was particularly problematic as many roads leading into the centre of the Island were compromised by the landslides of Hurricane Tomas in 2010 and the December Trough in 2013. Chateau Belair is situated at the head of a valley with only a small, unpaved road leading up to an adjacent hill where it is possible to obtain an overview of the site. To find suitable locations where a good overview of a site can be achieved on the ground is not an easy task in an environment blessed with exuberant vegetation (Figure 23). There is therefore much merit in the use of satellite images to enable interpretation of features at remote locations.

Figure 13 Landslides near the Roseau Dam. Multiple events in 2010 seriously affected the water quality of the reservoir and a large landslide occurred to the east of the dam (see inset photo from ECLAC 2011[6]). The remains of this landslide are still clearly visible. The direction of the photo is indicated by a red arrow on the satellite image of 2014. Compared with the 2011 image it is evident how much more detail can be observed. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved and material ©2014 BlackBridge, all rights reserved.
Figure 14 The Fond St Jacques/Migny area on the 2011 RapidEye image. The light coloured pixels indicate the result of a ‘bare earth’ classification (areas larger than 300m2). The red polygons represent the 2010 landslide inventory and the purple polygons the multi-temporal inventory where bare earth signatures in valleys and fields have not been included. The two landslide polygons in the centre were generated in 2014. Includes material ©2014 BlackBridge, all rights reserved.
Figure 15 Colombette Landslide. Top left is the state of the upper part of the landslide in October 2014 and top right shows the landslide in 2011. The lower images represent scenes from RapidEye (2011 and 2013) and Pleiades (2014). Despite the greater resolution of the Pleiades image, the landslide is barely visible. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 16 Landslide stabilisation works along the Vieux Fort-Castries road at Thomazo.
Figure 17 The landslide/debris flow event of Ti Rocher. To the left an oblique of the event is shown (source ECLAC 2011[6]). The event is clearly visible on the 2011 RapidEye image, while only a year later all landslide deposits below the road are covered by vegetation. Includes material ©2014 BlackBridge, all rights reserved.
Figure 18 Landsliding near Marc. The oblique photo on the left (source ECLAC 2011[6]) shows the extent of the area affected. The blue outlines in the 2011 image follow the outline of the valley. In the 2014 Pleiades image the landslide complex is barely recognisable. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved and material ©2014 BlackBridge, all rights reserved.
Figure 19 The Petit Piton with fresh trails of rockfalls and the 2011 rockfall trails superimposed on the 2014 Pleiades image. In the SE corner of the image landslide trails were observed in the field (Figure 17) but these could not be identified on the satellite images. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 20 An example of landslide scars along the main ridge connecting the two Pitons. These events were generated during Hurricane Tomas but could not be picked up in the satellite images because of size, terrain steepness, shadow effects and overhanging vegetation.
Figure 21 An example of mapping at different scales at Chateau Belair using the 2014 Pleiades satellite image as an example (see also Figure xx for an indication of image variability across the multi-temporal image stack). At 1:5 000 scale it is possible to create a detailed outline of freshly exposed soils of a landslide (A), at 1:10k it is possible to roughly outline a small event (C) while at 1:20 000 scale a small cluster of linked events is grouped together (B). Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 22 A geomorphological sketch map produced using Pleiades and a surface model. This illustrates the opportunities that are on offer given time to interrogate these information sources at their maximum level of detail.
Figure 23 The Chateau Belair landslide as seen from a vantage point during field verification.
Figure 24 The 2011 image of a landslide enables establishment of just an outline of a landslide feature near the Roseau Dam. However, the 2014 Pleiades image can be used to draw a tentative morphological map of a landslide complex. Field checks are required to ensure these interpretations are realistic. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved and material ©2014 BlackBridge, all rights reserved.

Landslide inventory establishment — Grenada

For the creation of a landslide inventory in Grenada RapidEye images from 2011 and 2012, a Pleiades image from 2013 and a Landsat TM from 2014 were used (Figure 25). The 2012 RapidEye image is particularly affected by significant cloud cover.

Figure 25 Satellite images used for landslide event identification for Grenada. Includes material ©2014 BlackBridge, all rights reserved.

