Lessons learned from regional temporal land cover mapping

By Debbie Jewitt from Ezemvelo KZN Wildlife, and Mark Thompson and Lungile Moyo from GeoTerraImage

Land cover maps provide a description of the Earth’s surface[1], characterising both anthropogenic activity and eco-climatic diversity[2]. Land use change is a key driver of global environmental change and biodiversity loss[3] and the rapid rate of change is undermining the capacity of the earth to provide essential, free ecosystem goods and services[4].

Understanding the patterns and drivers of land cover and land use is critical to investigate anthropogenic impacts on the environment[5]. Land use change may be due to socioeconomic changes or biophysical processes[6]. Multi-year land cover maps permit land-cover change analyses and can assist in identifying the drivers of change in the landscape[7] and therefore provide essential decision support for policy and management.

Ezemvelo KZN Wildlife (EKZNW) has developed a series of four provincial land cover maps (for 2005, 2008, 2011 and 2017) for conservation and to address rapid natural habitat loss in the region. EKZNW is the provincial conservation agency in KwaZulu-Natal, South Africa. Situated on the country’s east coast, the province has a biologically and physically diverse landscape, with anthropogenic land covers and land uses including agriculture, timber plantations, the built environment, mines and dams.

These four land cover maps are comparable and therefore useful in analysing land cover changes, conducting trend analyses and determining rates of habitat loss.

Improving satellite imagery and the rapid developments in free and open satellite data acquisition, including increased spectral, spatial and temporal resolutions[1], and cloud-based processing and analysis capabilities[9], are increasing the capacity to rapidly develop land cover maps. With each new iteration of land cover map, we have sought to improve the maps based on lessons learned during the development and analysis of these land cover maps.

Developing of the provincial land cover maps

A variety of classification techniques and ancillary data were used to develop the land cover maps. To ensure direct compatibility of the map series, the resolution of the land cover maps was standardised at 20 m irrespective of the source imagery resolution.

Higher resolution satellite imagery was spatially resampled to 20 m (using cubic convolution algorithms) from the source formats before classifying the imagery. The land cover datasets are suitable for 1:50 000 scale or coarse mapping.

The minimum mapping unit associated with the thematic land cover classes is ±0.25 ha, i.e. all classified land cover units smaller than this spatial extent were digitally dissolved into the appropriate adjacent land cover class.

All the maps include a 1 km buffer along the terrestrial borders of the province to minimise edge effects, especially for analyses sensitive to edge effects such as corridors and linkage development.

Where possible, errors discovered in the previous land cover maps were corrected before updating to newer time periods. Where better imagery existed, it was used to correct time-stable components of the previous maps, e.g. cloud cover or burn scars in an image.

The 2005 land cover map was derived from SPOT 2/4 satellite imagery[10], whereas the 2008[11] and 2011[12] maps were derived from SPOT 5 multispectral imagery.

The bulk of the new mapping relied on the availability of the existing land-cover dataset, against which areas of change were determined and classified. Updating existing maps is more time and cost efficient than developing new land cover maps from scratch.

The 2017/14 land cover map was a major departure from the previous maps, due to a change in source imagery from SPOT to Sentinel 2. ESA’s Sentinel 2 satellite imagery with its high temporal frequency has enable land cover mapping based on true multi-seasonal imagery. This allowed for significant improvements in landscape interpretation and thus mapping accuracy[13]. Similarly, cloud-based imagery archives from Google Earth Engine along with cloud-based processing significantly improved processing time.

Fig. 1: The 2017 land cover map of KwaZulu-Natal, South Africa.

Standardising legends

The legend categories were kept consistent across all the maps, with the exception of some sub-class categories that provided more detail but were hierarchical in nature, e.g. distinguishing natural water bodies from dams[11]. Temporally stable features such as dams that were identified at newer imagery were edited into the earlier maps.

South Africa recently adopted new hierarchical national land cover classes and definitions in order to standardise legend classification schemes across the country. While this is a good initiative, it poses a challenge for time series analyses of historical datasets, unless the older legend categories can be aligned to the new classification scheme. In most cases this is feasible using class amalgamations in one or both of the legend structures.

Improving accuracy assessments

Accuracy assessment techniques have improved over time.

The increased mapping accuracy of the 2017 land cover map is attributed to the enhanced spectral resolution of the Sentinel 2 imagery compared to SPOT imagery, the availability of true multi-seasonal imagery, and possibly the use of sample reference points for the accuracy assessment.

The 2005 land cover map combined reference data from pre-classified road and aerial transects, digital aerial photography and GIS coverages, to aid both image interpretation and class boundary mapping. Later land cover maps relied solely on aerial assessments of (semi-oblique) aerial surveys between 500 and 700 m above ground-level.

Mapping accuracies were determined using statistical analysis (standard contingency matrices) and field reference data comparisons, and ranged between 78.9% and 97.7% (see Table 1).

Class-specific errors were highest with the natural vegetation sub-classes, especially where gradients in cover density existed (e.g. the transition from “wooded grassland” to “grassland”), or with misclassification of structurally similar vegetation types. Similar errors were found between “rural clusters” and “urban settlement” where class definitions depend on user-defined thresholds.

YearNo. of reference pointsMapping accuracy at 90% confidence limitsKappa Index
200573883.06% (81.26 – 84.86)81.5
2008100178.92% (77.24 – 80.60%)78.14
201196683.51% (81.95 – 85.07%)82.92
2017226497.70% (96.26 – 98.15%)97.64
Table 1. The number of reference points used in the accuracy assessment, the mapping accuracies and Kappa Index for each land cover map.

