Quantifying urban land cover imperviousness as input for flood simulation using machine learning: South African case study Open Access

Hydraulic and hydrological models are often used to simulate, predict and forecast urban flooding (Guo et al. 2021; Mignot & Dewals 2022). These models are also used to design urban drainage systems for design events with selected durations and return periods (El Hattab et al. 2020; Ambrosio et al. 2022; Ferrans et al. 2022).

Urbanisation is often linked with more frequent and severe flooding (Jacobson 2011; Braud et al. 2013; Sakijege 2013; Ferreira et al. 2016; Saurav et al. 2021). It is reported that since 2005, more than half of the world's population lives in urban areas, with the proportion of urban dwellers set to rise to about 60% by 2030 (UNDP 2019). The developing continents of Africa and Asia are experiencing the fastest urbanisation rates (Cohen 2006; UNDP 2014; Zhang et al. 2015). Urban areas contain many different types of buildings and roads, resulting in large impervious areas. Total impervious area (TIA in this study) affects runoff generation and conveyance (Nguyen et al. 2022). Therefore, many widely used models, such as the Hydrologic Engineering Center's River Analysis System (HEC-RAS) and the United States Environmental Protection Agency's Storm Water Management Model (SWMM), require TIA as an input parameter when considering urban areas.

However, it is time-consuming for practitioners to quantify the TIA of large catchment areas. Time and resource constraints are prevalent worldwide, but especially problematic in developing countries like South Africa. Therefore, practitioners tend to estimate TIA visually using satellite photos or 1:50,000 drawn topographical maps (Topo maps) derived from aerial photos. However, this ad hoc method can lead to inconsistent results. For example, a study conducted in South Africa by Loots et al. (2023) used Topo maps to define the total extent of urban development and found varying runoff responses from catchments with similar urban development levels. The study concluded that the density of impervious areas in different development types could contribute to the inconsistent runoff responses. In many developing countries, the imperviousness and layout characteristics of formally planned, structured and developed suburbs (formal development) differ significantly from those in informal areas, where settlement occurs in an unplanned and disorganised manner (Gogate & Rawal 2015).

Most South African cities have experienced significant population increases in recent years (Mlambo 2018; Mubangizi 2021; StatsSA 2022). People migrate to urban areas hoping for improved livelihoods (Cohen 2006; Geyer et al. 2012; Capps et al. 2016; UNDP 2019). Geyer et al. (2012) found continued growth of informal townships on city outskirts. They also noted a trend in the post-Apartheid era where people migrated away from urban centres towards urban fringes. This combination has led to significant development of both formal and informal settlements on city outskirts. Thompson (2019) reported a 30% increase in the South African urban footprint between 1990 and 2018. In 2022, an estimated 88.8% of the population lived in formal settlements, 8.1% in informal dwellings and 3.1% in traditional dwellings. Gauteng is the most urbanised of South Africa's provinces, where 11% of households were informal dwellings in 2022 (StatsSA 2022). This rapid urban development makes it imperative to quantify the TIA of different development types in South Africa to improve the hydrological modelling of urban areas.

In order to simplify the estimation of impervious areas for hydraulic and hydrological modelling, the main aim of this study was to develop a new framework to quantify the imperviousness of urban land cover classes from satellite images using the Red, Green and Blue (RGB) bands. The first study objective was to develop a methodology that could be used to derive TIA in any region with a land cover dataset that differentiates between a range of urban land cover classes and with high-resolution satellite imagery containing RGB bands. The second objective was to use this methodology to derive representative estimates of TIA for different South African National Land Cover (SANLC) classes. The third objective was to apply the methodology to quantify temporal changes to TIA in selected study catchments.

This approach is described in the following paragraphs and then applied to case study areas in the city of Pretoria, South Africa. The methodology is first used to derive representative estimates of TIA for different SANLC classes (as set out in Objective 2). The machine learning land cover classification tool is then used to quantify temporal changes to TIA in selected study catchments (as set out in Objective 3). The following sections describe the methodology followed in this study in further detail.

