Original Research

Mapping vegetation fractions in South Africa’s protected semi-arid savanna ecosystems: Assessing the impact of fractional cover proportions on the prediction accuracy

Hilma S. Nghiyalwa, Eliakim Hamunyela, Tshililo Ramaswiela, Jussi Baade, Christiane Schmullius

Received: 07 Nov. 2025; Accepted: 29 Apr. 2026; Published: 30 July 2026

Copyright: © 2026. The Authors. Licensee: AOSIS.
This work is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license (https://creativecommons.org/licenses/by/4.0/).

Abstract

Accurate estimation of vegetation composition is essential for effective conservation management strategies in protected savannas. In this study, we integrated annual temporal metrics from Sentinel-1 (S1) and Sentinel-2 (S2) time series at a 10 m resolution for 2017-2021 in Benfontein Nature Reserve (BNR). We trained a Random Forest Regression (RFR) model using training data derived from very high-resolution imagery (VHR) from Google Earth Pro^© and later used for results accuracy assessment and validation. Error metrics for predicting trees, shrubs and grasses were derived at each year. We selected the most important features for the most accurate models. Model error metrics were further derived by simplifying pixel fractions into broader fraction categories. Shrubs S1+S2 and S2 models produced the lowest mean absolute error (MAE) of 3.91%, while grass produced the highest MAE of 13.71% from S1 data. Trees produced the second-best accuracies obtained consistently with S1+S2 data. Trees achieved the highest R² values of up to 0.76% from S1+S2 data, followed by grass highest R² of 0.74% from S2 data. When predicting based on broader pixel categories, errors decreased significantly for near-pure pixels; with the largest decrease observed for pixels with 76% – 100% grass cover. Contrastingly, MAE for woody covers (shrubs and trees) increased as fraction in the pixel cover increased. Visible green and blue bands, VH and SAVI were the most important features for estimating trees. In addition to blue, VH and Coherence ranked high for shrubs, while blue and green ranked high for grass as well.

Conservation implications: The study is conducted in a nature reserve. Accurate information on composition and dynamics of vegetation is essential for parks management. Predicting change in woody cover is critical for monitoring woody encroachment that impacts herbivory and fires.

Keywords: vegetation fractions cover; Benfontein Nature Reserve; time series; optical; radar; VHR; conservation; savannas.

Introduction

African semi-arid savannas consist of trees, shrubs and grass vegetation compositions, which are essential to biodiversity and ecosystem services (Scholes & Archer 1997). Savannas are defined and maintained by their trees-grass co-existence, and changes in these vegetation compositions can have far-reaching impacts on ecological functions. Changes in the dynamics of grasses and woody vegetation have effect on fires and grazing regimes, the grass-trees balance and therefore the biodiversity of savannas (Archibald et al. 2010; Simpson, Archibald & Osborne 2022; Snyman 2015). The impacts of changes in savanna woody and herbaceous vegetation composition could mean that fire occurrences and intensity are altered (Govender, Trollope & Van Wilgen 2006). In addition, changes in herbaceous vegetation composition of savannas could also have negative impacts on the carrying capacity for wildlife grazers, and therefore affect biodiversity and ecosystem services (Geißler et al. 2024). Increased grazing in grass-dominated savannas can reduce herbaceous biomass and lead to changes in grass layer composition and structures, resulting in fragmented patches with reduced flammability (Archer, Andersen & Predick et al. 2017; Huntley 2023; Smit Izak & Coetsee 2019). Increased grazing and browsing pressure in mixed trees and grass savannas can directly reduce fire frequency and intensity through removal of fuel and changes to the grass layer (Smit Izak & Coetsee 2019).

