DEVELOPMENT OF A NEW SOUTH AFRICAN LAND-COVER DATASET USING AUTOMATED MAPPING TECHINQUES. Mark Thompson 1 1 GeoTerraImage Pty Ltd, Pretoria, South Africa Abstract This talk will discuss the development of a new South African Land-Cover dataset that will be generated from remotely sensed imagery using automated mapping techniques. Automated modelling procedures offer significant advantages in terms of data standards, processing time, repeatability and accurate change detection; when compared to more conventional image mapping approaches. Building on GeoTerraImage s significant past experience in land-cover mapping, a new automated mapping approach has been successfully developed. This has now been implemented as an operational procedure for generating base-level land-cover patterns from multiseasonal Landsat 8 imagery circa 2013-14. This process allows rapid production of foundation cover classes that can be easily converted into more detailed land-cover information categories. Furthermore the technique can be retrospectively applied to historical Landsat imagery in order to generate historical land-cover datasets for change detection. This method is now being used to produce a new 2013 South African national land-cover dataset, which will be followed by comparable historical land-cover data. Introduction Land-cover is one of the key information requirements for a wide range of landscape planning and management activities, ranging from environmental resource management to telecommunication planning. The ability to rapidly generate, accurate, reliable and repetitive landcover data has therefore wide commercial applicability and interest. Landsat 8 Catalyst The recent global availability of Landsat 8 imagery (April 2013) was the primary catalyst behind the development of the new GeoTerraImage (GTI) automated land-cover modelling approach, since it offered a free and regular supply of radiometrically and geometrically standardised, medium resolution, multi-spectral data, suitable for medium to large area mapping. Collectively this offered an ideal opportunity to re-look at automated land-cover mapping techniques as an alternative to more conventional analyst-assisted per-pixel classifiers. Overview of the Automated Modelling Approach Automated modelling procedures offer significant advantages in terms of ensuring data standards, minimising processing time, allowing easy repeatability and facilitating accurate change detection; when compared to more conventional image mapping approaches where there is a greater reliance on individual image analyst knowledge and inputs. 1
Building on GTI s significant past experience in land-cover mapping and image-based landscape interpretation across South African and southern Africa; a series of automated modelling steps were developed that utilised the seasonal dynamics associated with broad landscape characteristics. These could then be used to rapidly produce a set of foundation cover classes that could be easily converted into more meaningful land-cover information categories, using pre-defined geographic masks in the GTI data libraries. The foundation cover classes were defined as the basic building blocks associated with all landscape characteristics, namely water, bare ground, grass and tree-bush-shrub cover types, with each being defined in terms of seasonal occurrence or permanence. These basic foundation cover classes represented the initial output from the automated modelling approach. The foundation cover classes are then converted into more conventional land-cover information classes, i.e. urban, forest plantation etc, as part of the post-automated modelling data processing steps. The foundation cover classes are referred to as spectrally dependent classes, since they are generated from automated modelling procedures that are based directly on the spectral characteristics associated with each image pixel over time (i.e. seasonal) and space (i.e. within an image frame). The final land-cover information classes are referred to as the spectrally independent classes since different cover classes typically share similar foundation class spectral characteristics in a one-to-many type relationship. For example, the bare ground spectrally dependent cover classes could represent non-vegetated built-up urban areas, natural rock exposures, beach sand or a mine pit and tailings dump. Similarly the tree-bush classes could represent a natural vegetation cover, a timber plantation or a fruit orchard. The advantage of this approach is that the conversion of the initial, spectrally dependent foundation cover classes into the spectrally independent land-cover information classes can be tailored to suit a variety of end-user information requirements; simply by using a different set of pre-determined masks and foundation class sub-divisions and amalgamations. For example the completed 2013-14 South African foundation class dataset has already been converted into a product containing information required for regional telecommunication planning (see Figure 1). An alternative version aligned with typical national land-cover information requirements for environmental applications is currently in process. 2