It was previously established that urban development influences hydrological processes in South African catchments (Loots et al. 2023). It is envisaged that the results from the current study could be used in future studies to further analyse the developmental influences in the context of South African development types and corresponding levels of imperviousness. Therefore, urban catchment areas with available flow data were chosen to quantify historical changes in imperviousness. Five catchments with at least 14% developed urban area as of 2020 were selected (Figure 2). As shown in Figure 2, the catchments include both formal and informal urban areas. The catchment sizes range from 30 (A2H063) to 357 km² (A2H027).

Figure 2 shows that Catchment A2H063 is the furthest west of the five catchments. It is also the closest to the historical urban centre of Pretoria. Moving progressively east are A2H054, A2H055, A2H027 and A2H029. While the three westerly catchments are almost entirely covered by formal development, the two most easterly catchments are less developed overall and contain informal settlements.

The EnMAP-Box machine learning plugin available in the QGIS (ENMAP-Box_Developers 2019) was applied to classify the land cover images in this study. The impervious areas of the high-resolution (1.5 m resolution pixels) Spot 6-7 South African National Space Agency (SANSA) satellite images with RGB bands, dated 2017, were used for this analysis. Land cover sample points were manually selected from visual inspection of the satellite images to achieve an acceptable distribution among the six initially classified general land cover types, being: (1) buildings, (2) trees, (3) grass, (4) bare, (5) roads, and (6) water. These six land cover types were chosen with the primary aim of distinguishing impervious areas from pervious areas. Training and independent validation points were selected to cover different chromatic shades representing these classes in the study area and processed using the EnMAP-Box machine learning tool. Between 50 and 200 points for each general land cover type were selected from each of the four SANSA satellite images used in the study. The number of points depended on the availability of land cover types in the available images. To simplify output, land cover types were grouped into three categories: impervious area (IMP) from buildings and roads, bare area (BARE) from bare and water and vegetated area (VEG) from trees and grass. Figure 1(b) shows an example of the IMP, BARE and VEG categories identified for a subset of the data from the image shown in Figure 1(a) using the machine learning tool.

Since the primary purpose in this case was to quantify impervious area, water was not retained in a separate category for this study. In the SANLC dataset, water is classified in its own class. Hence, the water bodies incorporated in the urban built-up land cover classes investigated in this study are negligible. When this method is applied in regions with many water bodies in the urban environment, it is recommended that the water land cover type be kept in its own classification and quantified separately from the other classification groups.

The grouped classification derived using the machine learning tool was overlaid with the sample zones from the SANLC data in order to calculate the IMP, BARE and VEG area of each sample zone (Figure 1(e)), using the Zonal Histogram tool in QGIS. The median, 25th and 75th percentile of impervious area and maximum and minimum percentages of each of the three grouped classifications were calculated for each SANLC class included in this study.

Verification zones were selected from each of the 10 SANLC urban built-up land use classes. Verification zones were selected randomly from areas in Pretoria, but in suburbs different from those where the sample zones were selected for imperviousness quantification. Verification zones were also selected considering only the land cover class. This was done to avoid bias in the results. The locations of the verification zones are shown in Figure 4.

The same methodology, as for sample zones, was applied for the imperviousness quantification of verification zones, as summarised in Figure 1(a)–1(e). The median values for the impervious area of the verification zones of each SANLC class were compared with the impervious area median values in the sample zones, and the differences were noted. The total verification areas for the different urban built-up land cover classes are summarised in Table 1.

Further verification was achieved through field inspections. Research team members identified and visited independent sample sites for all SANLC classes under investigation. TIA was physically measured at these sites. Sample sites from different neighbourhoods with similar SANLC classes were included in the sampling. Field verification sample sites were chosen independently from remote sensing sample sites or verification sites. The locations of the field verification sites are shown in Figure 4.

The total area with impervious land cover for each sample site was physically measured on-site on properties where access was granted. All paved and roofed areas were included in the TIA calculation. Bare areas were not included in the TIA measurement or calculation. Schematic maps were drawn of the impervious areas at sample sites, and the length and width measurements were taken using a 30-m measuring tape. On properties where access was denied, notes regarding the location of vegetation covering impervious areas were taken from the street, and the TIA was manually delineated using high-resolution (1.8 m) Google Earth images. The total sampled area for the field verification study was 0.45 km² from 292 individual sample sites (or stands). A number of the sampled sites cover multiple SANLC classes. These were divided into parcels based on SANLC class coverage. In total, 477 sample parcels were included, with distribution per SANLC class, as shown in Table 1.