Satellite image interpretation and landslide inventory establishment for Grenada followed the same process as used for St Lucia. However, it was found that the terrain in Grenada is substantially more complex and that landslide signatures are not as clear. The satellite interpretation resulted in identification of 109 sites potential landslides. However, confidence in all 109 mapped landslide polygons was extremely low.

The interpretation of the satellite images was made more difficult as agricultural landuse created a dense patchwork of cultivated fields, such as small banana plantations, that take on multi-temporal signatures resembling unstable terrain with a gradually re-establishing a vegetation cover. Even in forested areas of the upland regions of the Island there are patches of disturbed vegetation that could be interpreted (on the basis of experience gained at St Lucia) as old landslide scars.

Field checking involved a two-day field investigation on Saturday 27th and Sunday 28th September 2014. The routes taken are shown in the figure below. The aim of the fieldwork was to see as many potential landslide sites as possible and the focus was on the following clusters of potential landslides (Figure 26);

  • La Fortune and Levera region
  • The uplands of St Patrick between Union, Castle Hill and Tricolet
  • The hills around Mt Qua Qua and the Grand Etang
  • The valley from Marigot leading up to Concord Falls
  • The road between Gouyave and Grenville
Figure 26 Potential landslides (in purple outline) identified from RapidEye (acquired in 2011–2012) and Pleiades (2013–2014) satellite imagery. Following field checking (routes taken are outlined in dark blue and red) of the mapped polygons, only one landslide polygon remained — a red polygon near Grand Etang representing an old, inactive landslide.

La Fortune and Levera region. In the NE of the Island several coastal landslides were identified (tentatively) from satellite images on the basis of morphology and exposed soils. Field checking of these sites near Antoine Point indicated that the dominant processes involve surface erosion and that there was a lack of vegetation development along the coastal cliff (Figure 27). A cluster of potential landslides associated with an unpaved road system were identified (again tentatively) on the satellite images. Their features were quite persistent and some morphologies suggested rather substantial movement, on relatively gentle terrain (Figure 28). Instead it turned out that road construction near Levera point has resulted in extensive erosion in soft bedrock (tuffs) with overlying fractured basalts (members of the Levera volcanics) flowing down into erosion gullies. Topography and exposed soils observed in the vicinity of Lake Antoine could lead to classification as a potential landslide, but field inspection concluded that this site was not affected by landsliding (Figure 29).

The uplands of St Patrick between Union, Castle Hill and Tricolet. During the brief fieldwork phase this area was investigated at, but none of the sites offered sufficient evidence to warrant a landslide classification. Instead, the patterns observed in the satellite images from 2011, 12 and 13 are most likely the result of cultivation of these slopes. Many of the discounted landslides represented small banana plantations. It was checked whether these were coincident with old landslides (as was observed at a small number of sites in St Lucia), but this did not appear to be the case (Figure 30, Figure 31).

The hills around Mt Qua Qua and the Grand Etang comprise steep slopes in the Mount Granby Volcanics (Miocene-Pliocene). The vegetation cover is still re-establishing itself following the devastation caused by Hurricane Ivan of September 7, 2004. The relatively young vegetation enabled some views across the hills from the footpaths along the Grand Etang and leading up to Mt Qua Qua. It is apparent that the morphology of these steep slopes has involved slope deformation processes, but during the field visit it was not possible to determine positive indications of landsliding in this landscape, with the exception of one event on the southern slopes of Mt Qua Qua. The landslide represents the only polygon on the Island where there is confidence in the mapped product. Further interrogation of the satellite images following the field survey resulted in discounting all the other (tentatively) mapped polygons (Figure 32).

The valley from Marigot leading up to Concord Falls and the road between Gouyave and Grenville. Discussions with people at Concord Falls (two contractors working on the road and two people from the Visitor Centre) revealed that landslides are not recognized as posing a threat. The 1991 rock fall killing 14 people is still seen as an exceptional event. The only other event that was recounted involved a small rockfall in July/August ago that caused an accident, but details were very vague. There were no recollections of any significant landslides in the area, either along the road or on the higher slopes. Several small landslide events were observed during the field investigations. These include landslides along steeply incised river channels along the road to Concorde Falls (Figure 33, Figure 34). These shallow translational failures are generated following undercutting by a small stream. Often, the valley floor deposits in which these small failures develop are capable of sustaining near-vertical slopes and comprise angular blocks in a coarse matrix. This provides substantial interlocking, frictional resistance and drainage. Some degree of cementation may also assist the preservation of these steep slopes.

