Detection of land use / cover changes of the KOSH region over a period of 14 years using the South African National Land Cover datasets for 2000 and 2014

Simple algebraic change detection techniques viz. image difference and image ratio were applied to the South African national land use / cover (NLC) datasets of years 2000 and 2014, prepared in grid format covering the Klerksdorp–Orkney–Stilfontein–Hartebeestfontein (KOSH) region in order to assess land use/land cover changes. Both the 2000 and 2014 NLC datasets were generated from Landsat images using different classification schemes and the code values & attributes of the land cover classes of the two datasets were different/not comparable. In order to make these datasets comparable for change detection, the NLC2000 dataset was examined in ArcView GIS by superimposing it onto the NLC2014 dataset and similarities and differences were identified. For each cover type of the NLC2000 dataset, comparable cover type of the 2014 dataset was identified by making a query to the NLC2000 dataset and after viewing the spatial distributions of selected units in respect of the NLC2014 dataset. Suitable code values of NLC2014 dataset were identified for the NLC2000 dataset and it was later reclassified. The land use / cover change detection study reveals that increase in areas were observed for the cover types: Cultivated common fields (low), Cultivated common fields (med), Mines 2 semi-bare, Wetlands, Urban commercial and Plantations/woodlots mature. The Grassland, Thicket/dense bush, Urban residential (dense trees/bush), Mines 1 bare, and Cultivated common pivots (high) showed a decrease in places. During the 14 years, Grassland had decreased from 2,132.47 km2 (77.35% of the total area) to 1,629.78 km2 (59.11% of the total area) owing to landscape transformation to other land covers (e.g. Cultivated common fields and Urban residential) due to human activities. The percentage increase in areas observed for the Cultivated common fields (low and medium) were 8.21% and 2.96% while the Mines 2 semi-bare, Wetlands, Urban commercial, Plantations/woodlots mature showed increases of 0.67%, 0.32%, 0.28% and 0.23% respectively. The area of Thicket/dense bush decreased from 108.15 km2 to 56.71 km2 (change of 1.87%). Maps of land use/land cover changes and statistics obtained for the changed areas are very useful for identifying various changes occurring in different classes and for monitoring land use dynamics.