This is especially true for protected areas that are essential to conserving the unique biodiversity of the semi-arid savannas (Huntley 2023; Riggio et al. 2019; Timis-Gansac et al. 2025). Accurate estimation of these vegetation compositions is therefore essential for monitoring of change and for informed conservation and management strategies. Furthermore, vegetation compositions are responsible for the maintenance of savanna equilibrium, which is key to biodiversity conservation, especially in protected areas that are essential to conserving the unique biodiversity of the semi-arid savannas (Huntley 2023; Riggio et al. 2019; Timis-Gansac et al. 2025, Sankaran, Ratnam & Hanan 2008). Currently, protected semi-arid savannas are increasingly threatened by increasing climate variability and anthropogenic changes (Archibald et al. 2019; Geißler et al. 2024). Semi-arid savannas also face increased shrub encroachment, which affects the availability of grasses and threatens their resilience and their unique biodiversity (Archer et al. 2017; Belayneh & Tessema 2017; Scholtz et al. 2022; Stevens et al. 2016). Changes in grasses in semi-arid savannas further affect the dynamics of fire and herbivory dynamics, which could threaten the grass-tree dynamic and biodiversity of the savannas. Reductions in grass means that there is less available forage for wildlife but increases in grass may also result in increased fuel for wildfires in protected areas, which can lead to a depletion of woody vegetation (Archer et al. 2017). In trees and grass dominated savannas, increased grazing alters fire frequency and intensity directly and removes fuel because of changes in the herbaceous layer. Increased browsing also indirectly increases fire because of reduced woody recruitment and may reduce woody-grass competition (Scholtz et al. 2022; Smit Izak & Coetsee 2019).

To ensure informed management strategies and accurate monitoring of semi-arid savannas, there is a pressing need for efficient and accurate methodologies for estimating vegetation compositions. However, estimating woody vegetation compositions can be challenging because the savanna compositions are highly influenced by interacting dynamic factors such as fires and herbivory (Sankaran et al. 2008; Scholes & Archer 1997). This results in complex mixtures and patches of shrubs, trees and grass, with sudden changes in the landscape, and different vegetation gradients (Sankaran et al. 2005; Sankaran et al. 2008; Scholes & Archer 1997). In addition, this highlights the need to develop estimation methods that are capable of capturing the complexities of semi-arid savannas.

Estimating shrubs and trees separately is essential as different vegetation forms have different ecological functions. For example, dramatic changes in savanna shrub cover have been shown to affect the species diversity of animals in given areas (Sirami et al. 2009). Vegetation composition and structure is important for species diversity such as birds. Variations of birds’ species assemblages have been observed along a gradient of woody cover, and some species may differ between contrasting habitats of woodlands and grassland (Basile, Storch & Mikusiński 2021; Godoi et al. 2018).

Earth observation offers tools for mapping and monitoring of savanna ecosystems. Increased availability of freely available multispectral and multitemporal data archives coupled with big data processing abilities led to increased mapping at various temporal and spatial scales (Dritsas & Trigka 2025). Despite the advancements, there is still a lack of detailed vegetation composition data at local scales, especially those that separate trees from shrubs (Nghiyalwa et al. 2021). Mapping woody vegetation in semi-arid savanna can be a complex task. Previously, classifications methods coupled with medium resolution 30 m Landsat data have been used to map vegetation classes in savannas (Hüttich et al. 2011a; Schwieder et al. 2016). However, discrete classification methods such as pixel-based classifications, hard classifications or decision-trees and Random Forest decision trees fail to fully capture the savanna’s heterogeneity (Nghiyalwa et al. 2021). Spectral unmixing methods are therefore preferred to capture the full gradients of the savanna and have been used to derive vegetation fractions with optical 30 m Landsat (Gessner et al. 2013; Nagelkirk & Dahlin 2020), and recently at 10 m – 20 m resolution with Sentinel-2 (Harkort et al. 2025; Vermeulen, Munch & Palmer 2021).

Because optical data tend to be affected by clouds, integrating optical data with radar data is gradually gaining momentum (Borges et al. 2020; Lopes et al. 2020; Symeonakis et al. 2018). Radar data are minimally affected by clouds, and radar structural information can complement the spectral information from optical data to improve estimation. A few studies have leveraged the fusion of radar and optical data to map vegetation and found accuracy classification improvements (Higginbottom et al. 2018; Karakizi et al. 2023; Naidoo et al. 2016).

