Multi-Tier Land Use and Land Cover Mapping Framework and Its Application in Urbanization Analysis in Three African Countries

Abstract

The population of Africa is expected to rise to 2.5 billion by 2050, with more than 80% of this increase concentrated in cities. Africa’s anticipated population growth has serious implications for urban resource utilization and management, necessitating multi-level monitoring efforts that can inform planning and decision-making. Commonly, broad extent (e.g., country level) urban change analyses only examine a homogenous “developed” or “built-up” area, which may not capture patterns influenced by the heterogeneity of landscape features within urban areas. Contrarily, studies examining landscape heterogeneity at a finer resolution are typically limited in spatial extent (e.g., single city level). The goal of this study was to develop and test a hierarchical integrated mapping framework using globally available Earth Observation data (e.g., Landsat, Sentinel-2, Sentinel-1, and nightlight imagery) and accessible methodologies to produce national-level land use (LU) and urban-level land cover (LC) map products which may support a range of global and local monitoring and planning initiatives. We test our multi-tier methodology across three rapidly urbanizing African countries for the 2016–2020 period: Ethiopia, Nigeria, and South Africa. The initial output of our methodology includes annual national land use maps (Tier 1) for the purpose of delineating the dynamic boundaries of individual urban areas and monitoring national LU change. To complement Tier 1 LU maps, we detailed urban heterogeneity through LC classifications within urban areas (Tier 2) delineated using Tier 1 LU maps. Based on country-optimized sets of selected features that leverage spatial/texture and temporal dimensions of available data, we obtained an overall map accuracy of between 65 and 80% for Tier 1 maps and between 60 and 80% for Tier 2 maps, dependent on the evaluation country, although with consistent performance across study years providing a solid foundation for monitoring changes. We demonstrate the potential applications for our products through various analyses, including urbanization-driven LU change, and examine LC urban patterns across the three African study countries. While our findings allude to general differences in urban patterns across national scales, further analyses are needed to better understand the complex drivers behind urban LC configurations and their change patterns across different countries, city sizes, and rates of urbanization. Our multi-tier mapping framework is a viable strategy for producing harmonious, multi-level LULC products in developing countries using publicly available data and methodologies, which can serve as a basis for a wide range of informative and insightful monitoring analyses.

1. Introduction

The global urban population is growing considerably, mainly as a result of rural-to-urban migration and natural increase, in which urban births outnumber urban deaths [1,2]. In 2022, the global urban population was increasing at an annual rate of 1.6% [3], and projections estimate that 68% of the global population will live in urban areas by 2050 [1]. The increasing proportion of the population living in urban areas (i.e., urbanization) is transforming physical, social, and economic landscapes, particularly in developing nations where most anticipated urban growth will be concentrated. While urbanization can be a catalyst for beneficial development [4], rapid urbanization presents numerous management challenges and, when unaddressed, often leads to adverse social, ecological, and economic effects. Rapid urbanization and its concomitant impacts have been particularly noted in Africa, where the urban population is growing at an annual rate of nearly 4% [3] and urbanizing areas face significant challenges, including affordable housing shortages and ensuing development of informal settlements [5,6], inadequate access to basic services [7], environmental and ecological deterioration [8,9], and loss of important provisioning land use [10,11]. Dynamic African socio-ecological systems, particularly those related to current urbanization, are complex and diverse [12] and will require deliberate investments and catered solutions to support the prosperity of a transforming Africa.

Planning and monitoring initiatives, for instance the 2030 Agenda for Sustainable Development, have been ushered in on the global stage to address rapid urbanization effects and promote sustainable growth, particularly in developing regions such as Africa. Geospatial analyses of remotely sensed products have become critical tools to carry out urban monitoring efforts (e.g., the United Nations Sustainable Development Goal Indicators), which often require consistent spatial data at multiple scales, spatial resolutions, and recurring timesteps. Nevertheless, associated urban monitoring may fail to capture key patterns, trends, and characteristics of urbanization-driven land use (LU) and land cover (LC) change, primarily due to the spatial and thematic characteristics of underlying map data [13,14]. Broad extent (e.g., global or national) LULC mapping efforts often classify all human settlements and related areas as a single “developed” class, sometimes referred to as the “red blob” given the red color scheme commonly used to symbolize developed classes. The “red blob” typically encompasses all human development LULC features into a single class and presents the urban landscape as homogenous; in reality, the LULC comprising the urban landscape is often highly heterogeneous. When important urban LULC features are not adequately segregated and identified, the suite of related spatial analyses that can be performed is limited and the findings may be misrepresentative. In the case of Africa, LU products exist that are suitable for blanket estimations of urban LU areas and change [15,16,17,18], but a deeper understanding of urbanization effects requires products that characterize the heterogeneity of different LC types across the landscape [19,20].

High-resolution imagery supports the characterization of urban LC heterogeneity and, through spatial analysis, may improve our understanding of urban LC composition, configuration, and form, as well as the dynamics of urban LC patterns. Very high-resolution imagery (e.g., <5 m), in particular, can depict detailed physical features, structures, and land cover classes comprising urban areas. However, very high-resolution images are typically difficult to access and acquire, especially across larger temporal and spatial extents, thereby hindering their use in monitoring efforts across data-limited regions.

Alternatively, multi-resolution product suites based on widely available public remote sensing imagery sources may be able to provide information for comprehensive investigations of urbanization effects. For example, the Global Human Settlement Layer is a framework processing and mapping various levels of human settlement data from 10 m to 1 km in resolution from 1975 to 2030 [21]. This product suite has proven valuable for the consistent global monitoring and reporting of Sustainable Development Goals and Indicators, such as SDG 11.3.1 [22], and offers unique opportunities for multi-scale analyses [23]. Unfortunately, the majority of GHSL data covering larger temporal and spatial extents are only available at coarser spatial resolutions (e.g., ≥ 250 m). Multi-resolution LULC product suites of higher spatial resolutions and extending multiple years are limited in developing regions (e.g., Africa and Asia), although some imagery sources, such as Landsat and Sentinel-2, are consistently collected and globally available, permitting the development of such multi-resolution product suites at 30 m and 10 m spatial resolution across these developing regions.