Slope failures in weathered bedrock can contain a greater fines content which is likely to affect slope drainage negatively and will result in a quite different post-failure behaviour and morphologies (For example as observed on the slopes of Mt Qua Qua, Figure 32). Upon further investigation, none of the small failures that were observed along the roads could be identified on the satellite images and therefore these were not included in the final product (Figure 30).

Additional observations. A large landslide complex near Palmiste Bay has led to the closure of the coastal road and a lengthy diversion inland (the Mount Nesbit Detour; Figure 35). This complex was not picked up during the satellite investigation. Upon inspection of the closed road there were few visible signs of landsliding along this road. Some sections were not paved, probably because of recurring deformation, but at the time no cracks or other signs of landslide activity were observed. Further investigation of satellite images did not enable the drawing of a clear outline of this mass movement complex. Palmiste Clays are well known to generate landslides and the 2006 landslide inventory and hazard mapping exercise highlighted this geological material as scoring particularly in the hazard index (CDB and CDERA, 2006[7]).

Signs of small rockfalls are observed along many of the Island’s rock cliffs. Immediately opposite the arrivals/departures pick up area at the airport a clear sign of this potential risk is clearly indicated (Figure 36, Figure 37). This cliff in pyroclastic deposits is a clear example of the kind of cliff faces that are prevalent throughout the country. Widely spaced persistent and intersecting joints often provide for a fragmented rock face where blocks can get dislodged during or following adverse weather conditions. These events pose a significant hazard, but are too small and ephemeral to be identified through satellite image interpretation.

The ground-truthing thus resulted in the observation that none of the 109 mapped landslide polygons were landslides, with just one exception — a large inactive rotational slide on the upper slopes of Mt Qua Qua. The EO interpretation and associated field investigation exercise has shown that unless there are very clear signs of displacement, most landslide event signatures (that in St Lucia could be interpreted with some confidence as being landslide related) are in fact associated with cultivation of these slopes.

Figure 27 Southern cliffs of the headland at Antoine Bay and landslide polygons mapped using the Pleiades 2013 image. Vegetation alignment indicates strong coastal effects reducing vegetation growth potential near the edge of the cliffs and wind and water erosion of local soils leading to exposure of the underlying bedrock. The small landslide to the north of the headland was also dismissed from the landslide record. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 28 Levera Point. Initially mapped as landslides, ground truthing indicated that these are predominantly surface erosion features. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 29 Landscape looking north from the tufaceous explosive crater rim of Lake Antoine in NE Grenada.
Figure 30 Typical landscapes in St Patricks (Union/Castle Hill) showing a patchwork of cultivated fields where several potential landslides had been mapped (see inset map with landslide polygons in purple outlines and the route taken in blue). Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 31 A field in the St Patricks area. This was initially mapped as a landslide because of the position in the landscape and a bare earth signature that gradually re-vegetated.
Figure 32 View looking north of the upper slopes of Mt Qua Qua and the outlines of the only positive identification of a landslide. The steep slopes that make up this ridged landscape are characterised by morphologies that are indicative of phases of landsliding and surface erosion. However, during the current mapping exercise, no evidence of active mass movements was found.
Figure 33 Shallow translational and rotational landslides generated by undercutting along the road to Concorde Falls and, right, a small roadside failure NE of Constantine. These small failures were generated in 2014 but are not visible on the latest satellite image available for this study as these would have occurred since the images were taken. However, given the dimensions of these events, it is highly improbable that these will be picked up by an EO survey.
Figure 34 A field along the Concord falls road. This field is typical of many in this area that have been picked up during the satellite image interpretation as potentially being a landslide. The identification as potential landslide or cultivated field are affected by slope, shadows and overhanging, patchy vegetation.
Figure 35 Road closure at Palmiste Lane, south of Gouyave. This closure sign is positioned at the red dot on the coastal road, as indicated in the image scene on the right. A lengthy detour (the Mt Nesbit Detour) reconnects with the coast road, just south of Cuthbert Peters Park. Unfortunately, satellite interpretation alone is not sufficient to draw a boundary around a region of unstable slopes with confidence. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.
Figure 36 The potential for rockfall is apparent at many sites across the island. A large cliff face near Molinere Point provides an example of many steep cliff faces found throughout the Island from where it is not inconceivable that the occasional rock fall can be generated. Differences in coloration of the cliff faces provide an indication of the dynamic nature of these cliffs. However, dense vegetation along the base of these cliffs will reduce the magnitude of the zone away from the cliff where the effects of rockfalls will be felt. In the absence of vegetation, e.g. near the airport (centre) and along roads (e.g. at Levera, right) rockfalls can pose a hazard with a more direct impact pathway.
Figure 37 Some of the coastal road stretches, e.g. between Grand Roy and Dothan, are constructed along the base of old coastal cliffs. Even though wave action is no longer providing an input into the cliff dynamics, occasional rockfalls still occur.