Numerous studies have shown that full-year, single-season or multiseason spectral-temporal metrics (STMs) for combined Synthetic Aperture Radar (SAR) and optical data can be robust for mapping fractional cover in savannas (Gessner et al. 2013; Higginbottom et al. 2018; Wessels et al. 2019). Specifically, multiseason data for SAR alone or SAR combined with optical are reported to improve mapping accuracies (Higginbottom et al. 2018; Karakizi et al. 2023; Naidoo et al. 2016; Symeonakis et al. 2018). Naidoo et al. (2016) demonstrated that integrating seasonally appropriate optical and SAR data yields better woody fractional cover estimates. SAR imagery often produces better accuracies during dry season because of the high contrast between woody and herbaceous vegetation during that season (Higginbottom et al. 2018; Urbazaev et al. 2015). Studies have, however, mainly focused on a single woody vegetation class and have rather used the L-band PALSAR (Phased Array type L-band Synthetic Aperture Radar) data and focused at larger spatial resolution of 105 m (Naidoo et al. 2016), 30 m, 60 m, 90 m and 120 m (Higginbottom et al. 2018) and 50 m (Urbazaev et al. 2015), with very few studies focusing on the use of 10 m Sentinel-1 C-band for fractions in savannas (Baumann et al. 2018; Karakizi et al. 2023).

Our study conducts a spatio-temporal analysis over a period of 5 years (2017–2021) evaluating the performance of Sentinel-1, Sentinel-2 and their combination (Sentinel-1 + Sentinel-2) when estimating trees, shrubs and grass vegetation fractions. Random Forest Regression (RFR) is trained and validated to derive fractions estimation errors for each year. We addressed the following questions: (1) Does the integration of radar (Sentinel-1) and optical (Sentinel-2) improve the accuracy estimation of shrubs, grass and trees in a savanna environment? (2) What is the influence of vegetation cover percentage in the pixel on the prediction error? (3) Which multitemporal metrics are important for the accurate estimation of vegetation cover fractions?

Description of the study area

Location and climate

Benfontein Nature Reserve (BNR) is a 9486.67 hectare private wildlife reserve located approximately 10 km southeast of Kimberley, at the border of the Northern Cape and Free State Province, South Africa (Benfontein Private Nature Reserve – South Africa | DEIMS-SDR). The reserve is owned by De Beers Consolidated Mines but serves as a wildlife conservation site. Benfontein Nature Reserve is a South Africa Environmental Observation Network (SAEON) long-term ecological research site with established phenological studies. The site is classified as having a semi-arid climate, with mean annual rainfall ranging from 134 mm to 419 mm (Maluleke et al. 2025; Mogonong et al. 2023). The wet season starts in October to the end of April, while the dry season lasts from May to September (Figure 1). Minimum temperatures during the winter months (May to August) can reach -5 °C, while maximum temperatures during the summer months (October to March) can reach up to 40 °C.

FIGURE 1: Climate diagram for Kimberley showing mean monthly temperature and monthly precipitation.

Land cover and land use description

The reserve is located at the intersection of three major biomes: Nama-Karoo/Northern Upper Karoo Biome, the Grassland Biome and the Kimberley Thornveld Savanna Biome (Mucina & Rutherfold 2006). The Karoo Biome consists of very short dwarf shrubs, with occasional taller shrubs. The Grassland biome comprises of open grassland with diverse grass communities. The Savanna Thornveld contains scattered Vachellia erioloba (Camel thorn) trees with a continuous grassland understory and a sparse shrub layer (Mucina & Rutherfold 2006). Vegetation within the site (Figure 2) is dominated by an open savanna featuring a continuous herbaceous layer composed primarily of grasses such as Schmidtia pappophoroides (Kalahari grass) and Stipagrostis uniplumis (Bushman grass), interspersed with scattered tree species of Vachellia erioloba (Camel thorn) and Vachellia tortilis (Umbrella thorn), and shrubs such as Ziziphus mucronate (Buffalo thorn) and Grewia flava (Brandy bush) (Bezuidenhout et al. 2015). Vegetation is highly influenced by variations in terrain, underlying geology and soil types. Mild to severe wildfires occur sometimes during the dry season (September to October). Herbivory is characterised by a large number of indigenous ungulates (Maluleke et al. 2025).