We propose a multi-tiered mapping framework characterizing national-level LU composition complemented by urban-focused LC classifications to allow for multi-scale (national to city) and multi-resolution (coarse to fine) examinations of urbanization trends, patterns, and changes. Characterizing and monitoring urbanization effects at different scales can reveal novel information useful to global/national monitoring initiatives and regional management planning that may not be illustrated at a single scale or resolution. National LU maps and city-wide LC maps can help satisfy the needs of both large-scale monitoring efforts as well as supporting enhanced local management planning. The fusion of city-wide LC maps with local knowledge can drive a more nuanced assessment of urban landscapes that include the distribution of ecological resources (e.g., waterways, open spaces, and trees) and social services (e.g., water provisioning, wastewater treatment, and transportation networks), or as standalone maps which can be used to assess the composition and quality of significant land cover features, such as green spaces or impervious surfaces across the urban landscape.

Our overarching goal was to develop a consistent, multi-tier integrated LU and urban LC mapping framework utilizing publicly available imagery and accessible methodologies for the purpose of supporting multifaceted and nuanced examinations of urbanization. Our mapping framework seeks to characterize LU across national extents and dissect urban areas into additional, and more informative, LC classes. We use consistent and openly available remote sensing data sources, such as Landsat (30 m) and Sentinel-2 (10 m) imagery, and adaptable methodologies, to produce a multi-tier methodology with applicability across Africa and other changing regions of the Global South. We developed and tested our multi-tier LULC mapping framework for three African countries (Ethiopia, Nigeria, and South Africa) representing a range of ecosystem and societal dynamics, across a 5-year study period (2016–2020) to evaluate urbanization-related changes across spatial scales. We showcase the applicability of our multi-tier product suite by conducting LU change analyses and examining LC urban patterns across the three African study countries.

2. Methods

2.1. Study Area

Africa may be considered a “frontier” for urbanization, as an unprecedented 665 million people are projected to be added to Africa’s urban population from 2018 to 2050. Additionally, economic growth in Africa is outpacing the world with a gross domestic product (GDP) averaging 4% in 2023 [24,25]. We chose to develop our multi-tiered LULC framework across three African countries that represent a range of climatic, geographic, economic, and demographic properties to test the versatility of our methodology across a diversity of environments and urbanization patterns. Our focal countries included Ethiopia, Nigeria, and South Africa, which are some of the most populated and economically vibrant countries in the African continent (

Table 1. Selected study countries and their respective land area, population, and GDP in 2023 [3].

Country	Area (km2)	Population (in Millions); (% Growth Rate/Year)	Real Gross Domestic Product (GDP—Annual % Growth)
Ethiopia	1.1 million	123.4 (2.4–2.5)	6.2
Nigeria	0.9 million	218.5 (2.4–2.5)	3.1
South Africa	1.2 million	59.9 (0.8)	1.8

).

Ethiopia, Nigeria, and South Africa are located in three different geographical regions of Africa, each with novel climates that may impact ecological gradients and, therefore, LULC mapping efforts. Per the Köppen–Geiger climate model [26], Ethiopia contains arid desert and steppe in the eastern lowlands and a mix of tropical savanna in the west with some temperate climates at higher elevations in the center of the country. Nigeria is in the African west, and mostly falls in the tropical (savannah) climate zone, with a tropical (monsoon) climate in the south and an arid climate in the north. South Africa covers different subcategories of arid climate in the north and west, and temperate climates in the east (Figure 1).

2.2. General Overview

The technical approach we present represents a hierarchical framework of Earth Observation data integrations using data sources and methodologies that are widely available and applicable across countries and continents. Our aim was to develop a flexible remote sensing approach that builds on the current paradigm shift from two-date change detection to continuous LULC monitoring [27] and does so across spatial resolutions and scales. We fused complementary remote sensing datasets (i.e., optical, radar, and night-time lights) that have proven effective in mapping different aspects of LULC, exploiting their unique contributions for identifying different LU and urban LC features. We leveraged the landscape patterns identified at continuous national extents and 30 m resolutions to delineate urban boundaries and assess urbanization-driven LU changes across our 2016–2020 study period. Within the urban areas identified across the extent of our study countries, we calculated density and configuration-related landscape metrics using the 10 m resolution LC products to identify areas of management interest and quantify changes relevant to monitoring efforts.

2.3. Remote Sensing Data

The data sources and derived metrics leveraged within our multi-tiered models are summarized in

Table 2. Summary of datasets used in Tier 1 and Tier 2 land use and land cover products, and their derived features used in modeling efforts.

	Imagery Details
Sensor Type/Dataset	Tier 1 Land Use Products	Tier 2 Land Cover Products	Derived Features
Optical	Landsat Collection-2 Surface Reflectance @ 30 m, six bands of Blue, Green, Red, NIR, SWIR1, and SWIR2 Zonal statistics calculated for green and NIR bands plus TCB, TCG, and TCW indices BAI/BAEI/NBAI not used in LU product	Sentinel-2 Top of Atmosphere * @ 10 m, six bands of B2(blue), B3(green), B4(red), B8(NIR), B11(SWIR1), B12(SWIR2) Zonal statistics calculated for green and NIR bands plus TCB, TCG, and TCW indices GLCM metrics calculated using NIR band over radii of 5 and 9	NDVI Tasseled cap indices (TCA/TCB/TCG/TCW) UCI, BAI, BAEI, NBAI MNDWI, WI %Water Zonal (optical/radar source): Min/Mean/Max/StDev of the pixel values within a 3 × 3 and 5 × 5 neighborhood Context (optical/radar source): GLCM metrics of ASM, Contrast, Correlation, Variance, Sum average, Entropy, Information Measures of correlation (1 and 2), Dissimilarity, Cluster shade, and Cluster prominence
Synthetic Aperture Radar	Sentinel-1 SAR Ground Range Detected VV polarization ** @ 30 m GLCM metrics calculated over radii of 5, 9, 13, and 17 pixels	Sentinel-1 SAR Ground Range Detected VV polarization @ 10 m GLCM metrics calculated over radii of 5 and 9 pixels
Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) @ 460 m	Included	Included	Yearly median of average monthly radiance values (months with at least two observations are counted)
TerraClimate @1/24 degree (~4.5 km)	Not included	Included	Total year precipitation and yearly minimum/maximum temperature
WorldClim V2.1 @ 30 arcseconds (~1 km)	Included	Included	19 bioclimatic variables (bio_01 to bio_19) featuring normal (30 years) temperature and precipitation statistics, as defined in Appendix A
SRTM digital elevation data @ 30 m	Included	Included	Terrain parameter (elevation, slope, aspect)—static parameter
Continuous Heat-Insolation Load Index (CHILI_Index) @ 90 m	Included	Included	CHILI index (a number between 0 and 255)
iSDA soil texture class @ 30 m	Included	Included	USDA Texture Class at 0–20 cm depth (a number from 0 to 12)
World Ecoregions (RESOLVE), vector dataset	Included	Included	Ecoregion identifier (a 3-digit number)