Digital Elevation Models
Digital elevation models (DEMs) of St. Lucia and Grenada were also required to support landslide risk assessments for these two AOIs. The SOW stated that the DEMs were to be generated from stereo optical satellite imagery with a spatial resolution of 30m or better. Again, in keeping with the 'eoworld' framework, the preferred source of imagery from which to produce the DEMs was a European or Canada sensor. With no suitable archived imagery available, a tasking request was submitted to Airbus Defence & Space in early August 2014 in order to have fresh stereo Pleiades imagery acquired for both St. Lucia and Grenada. However, the timing of this request coincides with the hurricane season in the Caribbean. As a result, all attempted acquisitions to date have been affected by considerable cloud and haze cover (Figure 38), thus rendering them inadequate for the generation of DEMs.

Figure 38 Select attempts to acquire fresh Pleiades stereo imagery for St. Lucia and Grenada. Includes material ©CNES 2014, Distribution Airbus DS / SPOT Image S.A. France, all rights reserved.

In the absence of any other alternative stereo imagery, the DEMs for the AOIs were generated based on imagery acquired by the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor. The ASTER sensor has a stereo camera that acquires nadir and backward-looking images for band 3, which can be processed using a photogrammetric approach to extract DEMs. Using an optimised approach, NASA and Japan’s Ministry of Economy, Trade and Industry (METI) have already processed an extensive archive of ASTER stereo imagery for the purpose of producing the 30m ASTER Global DEM (ASTER GDEM); released in 2011. With a view to augmenting this ASTER-derived elevation data, different strategies were developed and implemented based on the ancillary data available for the two AOIs.

For St. Lucia, ancillary elevation data derived from contour maps was made available by the Physical Planning Office and University of the West Indies. In an attempt to increase the accuracy of the ASTER-derived elevation data, a vertical calibration approach utilising the contour data was implemented. To achieve this, 32 000 corresponding ASTER- and contour-derived elevation points were extracted and modelled (R2=0.97) using regression analysis (Figure 39). All ASTER-derived elevation values were subsequently vertically calibrated using:

Hcal= (-9×10-5) x2 + 1.01x - 6.475, (1)

where Hcal are the vertically calibrated ASTER elevation values and x are the original ASTER elevation values. Next, the calibrated ASTER-derived elevation point data were merged with the contour heights to create a single x-y-z dataset. These data points were then gridded using a Triangular Irregular Network with linear interpolation algorithm to generate a 30m DEM.

Figure 39 Vertical calibration of the ASTER-derived elevation data using regression analysis.

For Grenada, ancillary elevation data was available in the form of a 5m digital terrain model (DTM) generated from airborne Light Detection And Ranging (LiDAR) data. The ASTER-derived elevation data was vertically calibrated using this LiDAR data. To achieve this, 500 corresponding ASTER and LiDAR elevation points were extracted and modelled using regression analysis (R2=0.99). To avoid introducing errors by comparing data from a digital surface model and a DTM, care was taken to ensure that only points corresponding to bare ground were selected. All ASTER-derived elevation values were subsequently vertically calibrated using:

Hcal = 1.018x - 1.016, (2)

where Hcal are the vertically calibrated ASTER elevation values and x are the original ASTER elevation values. The quality of the resulting DEM was further enhanced by replacing elevation values relating to a sizeable 'pit' artefact using elevation values from a Shuttle Radar Topography Mission (SRTM) DEM.