Figure 1. Completed Telecommunication Planning Land-Cover Dataset for South Africa, Swaziland and Lesotho, generated from Landsat 8 imagery acquired between April 2013 to March 2014 (see Appendix 1 for legend information). Although the development and initial objectives of the automated modelling was focussed on using Landsat 8 imagery within South Africa, subsequent testing of the modelling approaches on Landsat 8 imagery in other parts of Africa (i.e. Sudan, Zimbabwe, Namibia and Mozambique), and on archival Landsat 5 imagery over South Africa indicates a high level of model portability. This should allow and support the production of directly comparable historical land-cover datasets for change detection, assuming of course that the required level of seasonal image coverage is available in the data archives. No attempt was made to use object-based modelling in the automated mapping approach, primarily because the medium resolution format of multi-spectral Landsat 8 imagery does not lend it self to this type of modelling, since the pixel resolution typically precludes the identification of true landscape objects in comparison to high and ultra high resolution image formats. The 30 metre Landat resolution format means that pixels typically represent a mix of land-cover characteristics rather than a pure cover surface, i.e. an urban pixel is typically a composite of building roofs, garden vegetation and / or road surfaces etc. However, object-based modelling using the foundation class dataset (or derived product versions), as an input, may be an ideal approach for helping to generate the 2 nd stage information land-cover classes, for example by separating water in rivers (i.e. natural) from water in dams (i.e. artificial), based on size, shape and context etc. 3
Landsat 8 Data Sourcing The automated modelling approach utilised a range of multi-date Landsat 8 images per frame in order to characterise the seasonal dymanics associated with the various foundation level cover classes. As far as possible, a minimum of 7 and maximum of 9 image acquisition dates per image frame between April 2013 and March 2014 were used, although localised, prolonged cloud cover problems resulted in some image frames having less than this. Flexible land-cover models were developed that could run on any number of input datasets, although preference was always to have as many inputs dates per frame as possible in order to better represent the seasonal landscape changes and generate a more reliable and representative output. Nine acquisition dates per frame per year was seem as the optimal target number, based on Landsat 8 s 16-day overpass schedule. In general it was possible to acquire between 7 and 9 acquisition dates for most image frames, although prolonged, local cloud cover problems in a few frames required either the use of less Landsat 8 frames, or the inclusion of seasonally suitable, archival Landsat 5 images. Note the Landsat 5 imagery was only used if it was from a suitable seasonal period to compliment the Landsat 8 data, and was 100% cloud free. Figure 2. Location of the 10 x image frames that required the use of archival Landsat 5 imagery as a result of no suitable Landsat 8 data acquisitions during the period April 2013 to March 2014. 76 x Landsat image frames were required to provide complete coverage of South Africa, Swaziland and Lesotho. 616 x individual Landsat images were used to model and produce the land-cover data, representing and average of 8 x acquisitions per image frame per year. Of the 616 x images, 592 x 4
images were from Landsat 8 (i.e. 96 %) and 24 x images (i.e. 4 %) were from Landsat 5. The Landsat 5 images were however only used in 10 of the 76 image frames defining the geographic extent of the land-cover dataset. The distribution and location of image frames that required the use of archival Landsat 5 imagery is shown in Figure 2. All Landsat 8 (and 5) imagery used in the land-cover data modelling was sourced from the on-line data archives of the United States Geological Survey (USGS, http://glovis.usgs.gov/). The data are provided in a precise geo-corrected UTM / WGS84 map projection format, and was used as-is without any further geo-correction. Landsat 8 Data Preparation All Landsat imagery was standardised to Top-of-Atmosphere (ToA) reflectance values prior to model use. As far as possible only cloud free or image dates with limited cloud cover were used in the modelling (i.e. maximum ± 20% terrestrial cloud cover in any one date). Any cloud affected regions were corrected by merging the ToA corrected data with cloud free ToA reflectance data from either the preceding or following overpass date, so that as far as possible, the final cloud-free merged imagery composite only represented a maximum difference of ± 16 days. This was deemed acceptable in terms of minimising any changes in local vegetation cover growth changes. Approximately 35% of the 616 x images used in the land-cover modelling were cloud-masked composites. No external atmospheric correction was applied to the datasets. Spectral Modelling Automated land-cover modelling was based on spectral indices (generated from the ToA data), rather than the original ToA reflectance values. A standardised set of spectral indices were identified from which the required foundation cover classes could be modelled, using predetermined generic spectral threshold values. The generic threshold values used in the indices were tested over several landscapes and seasons before being confirmed and accepted as such. The spectral indices included both existing algorithms such as the Normalised Difference Vegetation Index (NDVI) and Normalised Difference Water Index (NDWI), as well as algorithms developed in-house specifically for the GTI land-cover modelling requirements. All models were developed in ERDAS Imagine image processing software using the Model Maker function. The modelling and generation of each foundation cover class was undertaken as a separate modelling exercise, i.e. water was modelled separately from bare ground, which was modelled separately from tree and bush cover etc. This approach simplified the modelling steps and facilitated easier desk-top quality control of outputs; compared to attempting to model all foundation cover types simultaneously within a single model workflow. The final geographic extent of each foundation cover class were generally defined using the combined output from several different spectral indices since no single index was found to work well in all landscapes and all seasons. The primary foundation cover classes that were generated are: 5