* We used Google Earth Engine repository for data access, and Sentinel-2 surface reflectance products were not available there prior to 2017 for our study region. Therefore, we opted to use Sentinel-2 TOA product. TOA products are not atmospherically corrected but processed for radiometric and geometric quality. ** Other polarizations were not necessarily available over all times and regions of our study.

, with the data processing flowchart depicted in Figure 2. The Landsat Collection 2 Surface Reflectance product (30 m) and Sentinel 2 Level-1C Top of Atmosphere product (10 m) were the primary optical data sources for Tier 1 and Tier 2 classification models, respectively. For the Landsat data, we removed medium and high-confidence clouds, cirrus clouds, and cloud shadows using the pixel quality assurance band. We also removed saturated observations based on the radiometric saturation quality assessment band [28,29]. For Sentinel-2, we first filtered for low-quality scenes through scene metadata flags [30]. As we found the pixel-level indicators inadequate for Sentinel-2 cloud and cloud shadow removal for our study area, we used a more detailed procedure in which “clouds are identified from the Sentinel-2 cloud probability dataset and shadows are defined by cloud projection intersection with low-reflectance near-infrared (NIR) pixels” [31].

We calculated a variety of spectral indices for LULC modeling from the Landsat/Sentinel-2 time series (Figure 2). These indices include tasseled cap indices of brightness (TCB), greenness (TCG), wetness (TCW), and angle (TCA) [32]. We included the Normalized Difference Vegetation Index (NDVI) due to its well-established history in vegetation type identification [33]. To contribute to the delineation of built-up areas and roads, we incorporated the Urban Composition Index (UCI) [34], Built-up Area Index (BAI) [35], Normalized Built-up Area Index (NBAI) [36]), and Built-up Area Extraction Index (BAEI) [36]. We used the UCI index for the Tier 1 LU classifications but included the three additional indices (BAI/NBAI/BAEI) that may give more power to separate road and paved areas from buildings and bare land in the higher resolution Tier 2 urban LC classifications. For water and wet area identification, we included the Modified Normalized Difference Water Index (MNDWI) [37] and a custom Wetness Index (WI) we developed by calculating the distance from the origin in the NIR band-Red band plane [38], in addition to the percentage of time over the year that a pixel is identified as water in Landsat or Sentinel-2 standard quality assurance flags (%Water). Additional information on these indices is provided in Appendix A.

The remote sensing data pre-processing and extraction steps were conducted in Google Earth Engine (GEE) [39] as it provides easy and unified access to all of our remote sensing data sets. The final time series for each data source was reduced to the set of minimum, maximum, mean, and range of each band or spectral index for each pixel across the year (Figure 2). In some places, particularly in Nigeria, cloud filtering resulted in limited coverage across some years. In such cases, we used the average value of adjacent years to interpolate values for the missing year. Neighborhood statistics and selected Gray Level Co-occurrence Matrix (GLCM) contextual metrics [40] were also calculated over the annual composite median image of the NIR and green bands, TCB, TCG, and TCW (Table 2).

To complement our spectral predictor sources, we incorporated radar-based Sentinel-1 Ground Range Detected (GRD) products within our classification frameworks (Figure 2), which is increasingly used for LC analysis due to its almost weather-independent operation (see [41]). The Sentinel-1 products were processed for quality in GEE based on the procedure explained in [42], which includes three steps of additional border noise collection, speckle filtering, and radiometric terrain normalization. The Sentinel-1 products were reduced to the annual minimum, maximum, mean, and range of backscatter metrics across each study year. Neighborhood statistics and GLCM metrics were calculated from the annual median composites (Table 2).

To further expand the predictor information to aid in delineating different LULC classes, we incorporated topographic, bio-climatic, night-time light, and additional spatial environmental data in our classification frameworks (Table 2). For night-time light, which is an important variable in identifying human settlements, we used the Visible Infrared Imaging Radiometer Suite (VIIRS) [43]. For bio-climatic variables, we leveraged both the WorldClim normal 30-year climate normal dataset [44] and TerraClim monthly temperature/precipitation statistics [45]. Additional data sources included the Continuous Heat-Insolation Load Index (CHILI Index) [46] derived from 90 m elevation data from the Shuttle Radar Topography Mission, iSDA soil texture class [47], and a map of World Ecoregions (RESOLVE) [48]. Among these additional datasets, VIIRS and TerraClimate features were calculated for each year, while WorldClim, CHILI Index, iSAD soil texture, and Ecoregion classes were static (e.g., not changing over years; Table 2). As we leveraged data sources that represent a wide range of spatial resolutions, we resampled all input data to a common 30 m resolution for Tier 1, and 10 m for Tier 2, for data extraction.

2.4. Reference Data Sampling Design and Interpretation

We selected a set of training pixels for Tier 1 and Tier 2 classifications for use in model training and initial assessment, with an additional independent set of map validation pixels for each tier and country (

Table 3. Number of interpreted reference pixels for training and validation for Tier 1 and Tier 2 across each study country.

