Detecting land use and land cover change for a 28-year period using multi-temporal Landsat satellite images in the Jukskei River catchment, Gauteng, South Africa

The Jukskei River catchment is one of the urban catchments in the central part of Gauteng province covering a large part of City of Johannesburg Metropolitan Municipality and small part of Tshwane and Ekurhuleni Metropolitan Municipalities that have witnessed tremendous land use/land cover (LULC) change over time. Jukskei River catchment is one of the fastest growing catchments in terms of population and change in LULC over time. Therefore, it is very important to detect the nature and extent of these changes in order to identify the direction and future expansion of LULC within the catchment area. To accomplish that, multi-temporal satellite remotely sensed data acquired from Landsat-5 Thematic Mapper (TM) 1987, Landsat-5 Thematic Mapper (TM) 2001 and Landsat-8 Operational Land imager (OLI) 2015 were used to detect LULC change in Jukskei River catchment area. The Jukskei River catchment was classified into four major LULC classes including: Built-up area, bare surface, sparse vegetation and intact vegetation. The analysis of the results revealed that for the past 28 years (i.e., 1987-2015), built-up and bare surface areas have increased by 56.2% (42713.1 ha) and 8,2% (6225.1 ha) while intact and sparse vegetation have decreased by 9.8% (7455.0 ha) and 25.8% (19659.6 ha), respectively. The overall accuracies for 1987, 2001, and 2015, were 85.9%, 87.5%, and 92.5% respectively, with Kappa Index of Agreement (KIA) of 81.3%, 83.3%, and 90% which indicates the accuracy of classified images with the reference images. The results revealed by this study can be used for decision-making activities and policy development regarding land use management within the catchment.


Introduction
Land use/land cover (LULC) change is one of the most important processes related to global change (Foley et al., 2005). LULC has been acknowledged as chief drivers of environmental change at all spatial and temporal levels (Mishra et al., 2014). Humans play a major role as forces of change in the environment, influencing changes at all levels ranging from local to global (Paiboonvorachat and Oyana, 2011). Among others, there are many driving forces contributing to LULC change which could be social, economic, institutional, political or geographic in nature (Lepers et al., 2005). These direct drivers impact at several levels such as local conditions (subsistence livelihood, poverty and culture); at national levels (population growth, domestic markets, state policies, legislations and laws) and at the international level (world markets' demands, commodity prices for goods and services) (Koranteng et al., 2016). Therefore, reliable and updated information on the LULC maps and their dynamics can help to provide base information for further decision-making in catchment management activities (Alphan, 2003).
Long-term image archives of the earth resource satellites are a useful resource for LULC change detection studies. Yuan et al., (2005) developed a methodology to map and monitor LULC change using multi-temporal Landsat TM data in the seven-county Twin Cities Metropolitan Area of Minnesota for 1986Minnesota for , 1991Minnesota for , 1998Minnesota for and 2002 Their result showed that between 1986 and 2002 the amount of urban land increased from 23.7% to 32.8% of the total area, while rural cover types of agriculture, forest and wetland decreased from 69.6% to 60.5%. Amin et al., (2012) carried out a study on LULC cover mapping of Srinagar city in Kashmir Valley. They observed that the Srinagar city has experienced significant changes during 1990 to 2007. The analysis also showed that changes in land use pattern have resulted in the loss of forest area, open spaces, etc. Rawat et al., (2013) have carried out a study on LULC of five major towns (i.e., Ramnagar, Nainital, Bhimtal, Almora and Haldwani) of Kumaun Himalaya in Uttarakhand (India). Based on 20 years of satellite data from 1990 to 2010, they found that the built-up area has been increased about 8.9% and 3.9%, respectively, while area under other land categories such as vegetation, agricultural land and water body have decreased about 9.4%, 0.7% and 2.8%, respectively.
The Jukskei River catchment has experienced rapid industrial development, increase in built-up areas and population growth. Detecting the past and the present LULC conditions play a key role in decision-making when managing human activities within the catchment area. However, within the Jukskei |River catchment, there is limited number of research studies that has been made in analysing and detecting LULC change using GIS and RS to capture the patterns, trends and drivers of landscape changes over time. However, multi-temporal remote sensing data contained much more information which benefit for improving LULC classification accuracy, mainly because different vegetation types usually had different growth characteristics and presented different spectral features in different periods (Shao et al., 2001;Jia et al., 2013;Odindi et al., 2012;Li et al., 2014). Therefore, this paper is based on application of multi-temporal remotely sensed data using a GIS and RS techniques approach in mapping, detecting, analysing the annual LULC change rate for the past 28 years from understanding of drivers that contribute to LULC change within the catchment, which are required to assists with LULC practices for sustainable development.