Water Bare ground Grass dominated vegetation Tree and/or bush dominated vegetation Tree dominated vegetation All modelling was undertaken on an individual image frame-by-frame basis. This allowed the models to be adapted according to the number of images and associated seasonal ranges available for that frame. The model modifications reflected the need to standardise the outputs according to pre-defined definitions of seasonal permanence, for example, whether a foundation cover class occurred: in only one image date, in several, but not all image dates, in all image dates. Obviously the more image acquisition dates available per frame, and the wider the seasonal range, the more accurate the modelled interpretation of a particular cover classes seasonal characteristics. Examples of the separate model outputs for the bare and tree-bush foundation cover classes are shown in Figure 3 (a,b). The examples show eastern Pretoria, from within Landsat 8 frame 170-078. Figure 3. Modelled outputs for bare (left) and tree-bush (right) foundation cover classes, over eastern Pretoria (frame reference 170-078). Colour tones represent different levels of seasonal permanence in terms of pure bare ground (whites) versus sparse vegetated ground (purples), and tree (dark greens) versus tree-bush (light greens) categories. 6
In some of the far western arid areas, additional tree-bush modelling was undertaken using a slightly modified modelling approach in order to better detect the sparse, low bush and shrub covers in these landscapes. Terrain Modifications Terrain modifications were included within some of the foundation cover class modelling procedures (i.e. water, bare and tree-bush classes), in order to minimise seasonally induced terrain shadowing effects, using a combination of solar illumination and slope. These parameters were modelled independently from the 90m SRTM dataset, with outputs being re-sampled to a Landsat 8 comparable 30m resolution format, before being incorporated into the foundation cover class models. For example, solar illumination and DEM slope masks were used to minimise any spectral confusion between dark terrain shadow areas and comparable water body reflectance levels. Seasonally Defined Land-Cover per Image Frame The modelled outputs for each foundation cover class were then combined into a single output for that frame, where the output cover class combinations were described in terms of integrated seasonal occurrences. For example: water in all dates (permanent), water in many dates (seasonal), tree-dominated in all dates, tree-dominated in many but not all dates plus bare ground in one date, Figure 4 shows the output after combining all the primary foundation land-cover classes into the seasonally defined land-cover classes. The illustrated area is the same eastern Pretoria area as shown in Figure 3. The full list of 51 x seasonally defined foundation cover classes generated at this point is shown in Appendix 2. Note that in addition to the above primary foundation classes already, fire scar occurrence was also modelled and incorporated into the final combined foundation cover composite classes, in order to provide an indication of whether or not fire was a factor in the overall seasonal profile. 7
Figure 4. Seasonally defined foundation land-cover classes, eastern Pretoria (frame reference 170-078, same as Figure 3). See Annexure 2 for legend information Compiling the National Seasonally Defined Land-Cover The completed seasonally defined land-cover datatsets for each image frame were then merged into a single dataset covering all of South Africa, Swaziland and Lesotho. The mosaicing procedure was analyst defined to ensure, as far as possible that optimal overlap rules were used to achieve seamless matching of land-cover characteristics between adjacent image frames. Whilst the standardised modelling approaches supported generation of spatially and content comparable land-cover data in adjacent image frames (if both frames utilised comparable input image acquisition dates), if there was a significant difference in the number or seasonal range of input images, then adjacent images often showed edge matching differences. In such cases, the adjacent images were often spatially comparable, but not necessarily content comparable. In these cases, the edge matching process was set-up, as far as possible so that the most prominent and spatially extensive landscape pattern was maintained across the image borders, in order to provide a more visually seamless image frame overlap. Figure 5 illustrates the single, merged seasonally defined foundation land-cover classes covering South Africa, Swaziland and Lesotho. 8
Figure 5. Seasonally defined foundation land-cover dataset covering South Africa, Swaziland and Lesotho, based on Landsat 8 imagery, 2013-2014. As can be seen in Figure 5, there were still some seasonally induced boundary anomalies within the merged dataset. These were a result of limited availability of Landsat 8 within the 2013-2014 wet season, which, in terms of the generic modelling thresholds, resulted in an under-estimation of grassland and over-estimation of single or multi-date tree-bush cover in the affected image frames. This was corrected with a follow-up corrective grass model, based again on pre-defined generic spectral indices modelling, but with locally modified spectral thresholds that resulted in a more representative grassland delineation in the problem image frames. Figure 6 illustrates the grass-corrected version of the spectrally dependent foundation land-cover classes covering South Africa, Swaziland and Lesotho. The remaining edge anomaly visible across the Highveld and Lesotho regions represents the difference between 2 x levels of grass cover and associated seasonal profiles. In terms of the spectrally dependent foundation land-cover terminology they are separate classes, but in reality they actually represent different levels types of grass cover which can be amalgamated into a single grassland extent in subsequent modelling steps used to generate spectrally independent land-cover datasets, such as the illustrated telecommunication planning dataset shown in Figure 1. 9