Country	Tier	Training Pixels	Validation Pixels
Ethiopia	Tier 1, whole country	740	550
	Tier 2, final urban agglomerations	1172	700
Nigeria	Tier 1, whole country	687	700
	Tier 2, final urban agglomerations	1200	525
South Africa	Tier 1, whole country	957	1000
	Tier 2, final urban agglomerations	2897	1050

). Tier 1 LU classification training pixels were defined based on a pixel’s use for human purposes. Our LU classes included agriculture, bare, developed, forest, rangeland, water, and wetland. The LU classes were interpreted from the spatio-temporal context. For example, a pixel with grass LC may have a LU of natural grassland or farmland depending on the condition of other pixels around it and the condition of the pixel over a time span; we considered a 3-year time span, i.e., 1 year before and ahead of the current time, for determining LU. Tier 2 training pixels were selected within our delineated urban agglomerations (see Section 2.6.1 for more details on the delineation approach) across all three countries. The LC classes considered across our urban areas included barren, building, pavement, short vegetation, tall vegetation, water, and wetland. Full definitions of LU and LC classes are further presented in Appendix B. Note that, based on the availability of remote sensing data for each year within 2016–2020, each point produced up to five samples for training or validation. The training database was compiled by combining all years of samples, while validation was conducted for each year separately. The training databases for each classification tier were split using an 80/20 ratio for classifier training and testing for model selection. Once a candidate model was selected to be applied to the full national extent, we used the map at a chosen year to select a stratified random sample of pixels for independent map validation for accuracy assessments according to [49]. Aside from the sample selection approach, the same reference pixel interpretation methodologies were utilized for both training and validation data, hereafter referred to as reference data or pixels, although interpretation protocols varied between tiers.

We used a dual interpreter approach in reference data generation to determine LU (Tier 1) and urban LC (Tier 2) designations for a selected set of pixels for each tier across 5 years (2016–2020) and our three focal countries. A team of primary interpreters assigned initial labels which were then checked by a single secondary interpreter to ensure labeling consistency and to minimize data entry errors. A set of software tools aided in the reference data interpretations, which included the following: (1) Landsat time series viewer on GEE [50] for viewing spectral indices time series to help determine the LC of the selected pixel; (2) the TimeSync Plus program v4.7 [51] for reference label entry and additional visualization of annual Landsat image chips around the reference pixel; and (3) Google Earth Pro v7.3, which served as the main source of high-resolution imagery across years for the final Tier-1 LU, or Tier-2 urban LC, class designations. As there was no high-resolution imagery available for each year for many of the reference locations, and to facilitate identification of possible changes through the study years, interpreters viewed the reference pixels’ spectral indices time series in GEE and the annual context provided by the TimeSync Landsat image chips to aid in identification of LU or LC changes. The interpretation approaches are provided in more detail in Appendix C.

2.5. Model Selection and Map Validation

We employed a Random Forest (RF) classification modeling approach for both tiers of our LULC framework as it is widely used in remote sensing and known to provide high accuracy [52,53,54]. We tested a variety of parameter settings for the random forest models (scikit-learn v1.3.0) including the number of trees, maximum depth, and minimum leaf size. We found that 100 trees grown to unlimited depth with a minimum leaf size of four provided the best balance between complexity and model improvements. Initial model optimization was executed using the 80/20 train/test ratio of training data. After preliminary modeling, we decided to develop separate optimized models for each country (Figure 3).

Overall, 154 features for Tier 1 maps and 170 features for Tier 2 maps were calculated, but reduced to a limited set of less-correlated features (between 28 and 53 features) for the final models. To conduct feature selections, we filtered correlated features to reduce the complexity of each model. Correlated features are believed to have little impact on the performance of RF classifiers, but they do impact feature ranking [55]. For each country and product tier, we first calculated the correlation matrix for the training data and created a dendrogram (hierarchical clustering model) of feature correlations (Figure 3). Through an iterative process, we selected different thresholds to cut the dendrogram to generate a set of randomly selected features from each cluster, trained an RF model, and calculated model performance (Figure 3). A final set of 10–20 top-performing models were selected for further visual refining. This extra refining step was needed as we found that models with comparable model assessment results can still produce significantly different map products when applied spatially. We visually compared the model outputs within several heterogeneous test areas for the final model selection for each product tier and study country.

Within the post-processing steps for the Tier 1 LU maps, we removed small patches of less than 5 pixels with the same LU and replaced each pixel with the majority pixel value from the immediate eight neighbors. No post-processing was performed for the Tier 2 LC maps. For each country and year of our multi-tier product suite, we performed an accuracy assessment with the post-stratified validation sample. The overall and user/producer accuracies were then calculated as map-based area-adjusted values to represent the accuracy of the population (i.e., all mapped pixels) [49].

2.6. Multi-Tiered LULC Analyses

We highlight the potential value of our multi-tier LULC framework by (1) applying an automated urban delineation approach to identify urban areas across countries and consistently through time using the Tier 1 products, (2) quantifying urbanization-driven LU changes, and (3) further exploring urban LC densities and landscape configurations at individual city-scales, as well as for comparisons across national extents or between countries.

2.6.1. Delineating Urban Boundaries

As a basis for quantifying urbanization-driven LU changes, as well as for identifying urban areas for classifying our Tier 2 urban LC, we first needed to delineate urban boundaries consistently across national scales and through time. While studies and monitoring efforts may simply focus on the developed pixels within LULC map products or use administrative boundaries to delineate an urban area, these simplified approaches may fail to capture the complexities of urban extents, as well as introduce additional errors within the monitoring of urban extents across time. We employed a recently developed automated urban delineation approach [56] which we integrated to delineate urban boundaries to constrain our urban LC modeling and mapping. The automated urban delineation approach utilizes LULC data, gridded population data, OpenStreetMap data, and the openrouteservice tool to identify the dynamic boundaries of functionally connected urban agglomerations within country extents over time [56].

2.6.2. LU Change Analysis

We quantified LU changes that occurred during the expansion of all urban agglomerations within our study countries between 2016 and 2020. We used this tabulation to compare patterns of LU provisioning and change across countries.