Introduction
The whole surface area of a region may be classified in terms of land use/ cover.Land use/ cover refer to categories of features described by the vegetation, water, natural surface and cultural features on the land surface (Thomas, 2001).Land cover influences the hydrological cycle, energy balance and carbon budget, as many different physical characteristics such as albedo, emissivity, surface roughness, photosynthetic capacity, and transpiration change as a function of land cover (Zhu and Woodcock, 2014).
The land cover datasets have a wide range of usefulness including landscape planning, natural resource management activities and protection of natural environments.Land cover change may be natural or anthropogenic, but with increasing human activities, the earth's surface has been modified significantly in recent years as a result of changes in land cover and use.Knowledge of land cover and land use change is necessary in order to model the earth system and its environments (for example by studying aspects such as hydrological processes and climate change) and for many purposes related to management.
Remote sensing data consisting of airborne and satellite observations of the land surface provide insight into land changes, in order to identify the factors which cause these changes, and in order to predict future changes (Boriah et al., 2008).Land cover change detection essentially entails identifying when the land cover at a given location has been converted from one type to another.Examples include the conversion of forested land to barren land (possibly owing to deforestation or a fire), grasslands to golf courses or farmland and farmland to housing developments.The study of land cover change is quite important because of its impacts on local climate, hydrology, radiation balance, and the diversity and abundance of terrestrial species (Boriah et al., 2008).Remote sensing data are primary sources extensively used for change detection studies in recent decades (Lu et al., 2004).Often, the land use/cover data refer to data that result from the classification of satellite data into "land use and land cover" categories based on the reflectance value of the satellite image showing the use of the land and the cover types present therein (Essic, 2005).One important area where remote sensing plays a key role is the study of land use/land cover and change affecting these factors.
Land cover change detection using GIS and remote sensing techniques normally involves adopting a pre-classification change detection technique or a post-classification change detecting technique.In a pre-classification approach by Haque and Basak (2017), Change Vector Analysis (CVA), Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) analysis were implemented to assess the change scenario.The post classification change detection involves mainly image pre-processing, image classification (either supervised or unsupervised) analysis, interpretation, ground-truthing or field verifications, refinement of image classification and some GIS analysis/processing to come up with a land use / land cover map for a particular period and doing the same processing and analysis for another satellite data sets of a period of interest and later comparing the derived land cover datasets visually or analysing using digital change detection techniques to identify any changes in spatial extent of the mapped features.
A variety of digital change detection techniques have been developed over the past three decades and incorporated as algorithms of change detection in most of the remote sensing software (Coppin et al., 2004).The different algorithms may be grouped into the following categories: algebra (differencing, rationing, and regression), change vector analysis, transformation (for example principal component analysis, multivariate alteration detection and Chi-square transformation), classification (postclassification comparison, unsupervised change detection, expectation-maximisation algorithm) and hybrid methods.(Nori et al., 2008).A comprehensive exploration of all the major change detection approaches implemented as found in the literature is presented by Lu et al. (2004).
Mining and agricultural practices are major human activities on the land in South Africa.The change of land cover due to mining and associated development and in the area (changes in land use / land cover due to human activities that are linked to mining) has resulted in significant changes in the catchment hydrology in terms of increased surface runoff and associated pollution and also depletion of surface water resources due to reduction of natural infiltration ground water recharge.The increased surface runoff in some situation also causes formation of acid mine drainage (AMD) at a faster rate (due to water easily entering into abandoned mining areas having shafts and underground tunnels etc.) than the natural conditions.South Africa has experienced many incidences of mine water pollution especially AMD in areas of past and present mining activities.The KOSH (Klerksdorp-Orkney-Stilfontein-Hartebeestfontein) region is a typical in the sense that has many mines and has faced the impacts of mining on water resources.This region is currently facing challenges of mine water management due to AMD and water ingress and the Council for Geoscience (CGS) is engaged in mine water management projects in order to mitigate the issues of water ingress and AMD.An assessment of changes of land use / cover using existing land cover datasets will help in getting some knowledge on the significant change of land use / cover faster than analyzing satellite acquired in the past and in the recent time for change detections studies as such processing involves pre-processing and image analysis for change detections using the above mentioned algorithms of the remote sensing software.Therefore, an attempt was made to detect major land use / land cover changes of the KOSH region using post classification data (existing land cover datasets of the years 2000 and 2014) in order to see whether it will have significant impact on the catchment hydrology due to mining and agriculture related human activities.This paper presents the results obtained from an attempt made to identify changes in the land use / cover of the KOSH (Klerksdorp-Orkney-Stilfontein-Hartebeestfontein) region over a time span of 14 years using the South African national land cover (NLC) datasets of 2000 and 2014.

Characteristics of the study area
The area chosen for this land use/cover change detection study is the KOSH (Klerksdorp-Orkney-Stilfontein-Hartebeestfontein) region of Northwest Province of South Africa.The study area spans 2 755 km 2 and falls within the Vaal River catchment (Figure 1).The KOSH area is located approximately 111 160 km southwest of Johannesburg.The Vaal River flows through the southeastern part of the KOSH region.Most of the area has sandy loam soil texture with an undulating relief, whereas the region south of Orkney has a flat relief (Midgley et al., 1994).The KOSH area is underlain mainly by an intercalated assemblage of sedimentary rock and extrusive rocks, porous unconsolidated and consolidated sedimentary strata, acid and intermediate intrusive rocks and basic/mafic lavas (such as dolomite, goldbearing conglomerates, Black Reef quartzite, Ventersdorp lavas and dykes) and has shallow aquifers containing uncontaminated water relatively close to the surface (Midgley et al., 1994;Pulles et al., 2005;SAFLII, 2013).The KOSH region forms part of the Witwatersrand gold mining area.Gold mining operations by a number of different gold mining companies have been undertaken in the KOSH area since 1950s (SAFLII, 2013).

