We are testing a new beta website for gov.scot go to new site

Assessing the Extent and Severity of Erosion on the Upland Organic Soils of Scotland using Earth Observation: A GIFTSS Implementation Test: Final Report


4. Image Processing and Data Preparation

The successful classification and understanding of the opportunities presented by earth observation require careful initial preparation of the satellite imagery. This comprises three main technical elements; geometric, atmospheric and topographic corrections.

These processes allow us to align the satellite images with other datasets ( e.g. in a GIS), then it allows for a consistent application of classification techniques by removing the earth atmosphere effects from the imagery.

Errors in any of these corrections compromise the extraction of trajectories of reflectance and associated measures and the successful application of the decision rules, particularly where these require temporal imagery.

Technical details of the image processing that have been applied are presented in Annex 2.

4.1 Image correction

Geometric Correction

The key priority is to register images acquired by sensors with different orbital and observing configurations ( e.g., spatial resolutions, viewing angles). All imagery supplied were geo-referenced to the British National Grid.

To do this, we generated Erdas Imagine ground control point ( GCP) files. Using independent procedures available in Erdas Imagine we orthorectified and compared the corrected imagery by evaluating against known features ( e.g., cairns, wall boundaries, field boundaries, buildings, road junctions), as identified using OS MasterMap data, aerial photography and NEXTMap ORI data.

In the registration process, we sought to minimise resampling by using image to map registration for all scenes (and referring to the GCP database). All images were resampled to 5 m using near neighbour techniques. We consider that images resampled using cubic convolution might be preferable for segmentation as edges are better preserved but nearest neighbour resampling is still best for classification, even though pixels are often removed or duplicated.

Once completed the geometrically processed imagery is correctly positioned (see Figure 4) and is by definition, co-registered with all other imagery and spatially referenced datasets (Figure 5).

Figure 4. Registration of SPOT5

Figure 4. Registration of SPOT5

(a) Unregistered SPOT5

Figure 4. Registration of SPOT5

(b) Registered SPOT5

Figure 5. Co-registration of SPOT 5 and ASTER satellite imagery

Figure 5. Co-registration of SPOT 5 and ASTER satellite imagery

Atmospheric Correction

All imagery supplied was calibrated to radiance (W m 2 sr -1 µ -1) and subsequently to surface reflectance (%). According to the plan, it was vital that these stages were delivered, and delivered accurately, to ensure the successful development of the rule base. Our previous work has demonstrated that atmospheric correction of optical ( e.g., SPOT and IRS sensor) data to surface reflectance is a critical stage in developing consistent decision rules and classifications. Using ENVIFLAASH, the following tasks were completed:

  • A sensitivity analysis of key input parameters ( e.g., visibility) required by the FLAASH algorithm.
  • Reference to Meteorological Office data on the state of the atmosphere at the time of image acquisitions.
  • Inclusion of aerosol loadings (from 2000) reported on a daily basis based on observations from coarser resolution optical sensors, namely MODIS and MISR.
  • Identification of temporally invariant dark and bright reflectance that can be used to evaluate the success of the atmospheric correction.

The result of this processing is a set of imagery that has comparable reflectance data (see Figure 6), ready for classification, allowing;

  • Comparison of reflectance data acquired by different sensors observing on the same date.
  • The spectral reflectance curves associated with vegetation to be better interpreted in terms of the photosynthetic pigment amounts (in the visible channels), internal leaf structure (near infrared) and moisture content (Figure 7).
  • Vegetation indices and endmember fractions can be calculated with greater confidence and inconsistencies associated with the atmosphere were removed.

Figure 6. SPOT5 and ASTER reflectance curves for the Monadhliath Mountains

Figure 6. SPOT5 and ASTER reflectance curves for the Monadhliath Mountains

Figure 7. Typical spectral reflectance curve for vegetation.

Figure 7. Typical spectral reflectance curve for vegetation.

Topographic Correction

Topographic correction (presented as units of reflectance, %) help compensate for differences in illumination as a function of time of year, time of day and slope and aspect. This can prove particularly useful for mountainous areas. As assessment was made of the study areas at the start of the project as, whilst the Monadhliath is clearly upland, there were not significant areas of steep slope that were going to affect the classification process. As a result, this correction has not been applied to the imagery, but it is likely to be an important part of the image processing stages for a wider roll out of any EO solution.