Description of the Study Area
The Jukskei River catchment is one of the largest catchment areas in Gauteng Province with the City of Johannesburg Metropolitan Municipality covering a large portion of the catchment area followed by Tshwane Municipality and Ekurhuleni Metropolitan Municipality. The catchment is located at South Latitude 26° 08' and East Longitude 28° 14'. The elevation of the downstream catchment ranges from 1319.12 m to 1497.62 m and the upstream catchment is characterised by an elevation ranging between 1378.08 m and 1802 m. The catchment area covers an area of approximately 800 km 2 with the Jukskei River being the longest river among others ( Figure 1). The catchment is characterized by a relatively dry and sunny climate. Temperatures are usually fairly mild due to the city's altitude, with an average maximum daytime temperature of 25 ºC in the summer, dropping to around 17 ºC in winter. The study area has different land uses, including industrial, trade, educational facilities, tourism destinations, places of entertainment all dominated by vast areas of predominantly built-up areas. The Jukskei River is the largest river within the catchment area with a total length of about 68 km and is joined by numerous tributaries. Additionally, the catchment is gauged and has five gauging stations with scattered and one located at the catchment outlet: there is also one global weather station (i.e., Climate Forecast System Reanalysis) and two South African weather stations.

Satellite Data Acquisition
Landsat data provide the longest terrestrial remote sensing record and have a long history for land cover mapping because of their moderate spatial resolution and near global coverage (Roy et al., 2014;Wulder et al., 2016). With the freely available Landsat data, it is now possible to reconstruct the history of the Earth's surface back to 1972 (Pflugmacher et al., 2012). In this study, three cloudfree summer images acquired by Landsat satellites in 1987Landsat satellites in , 2001Landsat satellites in and 2015 (Table 1)

Atmospheric Correction of Landsat Images
Topographic correction is the process of reducing variation of image values resulting from differences in surface terrain illumination and shadows cast during image acquisition (Vanonckelen et al., 2013). Atmospheric correction aims at determining the true surface reflectance values by removing the atmospheric effects resulting from the scattering and absorption of electromagnetic radiation by gases and aerosols when passing through the atmosphere to the satellite sensor (Hadjimitsis et al., 2010). Many studies have acknowledged that atmospheric correction is one of the most important corrections in land cover classification, especially when working with multiple and multi-temporal scenes (Song et al., 2001;Young et al., 2017).
For the purpose of this study the atmospheric correction algorithm was performed using Dark Object Subtraction to remove the atmospheric effects resulting from the scattering and absorption of electromagnetic radiation by gases and aerosols when passing through the atmosphere to the satellite sensor (Hadjimitsis et al., 2010). The Dark Object Subtraction approach argues that there can be a substantial likelihood that at least a few pixels in an image should be black. (i.e., 0% Reflectance) (Chavez, 1988). Therefore, Semi-Automatic Classification plugin in Quantum Geographical Information System (QGIS) was used for pre-processing of Landsat images (i.e., Landsat-5 TM, and where: Lλ is the corrected DN value, Lsλ is the original DN value and Lp is the DN value of the dark pixels (Chavez, 1988;Song et al., 2001).