FIGURE 2: Benfontein Nature Reserve study area.

Methods and material

Sentinel-1 and Sentinel-2 data processing

Our observation data are based on Sentinel-1 (S1) and Sentinel-2 (S2) data available in the South Africa Land Degradation Monitor (SALDi) data cube, which assembled the Corpenicus Sentinel data provided by the European Space Agency (ESA) free of charge (Schmullius et al. 2024). All S1 and S2 data used in the methods workflow (Table 1) are pre-processed to the Analysis Ready Data (ARD) specifications detailed in (Yuan et al. 2022). All available S2 10 spectral bands along with Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), Normalised Difference Vegetation Index (NDVI), Sentinel-1 vertical transmit, vertical receive (VV) and vertical transmit, horizontal receive (VH) backscatter coefficients, VV-Coherence and Radar Vegetation Index (RVI), were downloaded from January 2017 to December 2021 (see Table 1). The optical S2 are already pre-processed to Level-2A ARD data and are cloud shadow masked and are radiometric and topographic corrected to remove terrain-induced illumination variations and atmospheric errors. All spectral bands were spatially resampled to a 10 m spatial resolution.

Sentinel-1 data have been radiometrically terrain corrected with the Copernicus 30 m Digital Elevation Model (DEM) to account for topographic distortions. The backscatter coefficient (β⁰) is normalised to gamma naught (γ⁰_rtc) using the calculated scattering area and terrain-induced variations in backscatter intensity were effectively removed (Yuan et al. 2022). The data have been projected onto the WGS 84 grid with a range Doppler terrain correction using bi-linear interpolation for a smooth output. The VV and VH polarised γ_rtc⁰ images in SALDi data cube contains essential per-pixel metadata layers (which include a radar shadow mask, local incidence angle image calculated based on the DEM and scattering area image). The full pre-processing workflow for Sentinel-2 band reflectance and Sentintel-1 VV and VH backscatter is fully detailed in Yuan et al. (2022).

We generated annual temporal metrics generated from all available images for Sentinel-2 10 spectral bands and from the calculated 3 optical vegetation indices and from Sentinel-1 VV, and VH backscatter, VV-coherence and the calculated RVI time series pre-processed images to train a Random Forest Regression (RFR) machine learning in R using the Caret library (Table 1). We calculated annual metrics (mean, median, standard deviation, minimum and maximum) for each dataset: Sentinel-2 and Sentinel-1 and the combination variables for each year (2017–2021). For Sentinel-2, 65 annual metrics were obtained from the 10 spectral bands and 3 vegetation indices (from the 13 variables) for each year (Figure 3).

FIGURE 3: Methodological workflow for generating the training data from the very high-resolution images and for the error estimation fractions.

We further obtained 20 annual metrics from the Sentinel-1, 4 variables (VV, VH, VV-Coherence and RVI) time series for each year. Consequently, we obtained 85 annual metrics for the combination model, which combines all metrics from Sentinel-1 and Sentinel-2 data. Finally, these metrics were then used as input predictors in the RFR to estimate fractions at each year from 2017 to 2021. We used the Caret library in R, an open-source R package that provide common machine learning tools for training and pre-processing, and evaluating to implement the RFR and predict the fractions (Siddiqui et al. 2025). Regression modelling used the lm() function in Caret R package (Lee, Wang & Leblon 2020).