Data used and methodology
The 2013-14 South African national land use/ cover (NLC) dataset, produced by GEOTERRAIMAGE (GTI, 2015) using multiseasonal Landsat 8 multispectral imagery, acquired between April 2013and March 2014, and NLC 2000data (published in 2005) were procured for this study.The 2013-2014 national land cover dataset provides 72 land cover/land use classes based on 30 x 30 m raster cells and is ideally suited to 1:75 000-1:100 000-scale GIS based mapping and modelling applications (GTI, 2015).The NLC 2000data (published in 2005) were generated from digital Landsat imagery having 30 resolution, acquired primarily from 2000 to 2001 (Schoeman et al., 2010).
The NLC 2000 data were captured as a digital raster dataset (.img file format), and it contained 49 land cover classes.
Change detection analyses describe or quantify differences between images or standardised land use/ cover datasets or classified images of the same area at different times.The classified datasets of two or three periods may be used to calculate different land cover areas at different periods while observing the changes taking place over the particular time span.As the data used in this study are two post classification products of already existing landcover datasets of the years 2000 and 2013-2014 derived from classification of reflectance values of Landsat images, there is no need to do any image preprocessing steps and image classification on these datasets.The normal approach adopted for change detection studies using post classification approach (using already prepared landcover datasets) involves algebraic change detection and comparisons of the extents of the land cover classes.The basic fast approach taken for the present land cover change assessment was to compare the standardized land cover datasets representing the two assessment years (2000 and 2014), using GIS based analysis and simple algebraic change detection techniques available in remote sensing software such as the image difference algorithm and the ratio algorithm of ERDAS IMAGINE software.
For this land use/ cover change detection analysis, the NLC 2000 dataset was considered to be the reference raster data (before image) while the NLC 2014 dataset was treated as the "after image" in that the land cover showed some changes.Change detection analyses using image difference or the ratio algorithms of remote sensing software (ERDAS IMAGINE) require comparable "before image" and "after image" datasets.In other words, the images acquired at different periods should be of the same type and the land cover classification grid datasets should have the same attribute values or codes and cover type descriptions.The NLC datasets of 2000 and 2014 had been generated from Landsat images using different classification schemes.Thus, their attribute values and land cover class codes were not the same/comparable.First, the NLC 2000 dataset was made comparable to the NLC 2014 dataset by reclassifying and recoding the comparable cover types of the reference raster dataset, based on the cover types of the changed land cover dataset.

Preparation of comparable land use/land cover data sets
The procured NLC 2013-14 data were first converted into ESRI grid format with a 30 m pixel size using a GIS.As these data had been supplied without any postclassification spatial filtering, the speckled appearance of the data attributable to the presence of isolated single pixels was reduced using the MajorityFilter command in ArcView GIS 3.3 (using a neighbourhood of the eight nearest cells with the option of obtaining a clear majority for the neighbourhood).Later, the resulting grid was clipped in ArcView GIS 3.3 using the Map Calculator function in the Analysis menu of the Spatial Analyst extension and by using the study area extent shapefile as a mask.The procured data of NLC 2000 had also been clipped/subset in ArcView GIS 3.3 using the study area shapefile as a mask and later projected onto the UTM (Zone 35 south projection) by specifying a grid cell size of 30 m using the Project Raster command in ArcGIS.The NLC 2014 dataset extracted for the KOSH region contained 53 land use/cover types (Figure 2 and Table 1), whereas the NLC 2000 dataset of this area contained only 28 land use/ cover types (Figure 3 and Table 2).Moreover, the grid values (code numbers) of NLC 2000 were completely different from the code values of the NLC2014 dataset.2) and assigned to its attribute table as a lookup table attribute.The NLC 2000 dataset was later reclassified in ArcView GIS using the lookup table of identified code values of NLC 2014 that had been matched with the older dataset and saved as a comparable grid dataset of the year 2000 (Figure 4) in order to apply change detection algorithms.