4.2 Generation of derived layers

For the classification of the landscape within the Monadhliath, over 30 EO-derived features were utilised, in the form of either a derived layer, or as a 'customised feature' created within Definiens Developer software.

Derived data layers to be used within the classification included vegetation indices (namely the Normalised Difference Vegetation Index; NDVI), endmember fractions (of bare soil, water/shade, photosynthetic vegetation and non-photosynthetic vegetation; NPV) and topographic features (slope, aspect and convexity). The former were generated through using the fully processed EO data and applying linear spectral un-mixing or band math procedures to each image. The topographic features were obtained from the NEXTMap DTM data using a specialist topographic feature extraction tool, within ENVI.

Other indices were generated within the Definiens Developer software itself by generating an 'arithmetic function' using both the equations of known vegetation indices, and through the use of indices created in previous landscape classifications by Environment Systems.

With vegetation indices being a combination of surface reflectance at two or more wavelengths that are designed to highlight a particular property of vegetation, each of the vegetation indices generated will accentuate a particular vegetation property and therefore aid in the separation of the various classification classes.

The topographic layers will be used to locate and distinguish between habitats located on steeper and shallower slopes and those inhabiting 'sink' areas in the landscape where water commonly accumulates. Particularly this will enable discrimination to vegetation groups such as bogs and mires from grassland areas.

4.3 Development of classification classes

In order to produce a classification of peat erosion it was necessary to split the vegetation up into classes which were of relevance to peat stability and the surrounding area. Several standard systems exist for describing detailed vegetation classes on areas such as the Monadhilaith Mountains e.g. National Vegetation Classification (Rodwell 1992) or Phase 1 habitats ( JNCC). However these classification systems are not suitable as they stand, for use with satellite imagery. Phase 1 mapping includes land use as well as land cover classes, which are problematic to map from space.

Experience on work in Wales (Lucas et al., (in press)) has also shown that the Phase 1 classes are too broad, and classes need to be established that can be broken down into areas of similar species coverage's, or similar mixes of species that are identifiable spectrally from the satellite data. The NVC classification on the other hand are too detailed and many classes depend on the presence of plants which are very small in size and infrequent in occurrence throughout the sward, and again these cannot be identified from space. Instead a pragmatic approach was taken to dealing with the vegetation classes as this project was about looking at soil erosion classes rather than habitats. We therefore chose the vegetation classes based on the amalgamation of land cover types that could be separated into the fewest groups by the spectral data from the remote sensing imagery and the air photography. The broad habitats that resulted from this were named with descriptive titles which outline the sort of plant communities that could be expected to fall within these areas.

A full description of all classes, including definition, photo and erosion risk category is presented in Annex 3.

4.4 Selection of testing and training sites

Field work planning

A systematic route was planned around the study area for identifying the erosion features and associated land covers. Analysis was carried out on the number of field samples that needed collecting for each class to ensure that enough replicates were collected to allow for a statistically valid accuracy assessment to be completed at the end of the project. It was established (Congalton and Green) that 50 replicates were needed for each class, with the opportunity to over sample the key classes ( e.g. eroding blanket bog) and to under sample classes that we are mapping simply due to their location in relation to peat erosion; but which we are less interested in e.g. acid grassland.

Working with knowledge of the Grieve study, we targeted areas that had been previously visited and surveyed for peat; these areas are NH3500, NN4090, NH5005, NH5505 and NN5095 (see Figure 8 for a subset and Figure 9 for the complete area).

Replicates were established for each of the classes (as mapped in draft stage during Phase 1 of the project) and a full set of field maps (based on topographic, satellite, derived imagery and air photos) and recording sheets were produced. Access permissions from the Estates and Glendoe Hydro Scheme were obtained and basic daily routes established for five full days in the field. A full health and safety risk assessment was undertaken.

Field working

The total number of points visited (267) is shown in Figure 8. Each field work day was designed to collect the best coverage of data across the whole study area, whilst collecting enough data for each replicate. The pre-defined sample points were used to structure each day, with supplementary points then collected during transects between each (Figure 8).

A Trimble Nomad GPS (~5 m accuracy) and Garmin eTrex Summit (~15 m) were used to record the location of in-situ data. The Nomad was used to record smaller features, with the Garmin used to record larger, more homogeneous areas.