2.6.3. Characterizing Urban LC Patterns and Change

The Tier 2 LC maps were used to summarize building and green space densities and landscape configurations within the urban agglomerations for our starting and ending map years, 2016 and 2020, to further quantify changes across our study period. Our analysis was limited to the largest cluster within each urban agglomeration and its urban core. Only urban core areas with a minimum area of 500 hectares in 2020 were considered, while suburban, open, and rural spaces were excluded to focus our analyses on the main urban center of each agglomeration. Each of these selected urban maps was converted to a set of contiguous patches of the same land cover type for pattern analysis. Patches of land cover smaller than 4 pixels (400 m²) were removed for landscape configuration analysis (but not for density analysis below).

Landscape configuration within these patches was analyzed using the contagion (CONTAG), clumpiness (CLUMP), and mean Euclidean nearest neighbor (ENN_MN) metrics calculated with the FragStat v4.2 software [57] (

Table 4. List of calculated landscape metrics and their purpose in our urban land cover analyses.

Metric Name	Symbol Definition	Purpose
Contagion index	CONTAG	Spatial distribution (dispersion) and mixing (interspersion) of all land cover classes
Clumpiness index	CLUMP	Level of aggregation (clumpiness), calculated for building land cover
Euclidean nearest neighbor distance—Mean	ENN_MN	Mean of Euclidean nearest-neighbor distance, calculated for vegetation land cover

). The CONTAG metric reports on the distribution and intermixing of different land cover patches, with values between 0 and 100. Lower values mean less aggregation (more uniform distribution and better intermixing) and higher values mean more aggregation. The CLUMP metric was calculated for the building class. This index ranges from −1 when the patch type is maximally disaggregated, to 1 when the patch type is maximally clumped, and returns a value of 0 for a random distribution. Clumpiness has also been used to measure the degree to which urban patches are aggregated instead of more diversely distributed [58]. The ENN metric was calculated for the combination of the short and tall vegetation classes. ENN has been used to quantify urban green space distribution for assessing ecosystem services, ease of access, and environmental justice [59]. Assuming that more diversity, better spatial distribution of different land covers, and less distance between green patches are favored by city inhabitants, lower values on the selected three metrics are preferred over higher values.

We also conducted a density analysis for buildings to provide insight into the degree and spatial patterns of denseness of city structures within and across cities. Building density was calculated as the number of pixels within the Tier 2 LC maps labeled as ‘building’ within a neighborhood moving window of 15 × 15 pixels. A window size of 15 × 15 pixels was selected as an appropriate scale after visual inspection of several window sizes. We calculated univariate statistics such as the mean and range of density values for each city. The density values were also binned into three groups representing low, medium, and high-density areas using the K-means clustering algorithm, and the centroid was calculated for each of the three bins. Using a clustering algorithm and not simply binning by a fixed threshold results in the range of density values in each bin varying from city to city, but it can help quantify and compare the degree of denseness of different cities. For instance, consider two cities with the same minimum and maximum pixel density values, but city A has more high-density pixels than city B. This causes the high-density bin centroid value of city A to be higher than the high-density bin value of city B and indicates that city A is “denser” than city B. If city A’s maximum density is higher than city B’s maximum density, then the high-density bin centroid value for city A will again be higher than city B. The benefit of using the K-means clustering algorithm over simple mean statistics is that it is less dependent on the total number of map pixels, while averaging density values over a large area with a majority of low-density pixels can effectively nullify the existence of patches of high-density pixels.

3. Results

3.1. Tier 1 LU, Tier 2 LC Map Assessments

Feature importance values, as reported by the RF classifier, differed considerably between countries (Appendix D). We did not find a common pattern for most of the important features, except that the night-time light data was found to be an important feature for the LU classifications across all countries, in particular, for identifying the developed class. However, including night-time light data led to blocky image artifacts because the original resolution was 460 m. To address this, we devised a hybrid modeling approach for the Tier 1 LU classifications to remove blocky artifacts (refer to Appendix D.1). Multi-resolution input data sets did not produce artifacts in the Tier 2 maps, so no hybrid model approaches were necessary.

We found comparable map validation results for all years within a given study country, thus we will focus on the presentation of 2020 performance results (full information is provided in Appendix E). We calculated map class user accuracy (Map UA), producer accuracy (Map PA), average F1, overall map accuracy (Map OA), and its 95% confidence interval (CI) for Tier 1 products (

Table 5. Map accuracy assessments for the Tier 1 land use maps for 2020 (values are reported as percentages).

	Ethiopia			Nigeria			South Africa
Class	Map UA	Map PA	Map F1	Map UA	Map PA	Map F1	Map UA	Map PA	Map F1
Agriculture	61.2	77.0	68.2	72.4	77.4	74.8	64.7	89.9	75.3
Bare	45.5	95.8	61.7	81.0	27.9	41.5	53.7	6.9	12.2
Developed	33.1	23.4	27.4	81.5	47.8	60.3	92.0	24.1	38.2
Forest	76.4	80.2	78.3	62.7	59.0	60.8	53.7	82.9	65.1
Range	86.4	72.0	78.5	57.9	57.4	57.6	77.4	88.5	82.6
Water	96.7	100.0	98.3	75.3	92.9	83.2	86.3	98.8	92.1
Wetland	57.7	64.7	61.0	78.3	49.7	60.8	67.6	8.3	14.8
	Map OA	95% CI		Map OA	95% CI		Map OA	95% CI
	74.6	7.3		65.9	5.4		73.6	6.8

). The best Map OA was obtained for Ethiopia at 74.6%, while Nigeria had the lowest Map OA (65.9%). In post-stratified area-weighted accuracy assessments, the Map OA is calculated as the strata-area weighted sum of individual class accuracies [49], which allowed us to examine the contribution of individual class accuracies to OA (presented in Appendix E). The developed class Map PA was low in all countries, as it was a rare class relative to other LU classes at national extents. Map PA is highly affected by the class coverage, and each misclassified pixel for a rare class will significantly lower Map PA. We also found that the developed class Map UA is significantly different between Ethiopia and other countries, which was due to higher confusion between the developed and rangeland classes in Ethiopia. Additionally, we found that the Ethiopia model generated more false positives for the developed class than the other two country models. These falsely classified developed pixels were suppressed in a later process when delineating urban agglomerations, and, therefore, this bias was minimized for our urban pattern analysis.