Accounting for technological change

Free, high-temporal and high-spatial resolution satellite images, and free cloud-based technologies such as Google Earth Engine, have revolutionised the development of image-generated land cover maps. These improvements resolve issues such as burn scar mapping, distinguish shadow areas, provide cloud free images, better delineate vegetation types and cultivation, and identify permanently bare areas.

Despite yielding an improved land cover map, this could present challenges for time-series analyses. To ensure consistent land cover maps for time series analysis, another version of the 2017 EKZNW land cover map was created that had been modified to retain mapping improvements of Sentinel 2 imagery, but still ensured compatibility and comparability with the SPOT-generated land cover datasets.

Keeping track of anthropogenic category changes

As later land cover maps were developed, it became evident that certain anthropogenic categories (e.g. old cultivated fields, rehabilitated mines and old timber plantations) were reverting to secondary natural vegetation classes. This is because spectrally the secondary vegetation classes are almost identical to primary vegetation classes, and therefore difficult to distinguish.

From a conservation perspective, these secondary grasslands do not hold the same biodiversity value as primary vegetation, especially in terms of specialised species such as terrestrial orchids and geophytes[8]. Hence it was essential to track these secondary vegetation areas over time. Additional legend categories were created to track secondary vegetation in the land scape.

Fig. 2: An example of additional agricultural fields identified in 1990 but not detected in the 2017 map.

Since Agricultural expansion was extensive in South Africa prior to the 1960’s due to agricultural subsidies and low selling prices[8], historical agricultural fields in the province circa 1960/70 were identified and mapped using scanned topographical maps[15]. This data was used to inform the change analysis (see Fig. 2).

Approaches to change analyses across multiple land cover maps

Change analyses errors could arise from real changes in the landscape or be because of map errors[16]. Good map accuracy is therefore essential[17]. Change analyses should also consider the accuracy of categorical transitions during a time interval, and not just at a single point in time[5]. Since errors are additive, change analysis using multiple land cover maps could increase total number of errors. It therefore becomes difficult to confirm the overall accuracy for the analysis over the full time period.

  • One solution is to use shifting rules between categories and analyse the change rationality in a post-classification comparison[18].
  • Another option is to only analyse the start and end time points – but then valuable information on drivers of change in the landscape could be lost. For example, KwaZulu-Natal’s timber plantations initially increased in the landscape, but then stabilised due to catchments being closed to further water extraction[8].
  • Another alternative is to aggregate classes, especially those with lower accuracy scores. Aggregation also helps to reduce the number of small classes but could lead to a loss of information[16].
Fig. 3: Loss of natural habitat in KwaZulu-Natal, South Africa, between 1990 and 2017.

In measuring the loss of natural habitat in the landscape over time (Fig. 3 & 4), we aggregated the land cover categories into two classes: natural vegetation (e.g. grasslands, savannas, forests, wetlands) and anthropogenically transformed classes (e.g. cultivation, mines, the built environment, timber plantations). We used the earliest available national land cover (1990)[19] as a baseline dataset and followed the methodology in Jewitt et al[8], which accumulates transformation over time, i.e. if a pixel was an anthropogenically transformed category in an earlier time it was not permitted to become natural at a later time; instead it is assigned to a secondary vegetation classification.

This was done specifically for biodiversity conservation to identify primary vegetation rather than secondary vegetation. Based on this approach, 8.1% of natural habitat was lost between 1990 and 2017. If earlier landscape transformations are ignored, natural habitat lost is calculated to be 13.90%.

Fig. 4: Loss of natural habitat in KwaZulu-Natal, South Africa, between 1990 and 2017.

Improving mapping accuracy with carefully considered rulesets

To increase mapping accuracy, rules restricting land cover transitions can be applied[18].

Several rulesets were used in the analyses and interpretation of the land cover maps. For example, water bodies were always mapped to their long-term maximum extent (excluding flooding), irrespective of potential drought conditions. This ensured direct comparability over time. It also ensured that when counting dam features in the province, that these were not over-estimated due to water feature fragmentation by emerging substrate deposits or vegetation in drier years.

Rulesets should be applied carefully though, as seemingly obvious transitions may be legitimate changes in the landscape, and could therefore introduce errors.

A transition rule could limit a category transition from a built environment to a natural environment. For example, in KwaZulu-Natal we have observed an abandonment of rural buildings, driven by socio-economic and safety factors. These buildings were often broken down, allowing for alien plants and secondary vegetation to replace them. Local insight and photographic assessment showed this legitimate transition in the landscape and helped limit the introduction of errors into the land cover map.

Scope for improvements

Future improvements in the land cover maps relate to adequately accounting for degradation in the landscape. The maps currently cannot discern alien invasive plant species, which are extensive in the province[20]. Similarly, we cannot detect a loss of floristic composition in landscape from inappropriate management practices such as grazing and fire. Currently we can only detect gully erosion, but mapping sheet erosion remains elusive at our scale of mapping. In more arid areas of the country such as the Karoo, land cover maps don’t reflect land use change which is extensive in these systems[21].

The value of comparable maps now and in years to come

Land use and land cover change is a leading driver of habitat loss and ecosystem degradation[22, 23], and it is therefore essential that accurate, consistent time series land cover maps are developed to identify and track the drivers and processes of land cover change.

As Wadsworth et al[24] point out, repeat land cover maps developed using same methods are rare due to changing technology, science and policy objectives. They suggest using consistent mapping so as to compare disparate maps.

So far, Ezemvelo KZN Wildlife has favoured map consistency and compatibility, but future technological advancements will need to integrate multiple technologies to meet new needs. Similarly, the application landscape will continue to evolve, driving the rapid advancement of land cover mapping technology.

Building on the lessons learned from previous land cover maps developed, developing rulesets and using ancillary data, have helped in developing a valuable time series of land cover maps for the region to guide conservation initiatives and policy.

References

Read more