Detection of land cover changes
The detection of land cover changes was performed using simple algebraic change detection techniques available in remote sensing software such as the image difference algorithm and the ratio algorithm of ERDAS IMAGINE software.In the image difference method, registered images or the raster dataset acquired at different times are subtracted to produce a residual image that represents the change between the two dates.The image difference method of ERDAS IMAGINE requires two datasets, for example, comprising the before image and the after image.The before image is the earlier of the two images.The after image represents the more recent of the two themes and reflects change over time.

Land cover statistics per assessment year
The land use/land cover distribution shown in Table 1 4.  Later, the changes in areas of land cover during a 14 year time span were calculated (Table 6).Table 6 shows that the land use/ cover classes that have undergone significant changes over 14 years are the following: Grassland, Cultivated common fields (low), Thicket/Dense bush, Cultivated common fields (med), Thicket/Dense bush, Mines 2 semi-bare, Urban residential (dense trees/bush), Mines 1 bare, Wetlands, Urban commercial, Plantations/Woodlots mature and Cultivated common pivots (high).It is evident from this table that the area of grassland has decreased significantly (from 2,132.47 km 2 to 1,629.78 km 2 or 77.35% to 59.11% of the total area) over the course of 14 years owing to landscape transformation (changed from a natural state) to other land cover types (Cultivated common fields, Urban residential, etc.) owing to human activities.The area of Thicket/dense bush also decreased from 108.15 km 2 to 56.71 km 2 during this time.
The total area of Cultivated common fields (low) was 328.67 km 2 (11.92% of the total area) in the year 2000 whereas it had increased to 555.05 km 2 (20.13% of the total area) in the year 2014.
Similarly, Cultivated common fields (med) increased from 0.43 km 2 (0.02% of the total area) to 82.07 km 2 (2.98% of the total area) during the course of 14 years.The total area of Mines and Quarries during the year 2000 was 47.26 km 2 (1.71% of the total area) whereas, in year 2014, the total area of Mines 1 bare and Mines 2 semi-bare was 54.11 km 2 (1.96% of the total area).The major land cover classes observed in the reclassified land cover dataset of the KOSH region during the year 2000 that have % area >0.5% and the corresponding % area extracted from the NLC 2013-14 datasets are shown in Table 7.

Maps indicating areas of changed land use/cover
The image difference method of ERDAS IMAGINE computes the differences between two images or grids, highlighting changes that exceed a user-specified threshold.The after image is subtracted     Figure 6 shows areas of change highlighted as having increased and decreased by more than 10%.
The highlighted green and red patches of this map match regions of larger positive and negative values shown in Figure 5. Validation of areas of major changes can be done through a comparison with some satellite images, e.g., Google Earth Pro software.An examination of the larger green patch seen in the southeast of Orkney (Figure 5) with Google Earth Pro image of Dec 2000 revealed that this green patch was under cultivation; later, it was used as a mine waste dump in Dec 2013.
Figure 7 is the outcome of the change detection obtained using the ratio method, highlighting areas that had changed between 2000 and 2014 and the direction and magnitude of change are expressed.
In general, the resulting patches of this map match the patches in Figure 5