Similar to the Tier 1 LU, map accuracy assessments for Tier 2 LC maps were carried out for all years, although we found consistent results across years. The results for target countries for the year 2020 are shown in

Table 6. Map accuracy assessments for the Tier 2 land cover maps for 2020 (values are reported as percentages).

	Ethiopia			Nigeria			South Africa
Class	Map UA	Map PA	Map F1	Map UA	Map PA	Map F1	Map UA	Map PA	Map F1
Barren	40.1	60.4	48.2	66.9	31.8	43.1	68.8	36.4	47.6
Building	68.1	55.6	61.2	58.4	85.2	69.3	48.0	85.8	61.5
Pavement	29.3	63.1	40.0	66.2	26.9	38.3	46.1	41.7	43.8
Short vegetation	89.8	85.7	87.7	75.2	74.3	74.7	85.4	67.5	75.4
Tall vegetation	62.7	63.2	63.0	65.7	70.1	67.8	46.5	55.9	50.8
Water	96.7	95.8	96.3	90.0	96.4	93.1	77.5	82.3	79.8
Wetland	52.3	58.8	55.4	86.8	58.9	70.2	16.9	90.2	28.4
	Map OA	95% CI		Map OA	95% CI		Map OA	95% CI
	78.3	5.1		66.3	5.3		62.8	3.7

(full information is provided in Appendix E). We found that Ethiopia had the highest Map OA (78.3%), followed by Nigeria (66.3%) and then South Africa (62.8%). Deficiencies in Map UA and Map PA were related to confusion between classes that were common to all countries. We encountered numerous cases of confusion between barren and building and also building and pavement when the pixels are mixed, or the material is similar (particularly due to the similarity of building roof material and pavement). Confusion between short and tall vegetation is also one of the more common confusions. There was limited confusion between vegetation classes and barren/impervious surface classes.

3.1.1. Tier 1 LU Classification and Change Analyses

Annual Tier 1 national-level LU maps were created based on separate models optimized for each country. While the models were optimized for each country, the same seven-class classification scheme was applied annually to the three study countries (Figure 4), allowing for LU change assessments across our 5-year study period and comparisons of LU patterns and change dynamics between countries.

3.1.2. LU Distribution and Its Change within Urban Expansion

We calculated the percentage of LU classes based on the Tier 1 maps across three years to depict temporal trends in each country (Figure 5; for full data see Appendix F). South Africa is dominated by rangelands (~75%), while in Nigeria the majority of land area is dedicated to agricultural land uses (~45%). Ethiopia falls in the middle of these extremes, with more area in rangelands than Nigeria (~53%) and more agricultural areas than South Africa (~20%). In all three countries, urban land area is quite small (less than 1.5%). Although Ethiopia has the lowest overall urban land area (0.5–1.1%), from 2016 to 2020 it exhibited the highest percentage of change compared to the other two countries. Furthermore, in Ethiopia, this increase in urban land area was mostly a function of agricultural land conversion (

Table 7. Summaries of developed land use changes within final urban agglomerations of our study countries from 2016 to 2020. The table contains the total developed area (in hectares) for 2016 and 2020 and details the conversion from other land use types to developed land across the study period. The values calculated in the last two columns are absolute change in developed area (in hectares) from 2016 to 2020 and relative change in developed area (percentage) across the study period.

	Developed (2016)	Net Conversion from Other Land Uses from 2016 to 2020						Developed (2020)	Absolute Change	Relative Change
		Agriculture	Bare	Forest	Rangeland	Water	Wetland
Ethiopia	153,779	54,029	16	3066	19,635	2	57	230,584	76,805	49.9%
Nigeria	658,438	89,162	0	22,940	15,839	−36	97	786,440	128,002	19.4%
South Africa	635,139	1790	664	3955	75,892	−3	26	717,463	82,324	13.0%

). Nigeria also experienced large urban changes over this time period, with more expansion of developed land in absolute area than the other two countries, primarily converted from agricultural lands. Although South Africa experienced the lowest percent of change in developed area (13%), more land area was converted to developed land uses than in Ethiopia, and most of this was due to conversion from rangelands. The changes that we highlight are specific to our study period of 2016–2020.

3.2. Characterizing Urban LC

3.2.1. Annual LC Classification Models and Maps

Delineation of urban agglomerations from the Tier 1 LU maps resulted in 192 agglomerations for Ethiopia (Appendix G Figure A12), 323 agglomerations for Nigeria (Figure 6), and 261 agglomerations for South Africa (Appendix G Figure A13). For example, the Nigerian agglomerations are presented in Figure 6, where we can see a high concentration of urban agglomerations in the central north around Kano, in the central south near Onitsha, and in the southwest near both Lagos and Ibadan. One example of a rapidly expanding city is Benin City, Nigeria, which has seen building expansion towards the edges of the urban agglomeration between 2016 and 2020 (Figure 6), where our Tier 2 LC products show the patterns of these expansions, including changes to green spaces and other land covers with the increase in building density.

3.2.2. Urban LC Patterns

Leveraging the Tier 2 urban LC products, we calculated landscape metrics (CONTAG, CLUMP, and ENN MN) for the years 2016 and 2020 to assess LC configurations (

Table 8. Mean and range of landscape metric values calculated for the urban areas of our study countries in 2016 and 2020. Note that, due to population size thresholds used in the identification of urban areas, some areas may be included in the 2020 assessment that were not included in the 2016 assessment. This is because an area may not have been identified as urban in the initial year (2016) but grew in population and met thresholds to be classified as urban by the final year (2020).