Figure 6. Grass-corrected, seasonally defined foundation land-cover dataset covering South Africa, Swaziland and Lesotho, based on Landsat 8 imagery, 2013-2014. The dataset illustrated in Figure 6 represents the base from which all other land-cover related products can now be generated, for example the telecommunication planning dataset illustrated in Figure 1. The spectrally dependent cover classes in this base dataset represent the building blocks, from which more information focused land-cover classes can be derived, especially if combined with ancillary datasets such as vegetation or land-use maps to facilitate further class subdivisions or area-based amalgamations. Modelling Speed and Data Processing Efficiency A distinct advantage of a model-based mapping approach is that it is significantly more time efficient than more conventional, analyst-assisted mapping techniques. The entire South African, Swaziland and Lesotho data coverage, from pre-modelling ToA and cloud-masking preparation of the Landsat 8 imagery to completion of the illustrated telecommunication planning dataset (see Figure 1) was completed within ¼ of the comparable time it would have taken using conventional desk-top image classification techniques. Human resources consisted of a dedicated team of 4 experienced image analysts, able to work full time on the project, with two additional analysts able to participate on a partial basis. Model Portability Outside of Southern Africa In order to verify the portability of the automated modelling procedures in landscapes outside of South Africa, several tests were done using comparable Landsat 8 imagery (again downloaded from the USGS data archives). The test results illustrated below in Figures 7 and 8 show a full Landsat frame processing over Khartoum, Sudan and sub-frame area over a region in Paraguay. 10
Both results clearly indicate the generic ability of the existing models to be used successfully elsewhere without any modification whatsoever to the existing spectral thresholds developed under South African conditions. Note that the illustrated examples have only been processed to the initial, spectrally-dependent foundation cover classes within a single frame, and do not represent final spectrally independent land-cover information classes. Figure 7 shows the modelled land-cover generated from 7 x Landsat 8 acquisition dates for frame 173-049, within which the city of Khartoum (Sudan) is in the lower left corner at the confluence of the White and Blue Nile rivers. The landscape is basically sand and rock desert, with extensive irrigated cultivation. Figure 7. Automated land-cover outputs over Khartoum, Sudan (Landsat frame 173-049). Figure 8 shows the modelled land-cover generated from 6 x Landsat 8 acquisition dates for frame 224-077, over a section of Paraguay, South America. The landscape contains patches of previous forest cover intersected with a mosaic of commercial cultivation, with a major urban area in the eastern section. 11
Figure 8. Automated land-cover outputs over Paraguay, South America (image subset from Landsat frame 244-077). 12
Appendix 1 Legend for telecommunication planning land-cover dataset illustrated in Figure 1. Class Class Name Description 1 Water All areas of open water that can be either man-made or natural in origin. Based on the maximum extent of water identified in all seasonal image acquisition dates. 2 Bare Bare, non vegetated areas dominated by loose soil, sand, rock or artificial surfaces. May include some very sparse scattered grass, low shrub and / or tree and bush cover. Can be either natural (i.e. beach) or man-made (i.e. mines or built-up areas). 3 Low Vegetation & Grassland Grass and low shrub dominated areas, typically with no or only scattered trees and bushes. Mainly natural or seminatural vegetation communities in both urban and rural environments. May also include some subsistence cropping fields 4 Tree / Bush Dominated Vegetation Low tree and/or bush dominated areas, typically with lower canopy heights and more open canopy densities (i.e. open to scattered) than class 5 (below). Includes natural, semi-natural and planted vegetation communities in both urban and rural environments. Will also include young planted forest plantation stands 5 Tree Dominated Vegetation Tall tree and bush dominated areas, typically with higher canopy heights and more compact canopy densities (i.e. open to closed) than class 4 (above). Note: may include dense thicket, even if not composed of tall trees & bushes. Includes natural, semi-natural and planted vegetation communities in both urban and rural environments. Will also include mature planted forest plantation stands and windbreaks. 6 Cultivated Large-scale, commercially cultivated fields used for the production of both annual and permanent crops (i.e. maize, sugarcane, orchards etc). The class includes both rain-fed and artificially irrigated fields. The class does not include small-scale subsistence type cultivation. 7 Sports, Golf and Parks Managed grassland areas associated with golf courses, sports fields and urban parks. 13
Appendix 2 Seasonally defined integrated foundation cover classes generated for each Landsat 8 image frame. 14