The TIA was quantified for each parcel and reported as a percentage of the total parcel area. The TIA percentages were grouped per SANLC class, and the median, 25th percentile and 75th percentile values were calculated. The results were compared with the remote sensing imperviousness median values.

In order to quantify changes in imperviousness of the study catchments, historical and recent estimates of the TIA were extracted using the EnMap-Box machine learning plugin in QGIS. Prior to 2002, the only available images in the study areas are greyscale aerial photos taken by the South African National Geo-spatial Information (NGI). Photos of the study area were taken in 1984, 1991 and 2001. They were digitised to a resolution of 2.5 m by 2.5 m pixels from the analogue data. Impervious built-up areas were identified from the digitised images through a combination of remote sensing and manual correction of misclassified areas.

For this analysis, the EnMAP-Box machine learning plugin was applied in QGIS to identify TIA from the historical images. The TIA was extracted for each study catchment from the classified data, for multiple years. The TIA results are described as a percentage of each catchment's area (TIA_%).

A comparison was undertaken between the 2001 NGI images and the 2002 Google Earth dataset. An average difference of 2.1% was found between the TIA measured using the two datasets for all study areas. The discrepancy was deemed sufficiently small to combine the results from the different datasets for this study.

This section summarises results obtained in the analysis and verification of TIA in the 10 studied SANLC classes. The derivation of generalised TIA percentages (TIA_%) for different urban SANLC classes is also described.

It is evident from Figure 6(a) that areas classified as Residential Formal (Trees) have higher percentages of vegetation than other classes. Vegetation percentages decrease progressively from Residential Formal (Tree) to Residential Formal (Bush) to Residential Formal (Grass) to Residential Formal (Bare). The same trend is evident for informal residential classes. This result was expected and confirmed that the methodology works as expected. The impervious area median for Residential Informal (Bare) was lower than expected when compared with other informal residential classes. Closer inspection of the satellite images showed that informal areas classified as Residential Informal (Bare) had high proportions of bare surfaces between dwellings, with small, impervious dwellings. As expected, commercial and industrial areas have high TIA_%.

The TIA_% results from the verification zones (Figure 6(b)) compare well with the TIA_% as calculated in the sample zones, with differences between the median IMP values of the sample zones and verification zones of between 0.9 and 3.0% for the four Residential Formal areas (denoted by the red IMP plots in Figure 6), 1.1% for Commercial areas and 1.5% for Industrial areas. The sampled area for Residential Informal (Tree) was the smallest of all 10 areas due to limited land cover with this classification in the study area. The area that was available for verification of TIA in Residential Informal (Tree) was significantly smaller than the sampled area. Less than 1 ha of this land cover class is located in the rest of Pretoria. Therefore, verification was not done for this land cover class. The differences between the median values of the sample zones and verification zones of the verification of impervious areas in the other three Residential Informal SANLC classes were between 1.5 and 3.0%.

It should be noted that seasonal variations between VEG and BARE areas are possible. However, the focus of this research was on quantifying IMP areas. Where possible, images from winter months, less vegetation covers impervious areas, were used to get as good a representation of impervious area as possible. If this methodology is applied in other studies where water bodies are also identified, seasonal variation of water body extent should also be considered.

The methodology followed in this study did not specify a direct comparison between remotely sensed and field-verified sample sites. However, based on the results of this example, it is recommended that future studies should consider direct comparison. It is postulated that the discrepancies will average out in larger datasets, but this needs to be confirmed in a future study.

As shown in Figure 8, the median values of all verification results in residential areas are within 10% of the median of measured results. Therefore, the median results from the field verification compared reasonably well with the TIA_% quantified using the machine learning methodology, with field verification values showing wider ranges of values between the 25th and 75th percentiles of TIA. The broader range of values in the field verification dataset is attributed to the fact that the field verification sites were chosen based on accessibility. This meant that, in contrast to the large homogeneous areas chosen as sample sites for the remote sensing procedure, one site generally contained multiple land cover classes, with some land cover class areas only being a few square metres (an example of this is shown in Figure 7). This heterogeneity, combined with the fact that the land cover layer's pixel resolution is 20 m, resulted in more variability of TIA per land cover class.