At each point a record sheet was filled including; unique ID, grid reference, class, description, picture (with photo ID). In addition, areas of relevant peat/associated habitat were mapped on the field sheets.

A key element of the field work was ecologists working with remote sensors. It was only by combining expert knowledge of the manifestation of different surfaces ( e.g., vegetation and soils) within remotely sensed data and derived products ( e.g., shade fractions, the NDVI) that a full understanding and therefore accurate recording could take place.

Post field work

After the field work, all the data collected was brought together and analysed for completeness. All in-situ points were georeferenced and associated photos checked. The data associated with these objects were then divided into a "training" dataset that can be used to refine and/or develop decision rules and a reserved "testing" dataset that was used to independently evaluate the success of the decision rules in classification.

Figure 8. Comparison of in-situ data points

(a) Survey points, with topographic map

Figure 8. Comparison of in-situ data points

(b) Survey points, with SPOT 5 image

Figure 8. Comparison of in-situ data points

Figure 9. Location of in-situ data collection

Figure 9. Location of in-situ data collection

4.5 Image segmentation

Segmentation techniques are not a new domain within computation, they are however a relatively new way of classifying remote sensing data. Historically, a process of manual digitisation of areas of interest following an extensive field campaign would be undertaken.

However with the increased necessity for high-spatial resolution imagery analysis, and the availability of commercial or non-commercial software packages, segmentation processes have become increasing useful as these enable a rapid separation of image pixels into a representative form of the real work component they depict (Blaschke et al., 2004).

In high spatial resolution imagery, pixels when grouped together can better represent the characteristics of land-cover groups far better than single pixels can. Therefore groups of adjacent pixels will be organized into objects which will in turn be treated as a minimum classification unit (Yu et al., 2006). These objects are defined as basic entities which are located within an image, where each pixel group is composed of similar digital values, and possesses an intrinsic size, shape, and geographic relationship with the real-world scene component it models. Therefore, the objects can be said to be spectrally more homogeneous within individual regions than between them and their neighbours.

The Definiens Developer multi-resolution segmentation algorithm creates image or grid data segments which are primarily based upon three criteria: scale, colour and shape (smoothness and compactness). Scale is the heterogeneity acceptance within a segment whilst colour, smoothness and compactness are all variables that accentuate the segments spectral homogeneity and spatial complexity.

A comparison of the segmentation process within Definiens Developer has been noted to have performed more to reality than many others have (Neubert and Meinal 2003). By emphasising the scale rather than compactness, it can allow for polygons to follow natural features more naturally. When shape information was strongly emphasized instead of colour, the resulting polygons were unstructured and did not closely follow feature boundaries.

As with previous work, segmentation has been undertaken within Definiens software. Most peat areas will occur in the uplands outside of any SIACS boundaries. For areas outside of the SIACS boundaries, objects of only a few pixels have been generated using the spectral information contained within the imagery. Within discrete areas, consideration will be given to the use of finer (< 5 m) spatial resolution e.g. aerial photography data in complex environments. Where available (given the open, upland nature of the area), we used the SIACS (field boundary) data to align the segments with the boundaries of farm units ( e.g., fields) and OS MasterMap layers relating to buildings, roads and rivers will also be used to guide the segmentation.

The segmentation procedure involves the creation of three levels which will be used to classify based upon the level of detail required. Within the upper level (Integration Level), all external dataset ( i.e.OS MasterMap) were synced within a chessboard segmentation procedure and classified out. The multi-resolution segmentation algorithm was then used to segment the remaining areas of the image to aid in the removal of all other areas not associated with peat erosion features ( i.e. Woodland, Non Peat vegetation) and to separate the areas of blanket bog.

This segmentation was then carried down to create an EO level and a finer multi-resolution segmentation within the SPOT data applied to areas of blanket bog to aid in the classification of smaller peat features within the EO data.

Finally within the bottom level (Air Photo Level), a very fine multi-resolution segmentation using the aerial photography was applied to pick out the peat erosion features present within the imagery. This segmentation was applied within the areas noted to be peat erosion classes within the EO layer.

Figure 10. Levels of segmentation employed.

Figure 10. Levels of segmentation employed.