Image Classification and Change Detection Techniques
In this study image classification was performed by capturing different pixels and assigning them to certain classes in accordance with the information categories provided by the users (Perumal & Bhaskaran, 2010). Anderson et al., (1976) point out that there is no ideal classification for LULC, the process is subjective, depending on the different perspectives in the classification process. For this study, a supervised maximum likelihood classification method was applied. The maximum likelihood method is a common method in remote sensing owing to its robustness and was therefore implemented to classify the images (Strahler, 1980;Ediriwickrema & Khorram, 1997;Zheng et al., 2005;Mawasha & Britz, 2020).
Prior to image classification, training areas were captured using a false colour composite for each of the satellite images of 1987, 2001 and 2015 by selecting a minimum of 10 pixels per polygon. In this study, four classes of LULC (i.e., built-up area, degraded vegetation, bare surface, and intact vegetation) were selected with the use to google earth image (see Table 2). To determine the changes in LULC in different years, a post-classification change detection algorithm was used in this study.
Post classification comparison is a common method used for change detection (Alphan et al., 2009;Mas, 1999;Chen et al., 2003). This method detects changes in land cover type by comparing pixel by pixel for the classified images. The use of the post classification comparison technique resulted in a cross-tabulation matrix (i.e., LULC change transition matrix) which was computed using the CROSSTAB module in the TerrSet Geospatial Monitoring and Modelling software. A transition matrix is a fundamental starting point for the analysis of land change (Pontius et al., 2004). Gross gains and losses were calculated for one period covering 28 years (i.e., 1987-2015) to detect changes and LULC transformation among LULC classes and to identify. 2. Bare surface This group covers regions not protected by trees, fields with exposed soils, degraded soils and mined areas.

Intact vegetation
This category applies to areas covered by dense vegetation, cropland, agricultural land and natural landscaping.

Sparse vegetation
It consists of areas of scattered vegetation, areas of shrubs covering and small wooded trees mixed with grasses. This group applies to areas with very little vegetation coverage.
As shown in Table 3, the computed LULC change transition matrix consisted of rows (displaying LULC class category for time 1, T1) and columns (displaying LULC class category for time 2, T2).
The notation A12 denotes the proportion of the landscapes which experiences a transition from LULC1 to LULC2. The diagonal entries indicate the total amount of persistence, which dominates most landscapes, including those where authors claim that the change is important and/or large (Wear & Bolstad, 1998;Mertens and Lambin, 2000;Geoghegan et al., 2001;Chen et al., 2002). In the Total column (i.e., total T1), the notation A1+ denotes the proportion of the landscape in LULC1 in time 1, which is the sum over all 1. In the Total row, the notation A+ 1 denotes the proportion of the landscape in LULC1 in time 2, which is the sum over all 1. The additional column on the right indicates the The percentage and rate of LULC change were computed by the following formula (Kindu et al., 2013;Demissie et al., 2017): where: A and B are the final and initial area coverage of LULC types during that specific period, and T represents year difference between the initial and final period. A LULC classified image for 1987 was used as the initial image and the 2015 image as the final image and a change statistics table was created for analysis.

Accuracy Assessment
In remote sensing, accuracy assessment is mandatory and is important for providing information about the quality of the produced classification (Okeke & Karnieli, 2006;Rwanga & Ndambuki, 2017). For the purpose of this study, a confusion error matrix was used for the assessment. The confusion matrix shows errors resulting from missing and mixing of information, various accuracy measurements, such as the overall accuracy, producer's accuracy, user's accuracy and Kappa Index Agreement (KIA) (Zhou & Xiong, 2012). In order to assess the accuracy of the classification method, a stratified random sampling method was applied to select 50 points per class i.e., 200 points for each classified image. With the aid of the aerial photo mosaic (1:25 000), invariant feature points and topographic sheets, the accuracy of the images was assessed using the error matrix technique described by Congalton & Plourde (2002). The user's accuracy, producer's accuracy, overall accuracy as well as Kappa coefficient were calculated as follows: User accuracy is the likelihood that a pixel relates to a certain group of LULCs as defined by Petropoulos, Kalaitzidis & Vadrevu (2012); the algorithm properly identifies the pixels within the same class. It displays the percentage of likelihood the that class on the ground is actually defined by the class in which a pixel is identified into an image (Story & Congalton, 1986). The sum of the row is calculated by dividing each diagonal part into one confusing matrix: User Accuracy = Total number of correctly correct pixels in a category Total reference pixels (i.e., column total) × 100 [4] Although the accuracy of the producer is an indicator of the accuracy of a specific classification system, it indicates the correctly categorized percentage of a particular ground class. The diagonal elements in the table are divided by the sum of each column by dividing: Producer Accuracy =