The results were, therefore, reasonable, and median values were chosen as recommended estimates of TIA_%, with the 25th and 75th percentile values of the median (M) suggested as lower (L) and upper (U) confidence intervals. The recommended TIA_% values for urban SANLC classes are summarised in Table 2. It is emphasised that the results from this study are limited to one geographic area due to the limited availability of observed data in urban areas in South Africa. However, it is postulated that the results apply other regions, given the relationships developed with national land cover classes in this study. Thus, until additional data are available and further research is undertaken to improve the results, the TIA_% values in Table 2 are proposed for the hydrological modelling of South African urban areas.

The methodology applied in this study can also be applied in any area with high-resolution (1.5 m resolution or better) satellite imagery with visible RGB bands and a land cover dataset with multiple classes of urban built-up areas. Site-specific values are recommended for hydraulic and hydrological modelling where possible.

These results were expected, based on previous reports that urbanisation is still increasing in developing countries (Cohen 2006; Geyer et al. 2012; Capps et al. 2016; UNDP 2019), with the continued growth of informal townships on city outskirts (Geyer et al. 2012) and a trend in the post-Apartheid era (post-1994) where people migrated away from urban centres, towards urban fringes (Geyer et al. 2012).

All five study catchments have observed flow data dating back to 1985 or earlier. Therefore, runoff from these catchments should be compared with impervious areas, to determine if trends exist and quantify the effect of urban built-up areas in developing countries on runoff.

Previous studies have found that imperviousness significantly influences runoff characteristics. However, determining imperviousness at a catchment scale can be time-consuming and expensive. In order to simplify the estimation of impervious areas for hydraulic and hydrological modelling, this study aimed to develop a methodology that would provide consistent estimates of the imperviousness of built-up urban land cover classes by applying a machine learning tool.

The first objective was achieved by developing the methodology for this study. The methodology applied a combination of remote sensing analysis using the EnMAP-Box QGIS plugin machine learning tool and field verification. This methodology can be applied to derive imperviousness in any region with a land cover dataset that differentiates between a range of urban land cover classes, and with high-resolution satellite images with RGB bands.

The second objective was to use this methodology to derive representative estimates of TIA for different SANLC classes. Major differences were found in median total impervious area percentages (⁠_TIA%) between the different urban built-up land cover classes studied, with urban residential (⁠_TIA% = 38%) having lower median imperviousness than commercial (⁠_TIA% = 81%) and industrial (⁠_TIA% = 89%) classes. Furthermore, _TIA% varies between 17 and 61% for different formal residential classes and between 14 and 43% for different informal residential classes. TIA_% of formal and informal urban land cover types were generalised based on a South African case study, to derive typical imperviousness values for different SANLC classes. These values are recommended as default values for hydraulic and hydrological modelling of South African formal and informal urban areas, or for use in similar land cover classes in other developing countries, where site-specific information is not readily available. It is recommended that direct comparison of remotely sensed and field-verified data be expanded in future studies.

The third objective was to apply the methodology to quantify temporal changes to TIA in selected study catchments. The results showed that the most western catchment had the highest level of TIA over the study period, with catchments further east having progressively less TIA. This result was expected, with the western catchment being closest to the city centre. As expected from the literature, the results also showed increasing TIA in all the study catchments over time. All five study catchments have observed flow data dating back to 1985 or earlier. Therefore, runoff from these catchments should be compared with impervious areas to determine trends and quantify the effect of urban built-up areas in developing countries on runoff.

The contributions of G. Mogale, L. Moleka and T. Mposhomali, who assisted with data collection in the field and processing of remotely sensed imperviousness data, are gratefully acknowledged. The contributions of the anonymous reviewers are gratefully acknowledged.

The research presented in this manuscript emanated from a study funded by the Water Research Commission (WRC) (Project K5/2747), whose support is acknowledged with gratitude.

I.L. contributed formulating the research question, designing the methodology, collecting and processing data, analysing data and preparing the manuscript. J.S. contributed to formulating the research question, designing the methodology and preparing the manuscript. T.K. contributed to the design of the methodology and preparing the manuscript.

The SANLC data used for this study is available from https://egis.environment.gov.za/sa_national_land_cover_datasets. The SANSA satellite images can be requested from https://www.sansa.org.za/research/.

The authors declare there is no conflict.