Conclusions
Detection of land use/ cover changes in the KOSH region over a period of 14 years was successfully mapped using the postclassification datasets of land cover distribution for the years 2000 and 2014.The land use/cover change detection study reveals that significant landscape transformation has occurred in the KOSH region during the course of 14 years with some noticeable increases and decreases in respect of land cover classes.Noticeable increases are observed for the following classes: Cultivated common fields (low) (11.92%),Cultivated common fields (med) (2.96%), Mines 2 semibare (1.71%), Wetlands (0.67%), Urban commercial (0.28%), and Plantations/woodlots mature (0.23%), whereas a major decrease is observed in respect of Grassland and Thicket/dense bush.
Grassland decreased significantly, from 2,132.47 km 2 to 1,629.78 km 2 (a decrease of 18%), mainly as a result of human activities.Similarly, the area of Thicket/dense bush decreased from 108.15 km 2 to 56.71 km 2 (a decrease of 1.87%).This study revealed land use/cover changes in the form of change detection difference maps, with areas and changes in area percentages by applying the image subtraction and image ratioing methods on postclassification images (datasets).The results seen in the resulting change detection difference maps depend on pixel-for-pixel comparisons and the accuracies of the classification datasets.The method of change detection adopted for this study generated acceptable results within the accuracies of the datasets used.The overall map accuracy for the 2013-14 South African NLC dataset is 81.73%, with a mean land cover/land use class accuracy of 91.27% (GTI, 2015).The reported map accuracy for the NLC 2000 dataset was 65.8%, (Van den Berg et al., 2008).The analysis for change of land use/cover is very helpful in monitoring the dynamics of land use/cover and in identifying various changes occurring in respect of different land use/cover classes such as an increase in urban built-up areas or a decrease in agricultural land.

Acknowledgments
The

Figure 1 .
Figure 1.Location of the KOSH area

Figure 2 .
Figure 2. Land use/cover map of the KOSH area based on NLC 2013-14

Figure 2
Figure 2 reveals that Grassland is the main category present in most of the KOSH area, whereas the Cultivated common fields (low) are seen mainly in the southeastern and northwestern parts, with scattered patches in the northern, northeastern, and southwestern parts of the study area.The Low shrubland category is seen mainly in the eastern and southwestern parts.The Thicket/dense bush is seen mainly north of Orkney, along the Vaal River and in the north.Mines are located mainly in the eastern, central and northeastern part of the KOSH region.Table3shows the area and percent area coverage of the land use/cover classes of KOSH area based on NLC 2000.A comparison of Figures2 and 3and Tables1 and 3reveal that the area coverage and spatial distribution/extent of the major land use/land covers of year 2000 are not the same as seen in the years 2013-14.The major land use/ cover categories observed in the KOSH region for year 2000 that have % area >0.5% are shown in Table4.

Figure 4 .
Figure 4. Reclassified NLC 2000 dataset of the KOSH area based on NLC 2013-14 Figures 5 to 7 show the results obtained from the image difference and image ratio methods with spatial distribution and aerial extent of the changed land use/land cover units of the KOSH region over a 14-year period.The image difference map (Figure 5) shows the differences of grid values obtained from the subtraction of the NLC 2000 dataset from the NLC 2014 dataset.A region of zero value indicates an area which did not change and which mainly represents grassland for the period 2000-2014.Areas showing positive values of 1 to 30 (blue, purple and pink) regions represent areas that were previously grassland and that had changed to other land covers in the NLC 2014 dataset.These areas have grid values greater than 6 (such as Low shrubland, Woodland/open bush, Plantations/woodlots, different types of cultivated areas and mines).Light pink regions (difference values of 31 to 70) represent areas which were previously grassland and which had changed to other land covers in the NLC 2014 dataset.These areas have grid values greater than 36 such as Urban residential (dense trees/bush), Urban township (low veg/grass), Urban commercial, Urban industrial, Urban informal (low veg/grass), Urban smallholding (low veg/grass), and Urban residential (open trees/bush).

Figure 5 .
Figure 5. Changes in land use/land cover of the KOSH region over a period of 14 years obtained using the image difference method.

Figure 6 .
Figure 6.Areas of major change in land use/land cover in the KOSH region over a period of 14 years, based on the NLC 2000 and NLC 2013-14 datasets.