The resulting enhanced DEMs for St. Lucia and Grenada are shown in Figure 40.

Figure 40 The 30m DEMs generated for (A) St. Lucia and (B) Grenada.

The landslide inventory polygon datasets for Grenada and St Lucia were also compiled into map format. An example (for St Lucia) is illustrated in Figure 41.

Figure 41 Map showing the landslide inventory (2014) and digital elevation model of St Lucia.

Service 3: Digital Elevation Models of Belize

The objective of Service 3 is to generate a national DEM of Belize and a precise DEM for a subset of Belize to support hazard/risk assessment.

National Digital Elevation Model

The SOW stated that the national DEM of Belize should be generated from stereo optical satellite imagery with a spatial resolution of 30m or better. Again, in keeping with the 'eoworld' framework, the preferred source of imagery was a European or Canada sensor. Accordingly, a 20m DEM produced from SPOT-5 stereo satellite imagery was identified and acquired from Airbus Defence & Space. However, this DEM provided only 40% coverage of the total Belize AOI.

Additionally, a 30m DEM covering the entirety of Belize was generated based on existing ASTER-derived elevation data. The accuracy of the ASTER-derived elevation data was enhanced by vertically calibrating it with the higher-resolution 20m DEM. To achieve this, 32 000 corresponding ASTER- and SPOT-derived elevation points were extracted and modelled using regression analysis (R2=0.99). All ASTER-derived elevation values were subsequently vertically calibrated using:

Hcal = 0.989x – 2.610 , (3)

where Hcal are the vertically calibrated ASTER elevation values and x are the original ASTER elevation values. The quality of the calibrated DEM was further enhanced by applying a 3×3 pixel moving average filter. This filter helped to reduce “noisy” data values as well as notable 'spike' artefacts in the DEM. The 20m and 30m DEMs generated for Belize are shown in Figure 42in the form of a map also produced as part of the project.

Figure 42 The national 30m DEM and 20m DEM (in blue window) of Belize.

Precise Digital Elevation Model

The second aspect of Service 3 was to generate a high precision 1m resolution DEM from very high-resolution optical satellite imagery for a subset of Belize encompassing an area to the north of Belize City. To satisfy these requirements, a tasking request was submitted to Airbus Defence & Space in early August 2014 in order to have fresh Pleiades tri-stereo (triplet) imagery acquired for an area of 100km2 situated to the north of Belize City, encompassing the town of Ladyville. However, the timing of this request coincides with the hurricane season in the Caribbean region. As a result, all attempted acquisitions to date have been affected by considerable cloud and haze cover (Figure 43), thus rendering them inadequate for the generation of a precise 1m DEM. Acquisition attempts for this area of Belize are still ongoing, and it is intended to generate the precise DEM upon acquisition of suitable tri-stereo imagery.

Figure 43 Selected attempts to acquire fresh Pleiades tri-stereo imagery for an area of Belize. Includes material ©CNES 2014, Distribution Airbus DS/SPOT Image S.A. France, all rights reserved.

Metadata

The primary metadata for all of the products generated for the three services is summarised in Table 7. More information on the products can be found in the relevant sections of this document.

Metadata are included with the digital services, where appropriate. For example, the landslide inventory includes an attribute table of information relating to each polygon of the inventory.