Country	Year	CONTAG (Unitless)	CLUMP (Building Class, Unitless)	ENN MN (Vegetation Class, Meters)
Ethiopia	2016	39, 31–52	0.7, 0.58–0.83	32, 25–43
	2020	41, 32–53	0.7, 0.56–0.84	30, 23–43
Nigeria	2016	58, 39–82	0.85, 0.76–0.92	38, 27–58
	2020	58, 41–76	0.83, 0.73–0.92	35, 26–57
South Africa	2016	47, 32–64	0.75, 0.57–0.86	38, 25–65
	2020	49, 34–67	0.76, 0.56–0.89	36, 24–62

). We found the CONTAG maximum value was stable for Ethiopia and South Africa from 2016 to 2020, while for Nigeria it decreased. The CLUMP and ENN_MN metric ranges remained relatively stable over the study period for all countries, which may mean that in general, the building densities and the average distance between vegetated patches did not change significantly across our study period (Figure 7). Based on the values of calculated metrics over different cities in each country, we did not see a correlation between calculated metric values and the size of cities.

However, we found a relatively higher range and mean for CONTAG values across Nigerian cities in both 2016 and 2020, suggesting that the LC types in Nigerian cities tended to be more aggregated and less proportionally distributed compared to the other two countries. While average CONTAG values for Nigeria were the same for 2016 and 2020, the range was shrinking, which may indicate that some Nigerian cities have greater land cover intermixing (lower CONTAG). This is consistent with Nigeria’s higher CLUMP values for the building class compared to Ethiopia and South Africa. For Ethiopia, the maximum CONTAG, CLUMP, and ENN_MN values were lower than for the other two countries, exposing less dense and more interspersed LC configurations.

In addition to landscape metrics, we calculated building density statistics over each city and the centroid value of each density bin for the study countries. Our results for the year 2020 were comparable to those of 2016, with the distribution of high-density bin centroid value for the year 2020 depicted in Figure 8. Ethiopian cities show both lower overall mean density and lower prevalence of high-density neighborhoods (as shown by the centroid of the high-density bins) compared to South Africa and Nigeria, while the Nigerian cities exhibit the highest densities. This result aligns with the CLUMP results (Figure 7).

4. Discussion

The aim of our work was to develop an integrated mapping framework to support multi-scale urbanization-related LULC assessments. The synchronized tiers within our developed methodology allow for the identification of dynamic urban boundaries and quantification of urbanization-driven change using LU data from Tier 1, and an improved representation and analysis of the “red blob” (urban areas) using higher resolution LC classifications from Tier 2. In testing the robustness of our mapping framework, we obtained map accuracies between 60 and 80%, depending on the specific country and product tier, using single country-specific models across all years. Other available LULC map products at the regional/global extent have achieved a comparable overall accuracy range, although typically developed and applied for a single year [15,17,18,60,61,62]. The consistency in map accuracies across years in our study illustrates the potential of our framework for use in ongoing monitoring and is further supported by the flexibility and adaptability of our methodologies which can be applied to new years of imagery, updated or enhanced using additional training data, and applied across new locations.

The outputs of our multi-tier mapping framework and analyses add to the available knowledge of LULC over time, as well as to the data needed for near-recent, retrospective, and comparative LULC analyses in Ethiopia, Nigeria, and South Africa. In our applications, we described the LU composition of the study countries across our study period and revealed differences in LU conversions. Notable findings over the short 5-year period included Ethiopia experiencing the largest relative increase in developed land, Nigeria experiencing the largest absolute increase in developed land, and South Africa displaying the smallest relative increase. In addition to the analyses presented here, [56] incorporated the Tier 1 LU data products with auxiliary datasets across our study countries to develop an automated method for capturing dynamic urban boundaries, calculating SDG Indicator 11.3.1, and supporting urbanization metrics, examining spatial patterns of new development, and identifying hotspots of urban land use expansion [56]. The analyses of Cardenas-Ritzert et al. (2024) not only contributed to the broader knowledge of recent urbanization trends and patterns in the three African countries, but also highlighted the potential use of our LU product when paired with additional open-source datasets [56]. There remain significant opportunities for coordinated LULC analyses utilizing both tiers of our African map products, with further investigation needed to extract the full potential of the higher resolution LC classifications of our Tier 2 products. It will also be important to test the applicability of our multi-tier methodology to countries outside of Africa.

The LU and urban LC products produced through our integrated framework fill the spatial data gap for annual multi-resolution LULC analyses and change information specifically focused on urban landscapes and dissecting the “red blob” of a single developed LU class. Our urban LC pattern analysis results indicate that cities with similar geographical extents were found to have a wide range of building density and other landscape metrics. We did not find any change in the range of landscape or density metrics in the selected countries within our study period of 2016–2020. Nor did we find any trends indicating a correlation between the calculated urban LC pattern metrics and the overall size of the city. This suggests more nuanced drivers of urbanization patterns that warrant further investigation, where our multi-tiered product suite is well suited to support such investigations. While our findings allude to general differences in urban patterns across national scales, further analyses are needed to better understand the complex drivers behind urban LC configurations and their change patterns across different countries, city sizes, and rates of urbanization.

The worldwide availability of 10 m resolution Sentinel-2 imagery (for select bands) serves as a valuable and accessible resource for generating LC mapping products, which can be crucial for eliciting information on urban quality and function in data-limited regions. Our annual 10 m resolution urban LC products highlight the utility of Sentinel-2 imagery, complemented by Sentinel-1 SAR data, for consistent characterizations of LC across urban areas within national extents and through time at appropriate resolutions for urban pattern analyses and planning. While spatial products that classify urban LC features to elicit information on urban quality and function are critical for monitoring and management, many previous studies have relied on very high-resolution imagery sources (≤5 m resolutions). For example, Huang et al. (2022) used GaoFen-2 imagery to map urban land cover in a district in Beijing, China, combining six ecological indicators into the development of an urban ecological quality metric [63]. Milošević et al. (2023) used WorldView-3 imagery to map the city of Split, Croatia, and calculated several urban planning indicators to inform the spatial-functional development of the city [64]. While these very high-resolution assessments of urban landscapes can provide valuable information for landscape planning and analyses of urban patterns, the very high-resolution imagery sources are often not consistently available across large extents and/or are expensive to acquire, limiting the accessibility of such methods. Other global or continental mapping efforts have developed large extent LU products using widely accessible data sources with comparable accuracies as our product suite [15,17,18], some even starting to embrace the value and resolutions of Sentinel-2 imagery for more recent years [60,65,66,67], although many of these efforts are limited to single years and/or limited change analyses [60,68], lower resolutions less appropriate for urban LC analyses [69,70,71], or may not provide the nuanced LC classes appropriate for assessing heterogeneous urban landscapes [16,18].