Total number of correct pixels in a category
Total reference pixels (i.e., row total)

× 100 [5]
Overall accuracy is the number of pixels that have been accurately identified from the validity data collection over the overall number of pixels used as a percentage for the accurate measurement (Petropoulos, Kalaitzidis & Vadrevu, 2012).
Overall Accuracy = The Kappa Index of Agreement (KIA) is the most widely used metric to determine the degree of agreement between the comparison and validation datasets, as stated by Foody (2010). It indicates the degree of reliability of image classification. The KIA statistical range is between zero and one.
The KIA is determined according to Congalton (1991) where: (i.e., < 40%) signifies poor agreement (Congalton and Green, 2009), and this ranges were employed in this study to assess the accuracy.

Accuracy assessment of the classified images
The classification accuracy for the LULC maps derived from satellite data was assessed using three different time periods by means of a confusion error matrix. Assessment of the classification results proved that the supervised maximum likelihood classification produced valid and reliable LULC maps since all the overall accuracies and the KIA were higher than 80% (Congalton and Green, 2009). The overall accuracies for 1987, 2001, and 2015, were 85.9%, 87.5%, and 92.5% respectively, with KIA of 81.3%, 83.3%, and 90% which indicates the accuracy of classified images with the reference images. Producer's and user's accuracies were also high, ranging from 70% to 100%. The Landsat-5 TM data resulted in the lowest overall accuracy (85.9%) and KIA (81.3%) of the three datasets. The low spatial resolution combined with the spectral similarity of built-up and bare surface areas made it difficult to separate these two classes on the Landsat-5 TM imagery. Furthermore, a decrease of image spatial resolution leads to spectral mixing of different categories which produce spectral confusion between cover materials (Yang and Lo, 2002). As stated by Anderson et al., (1976) a reliable land cover classification needs the minimum overall accuracy value computed from an error matrix to be 85%. Therefore, the overall accuracies for the study periods are in agreement with findings of other studies. Hence, the results revealed by this study can be used by land use policy makers and catchment managers for decision-making activities and policy development regarding land use management within the catchment.

Spatio-temporal changes over time (1987-2001, 2001-2015 and 1987-2015)
In this study, LULC change detection of each LULC for the last three decades was examined between years (i.e. , 1987-2001, 2001-2015 and 1987-2015) and the relative changes were obtained in both hectares and percentage see Table 4 and Figure 3. The analysis of the results reveals that there was increase in areal coverage of built-up areas during 1987-2001 (Table 4)  These radical changes in natural land cover are mostly attributable to human activities and urbanization and may have resulted in an increase in demand for arable land for administrative or settlement development, as is common practice in the study area (see Figure 2 & 3).

LULC Transformation for the past 28-years (1987-2015)
The classified images of 1987 and 2015 were further analysed using a cross-tabulation (CROSSTAB) module of TerrSet software to detect LULC transformation that occurred within the Jukskei River catchment area. Table 6 Table   6). The analysis of the classified images over time revealed that, built-up areas are generally expected to increase in aerial coverage compared to other LULC classes due to residential development or demand within the Jukskei River catchment.

Conclusion
This study demonstrated the efficiency of remote sensing Landsat data and GIS tools to develop a model that was used to detect LULC changes for a period of 28 years (i.e., 1987-2015). The analysis of the study revealed that there has been a change in LULC classes within the catchment area.
Compared to other LULC classes such as bare surface, sparse vegetation and dense vegetation, the built-up area class increased in spatial extent throughout the period. This was due to industrial and residential development that is continuously increasing due to increase in population density and industrial development where impervious surfaces are replaced by pervious surfaces. The trend however, shows that sparse vegetation and intact vegetation has no potential to recover to its original state, as revealed by high conversion of sparse and intact vegetation to built-up area for the last 28years. Additionally, the analysis of the results revealed that more than half of the total catchment area is covered by built-up area in 2015 compared to 37.7% in 1987. This increases environmental risk such as flash floods due to high surface runoff due to LULC change. The multi-temporal classified images provides details of the spatial distribution of LULC change that can be used for decisionmaking for developing policies for land use practices within the catchment area.