Figure 7 .
Figure 7. Ratios of land cover codes showing changed areas of land use/land cover in the KOSH region over a period of 14 years.
very well, thus indicating the same pattern in the observed changes.The region showing a ratio value of 1 indicates that the area did not change and that it remained as Grassland from 2000 to 2014.The patches showing higher ratio values (>1) represents areas where there are significant increase in area of the land cover types for the year 2013-14 as compared to year 2000 (some of these patches are areas of different types of urban townships that have increased around the towns and some other patches are due to increase in areas of cultivated common fields (low)).Certain areas of low ratio values indicate areas where is decrease in the extent of grassland, and thicket/dense bush areas due to human activities.There are more patches seen in the ratio map than in the image difference algorithm, making the ratio map more useful when identifying locations of patches indicating changes in land use/land cover, especially in urban regions.
author would like to thank the Department of Environment (DEA) for making the 2013-2014 South African NLC dataset available, and Mr Harold Weepener of the Institute for Climate and Water of the Agricultural Research Council (ARC) in Pretoria for providing the South African NLC 2000 dataset.

Table 1 .
reveals that the total area of the identified area of extent for the KOSH region is 2757 km 2 .The major land use/land cover categories observed in NLC 2013-14 dataset having % area >0.5% are shown in Table2.The other categories of land use/over units including Plantations/woodland mature, Water permanent, Cultivated common pivots, Urban -Land use/cover classes of the KOSH area based on NLC 2014

Table 2 .
Major land use/land cover categories observed in NLC 2013-14 dataset

Table 3
shows that the other categories of land use/cover units of NLC 2000, including Urban/Builtup: residential, formal township and different forms of Mines and Quarries, Forest Plantations (Eucalyptus spp) etc. occupy less than 0.4% of the total area.Table 2 also shows the names and codes of the classes identified from the NLC 2014 dataset that match the classes of the NLC 2000 dataset for reclassification.Such identified classes were reclassified by using the Reclassify menu of the Spatial Analyst extension.After this reclassification of the grid data with the identified comparable classes, the final generated land use/land cover map (Figure4) shows 19 classes and their areas statistics are given in Table5.

Table 3 .
Land use/cover classes of KOSH area based on NLC 2000

Table 4 .
Major land use/cover categories observed in NLC 2000 dataset

Table 5 .
Reclassified land use/cover of the KOSH area during 2000 (based on classes of NLC 2014)

2,757,059,100 100.00
The attribute table of the reclassified NLC 2000 dataset was joined to the attribute table of the NLC 2014 dataset in order to display the total areas, percentage of each of the land cover classes of the NLC 2014 dataset compared with the corresponding classes available in the NLC 2000 dataset.

Table 6 .
Changes in the land use/land cover of the KOSH area during a time span of 14 years(Based on NLC 2014 and reclassified NLC 2000 datasets)

Table 7 .
Major land use/cover categories observed in the reclassified NLC 2000 dataset.

Table 7
reveals that grassland has decreased by 18.24% in a span of 14 years whereas the cultivated common fields (low) increased by 8.21%.The % area of Thicket/Dense bush also decreased by 1.86% in this time span.The other land cover types that has significant increase in area in a span of 14 years is wetlands (increase of 0.32%) whereas the extent of permanent water bodies has decreased in this timespan.The expansion of built up areas due to urbanization has resulted in decrease in % areas of Urban residential (dense trees/bush) and Urban township (low veg/grass).The extents of mining areas were more in the year 2000 as compared to 2013-14.There were more areas under Cultivated common pivots (high) in the year 2000 than in the year 2013-14.
Areas showing negative values of -2 to -4 (green) represent areas which were initially Grassland, Low shrubland, Thicket/Dense bush, Cultivated common fields (low), and Cultivated common pivots (high) in 2000 and which had changed to Cultivated common fields (high), Low shrubland, Wetlands, Urban sports and golf (dense tree/bush), and Cultivated common fields (med) by 2014.Areas showing negative values of -5 to -14 (yellow) represent areas that were initially Cultivated common fields Cultivated common fields (high), Low shrubland, Grassland, Woodland/open bush, Thicket/dense bush, Wetlands, Plantations/woodlots mature.Similarly, areas in brown with values of -15 to -39 were initially Urban township (low veg/grass) and Urban sports and golf (low veg/grass) and changed to Urban residential (dense trees/bush), Thicket/dense bush, Cultivated common fields (low), Grasslands, Wetlands, and Low shrubland.