Our integrated approach to classifying national-level moderate resolution LU and higher resolution targeted urban LC across all delineated cities shows value for supporting global monitoring initiatives, such as objectives under the Sustainable Development Goals (SDGs). For instance, SDG 11.7 focuses on the proportion and accessibility of green spaces within cities, and SDG Indicator 11.3.1 aims to track urban expansion and population change [72]. Furthermore, our products can support regional initiatives such as the Cities Alliance (https://www.citiesalliance.org/; last accessed 26 June 2024) or the Secondary Cities initiative (https://secondarycities.state.gov/; last accessed 26 June 2024). The LC maps, as standalone products or paired with auxiliary data sources (e.g., demographic data) and spatial analysis tools, can be utilized to carry out meaningful assessments for localized urban-related evaluations. These assessments can include patterns of building densities and distributions of green spaces, and the maps can be used consistently across broad extents as well as for monitoring changes through time.

In particular, geospatial examinations of natural areas and their provisioning of social, ecological, and ecosystem services within urban environments are fundamental for supporting urban land management decisions and improving urban quality of life. Urban sustainability research of this kind frequently exploits publicly available satellite data to examine urban green spaces [73,74], urban heat island effects [75], natural land fragmentation [76], ecosystem service mapping [77], and urban habitat conservation [78]. The resolutions and classes included in our Tier 2 urban LC products can support this variety of studies. For example, short and tall vegetation, which are classes in our Tier 2 products, have been studied regularly for their influence on land surface temperature [79,80,81,82]. Considering urbanization and climate change projections, products such as ours will be crucial for the continued monitoring of the impacts of intensified urban heat islands and the alleviating functions of natural spaces [83], especially in Africa where these studies have been less common [84]. Furthermore, urban green and open spaces benefit the physical and mental health of urban dwellers through recreation, social interaction, and exposure to nature [85,86] and spatially explicit information on short and tall vegetation can aid in evaluations of the access, quality, and distribution of green and open spaces in urban areas [87,88]. These examples are only a few of the possible usages of our LULC products for various ecosystem service analyses, as a multitude of opportunities exist, demonstrating their potential in guiding urbanization monitoring and urban landscape planning around the globe. In particular, the integrated multi-tier framework elicits information on broader urbanization-driven LU changes and trade-offs in LU provisioning, while supporting the delineation of urban agglomerations and the characterization of heterogeneous urban landscapes. The enhanced Tier-2 urban LC products then allow for simultaneous assessment of the complex changes occurring within urban agglomerations (e.g., infilling, densification, and changes to urban green spaces) to complement broader urban LU expansion monitoring.

While our products and framework can be advantageous to future LULC mapping efforts, we observed various modeling caveats in our work. Firstly, map accuracies were consistent across all years of our study, but model performances did vary across countries. During preliminary model assessments, we found that country-specific models outperformed a single, generalized model, at least partially due to differences in data quality and availability. The attainability of cloud-free, high-quality remote sensing data differed across the three countries, and was a particular challenge in specific regions, such as southern Nigeria, which required the interpolation of predictors from surrounding years. Careful consideration is suggested when applying our framework to areas that are likely to face the same limitation. Additionally, we chose to leverage the flexible and widely used Random Forest modeling approach within our integrated framework, although more computationally intensive deep neural networks may provide favorable performances but require significantly larger reference datasets than attainable within the scope of this study. Lastly, additional factors may have contributed to varying performance across our multi-sensor models, including inaccurate co-registration of sources across some areas. For example, Stumpf et al. (2018) reported a misregistration of up to 38 m between Landsat-8 and Sentinel-2 imagery [89]. Localized studies in time and space may conduct image preprocessing enhancements, such as co-registration. Within our study conducted at national extents and annually across five years, we were focused on developing a reproducible methodology applicable across large areas and through time. While some limitations exist within our approach that are inherent to studies incorporating multiple sources of remote sensing imagery across multiple years, we conducted robust annual map validations exhibiting consistent performance across years for all products and optimized models for each country to attempt to maximize accuracies. Not only does our work demonstrate the validity of our Africa LULC map products but illustrates the viability of applying our framework to other regions.

5. Conclusions

Incorporating multiple resolutions of analyses when studying urbanization patterns may help elicit more nuanced depictions of how urbanization is manifesting on the landscape. Studies have shown that relying on basic relationships between LU and services can lead to decision-making that amplifies social-environmental inequities and the development of biased urban planning policies [90]. Our multi-tiered, complementary LULC mapping framework allows us to gauge the “bigger picture” of urbanization in a country while also examining country-level characteristics of LU change and diving deeper into heterogeneous LC patterns within cities. Each data set captures unique patterns and characteristics of change at its respective scale, but together can be used tangentially to support the findings derived under each other, illuminating potentially new information. Not only are we able to derive new information at different scales or among scales but we can also provide insights into the significance of resolution in these assessments and the suitability of map characteristics for specific LULC-related objectives.

Further research is required to explore improvements in mapping efforts using more advanced AI models, full yearly time series for the sensors (e.g., phenological coefficients), and other data sources such as road networks, administrative boundaries, and building heights. We also acknowledge that there is a large suite of additional urban pattern analyses possible beyond those tested in this study and stress the importance of testing for appropriate data resolutions for different pattern metrics or changes in patterns observed depending on the base image resolutions. Future work should expand on the examinations within this study by integrating local knowledge and on-the-ground data related to services and demographics with the remote sensing data products to continue to test the potential value of the mapping products and multi-tier framework. Additional work is needed to balance this approach with the need to assess urban function—those attributes of the landscape that remain intangible in the remote sensing toolbox—to understand urban quality related to equity and access to basic services.