Table 7 Summary of the metadata for the products.
Area of interest Product Data description Geographical coordinate system Spatial resolution/scale Thematic accuracy
Service 1 St. Lucia Land use/land cover map Attributed raster detailing distribution of 14 land use/land cover classes WGS84 UTM
Zone 20N
2m< 84.9%
St. Lucia Water bodies Shapefile (polygons) of lakes and ponds WGS84 UTM
Zone 20N
1:10 000 N/A
St. Lucia Rivers and streams Shapefile (polylines) of rivers and streams WGS84 UTM
Zone 20N
1:10 000 N/A
Grenada Land use/land cover map Attributed raster detailing distribution of 15 land use/land cover classes WGS84 UTM
Zone 20N
2m 84.8%
Grenada Water bodies Shapefile (polygons) of lakes and ponds WGS84 UTM
Zone 20N
1:10 000 N/A
Grenada Rivers and streams Shapefile (polylines) of rivers and streams WGS84 UTM
Zone 20N
1:10 000 N/A
St. Vincent and the Grenadines Land use/land cover map Attributed raster detailing distribution of 15 land use/land cover classes WGS84 UTM
Zone 20N
2m 80.8%
St. Vincent and the Grenadines Water bodies Shapefile (polygons) of lakes and ponds WGS84 UTM
Zone 20N
1:10 000 N/A
St. Vincent and the Grenadines Rivers and streams Shapefile (polylines) of rivers and streams WGS84 UTM
Zone 20N
1:10 000 N/A
Service 2 St. Lucia Landslide inventory Shapefile (polygon) attributed with landslide activity status in 2010–2014 WGS84 UTM
Zone 20N
1:20 000 (50% at 1:10 000) N/A
St. Lucia DEM Digital elevation model derived using combination of ASTER satellite imagery and height contours Horizontal datum:
WGS84 UTM Zone 20N
Vertical datum: EGM96 geoid
30m 1.4m (RMS)
Grenada Landslide inventory Shapefile (polygon) attributed with landslide activity status in 2011–2013 WGS84 UTM
Zone 20N
1:20 000 N/A<
Grenada DEM Digital elevation model derived using combination of ASTER satellite imagery and airborne LiDAR data Horizontal datum:
WGS84 UTM Zone 20N
Vertical datum:
EGM96 geoid
30m 4.9m (RMS)
Service 3 Belize DEM Digital elevation model derived using SPOT satellite imagery Horizontal datum:
WGS84 UTM Zone 16N
Vertical datum:
EGM96 geoid
20m 7m (RMS)
Belize National DEM Digital elevation model derived using combination of ASTER and SPOT satellite imagery Horizontal datum:
WGS84 UTM Zone 16N
Vertical datum:
EGM96 geoid
30m 9.8m (RMS)

References

  1. Helmer, E H, Schill, S, Pedreror, D H, Kennaway, T, Cushing, W M, Coan, M J, Wood, E C, Ruzycki, T and Tieszen, L L. (2007). Forest formation and land cover map series-Caribbean Islands. U.S. Geological Survey/Earth Resources Observation and Science (EROS), Land Cover Applications and Global Change, International Land Cover and Biodiversity, Caribbean Land Cover Analyses.
  2. Helmer E H, Kennaway T A, Pedreros, D H, Clark, M L, Marcano-Vega, H, Tieszen, L L, Ruzycki, T R, Schill, S R and Carrington, C M. 2008. Land cover and forest formation distribution for St. Kitts, Nevis, St. Eustatius, Grenada and Barbados from Decision Trees classification of cloud-cleared satellite imagery. Caribbean Journal of Science, 44, 175–189.
  3. Haralick, R M, Shanmugan, K and Dinstein, I. (1973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, 3, 610–621.
  4. Herold, M, Liu, X and Clarke, KC. (2003). Spatial metrics and image texture for mapping urban land use. Photogrammetric Engineering & Remote Sensing, 69, 991–1001.
  5. Richards, J. A., & Jia, X. (2006). Remote Sensing Digital Image Analysis, Fourth edition, Springer-Verlag, Berlin, pp. 232–242.
  6. 6.0 6.1 6.2 6.3 6.4 6.5 6.6 6.7 ECLAC Report (2011). Saint Lucia: Macro socio-economic and environmental assessment of the damage and losses caused by Hurricane Tomas: a geo-environmental disaster. Towards Resilience. Economic Commission for Latin America and the Caribbean. 183pp.
  7. CDB & CDERA (2006). Final Project Report: Development of landslide hazard maps for St Lucia and Grenada. The Caribbean Development Bank (CDB) and the Caribbean Emergency Response Agency (CDERA). February 2006, 53p