Training data

To create training data for the regression analysis, we used reference cover fraction estimated from very-high resolution (VHR) images in Google Earth Pro^©. Firstly, a 10 m x 10 m grid was generated from a Sentinel-2 image dataset and intersected with transects of homogenous polygon covers of shrubs, trees and grass obtained with a GPS from a brief fieldwork that was conducted in March 2020 – April 2020 to identify shrubs from woody.

Secondly, grids that intersected with the polygons of homogenous cover of trees, shrubs and grass obtained during the fieldwork were then extracted using the Sentinel-2-pixel resolution as the basis. At each 10 m x 10 m pixel grid, fractional covers were estimated for each land cover using the VHR images available in Google Earth Pro^© for each year. This was performed by visually inspecting the respective grids in the available VHR image and consequently estimating the fraction percentage cover for trees, shrubs and grass at each target grid. This process yielded fractional cover estimates ranging from 0% to 100% for each target class at each 10 m × 10 m pixel. A total of 10 511-pixel grids of estimated fractions from the VHR images dataset not only formed the comprehensive training data for the RFR models for each year, but also served as reference data to derive error estimations for the predicted fractions. Given the open grassland nature of BNR, several months of laborious work to estimate fractions was possible to estimate 10 511-pixels per year. This was made easier by the dominance of grass pixels over trees and shrubs, which made the process to estimate these many fractions manually slightly easier, and which may not be possible for other sites with complicated savanna vegetation classes. Furthermore, as data availability is a known issue, for periods where Google Earth Pro Google Earth Pro^© images were not available, data availability was also complemented by 0.25 m colour aerial photographs from the National Geospatial Information (NGI) agency of the Department of Rural Development and Land Reform in South Africa (https://ngi.dlrrd.gov.za/). Estimation errors were then generated by carrying out an out-of-bag (OOB) process, whereby a portion of the training data was left out for model validation (Breiman 2001; Ibrahim 2023).

Random forest regression for vegetation fractions estimation

Random Forest Regression (RFR) is a non-parametric predictive model based on multiple decision trees, and has been used in various remote sensing studies (Walton 2008). The estimation of fractions is based on an RFR trained for each fraction cover and for each year (Figure 3). It is an ensemble learning method that constructs a multitude of decision trees at training time and outputs the mean prediction of the individual trees, known for its high accuracy, robustness to outliers and ability to handle high-dimensional data (Zhou, Qiu & Zhang 2023). They are available as classifications and regression trees, and are suitable to predict both discrete and continuous variables. An OOB bootstrap method (Breiman 2001) was adopted for model training and validation. Using Caret libraries in R (Siddiqui et al. 2025), 80% of the training data and predictors were randomly selected for model training, while the remaining 20% were held out for model validation of the fractions for each target class (trees, shrubs, grass) under three distinct sensor datasets configurations: Sentinel-1 (S1) Model (trained using only spatio-temporal metrics from Sentinel-1), Sentinel-2 (S2) Model (trained using only spatio-temporal metrics from Sentinel-2) and Sentinel-1 + Sentinel-2 (S1+S2) combined model (trained using a combined set of all spatio-temporal metrics from Sentinel-1 and Sentinel-2) for each year (Figure 3). Standard regression metrics such as coefficient of determination (R²), root mean squared error (RMSE), mean absolute error (MAE), prediction bias and y-intercept were calculated using the Caret package (Siddiqui et al. 2025) in R to quantify the accuracy and precision of the predicted fractions for each target class, per year. Furthermore, fractions models were generated for each year (2017–2021) for each fraction cover type, and later for fractions cover categories. With fractions categories, pixel grids were aggregated according to broader classes and fractions estimated again in R for each pixel categories, and prediction errors were generated. In the end, trees, shrubs and grass fractions were estimated for all 10 519-pixel grids, but the 10 519-pixel grids were also later generalised into a total of four broader fractions categories for each fraction cover type: 0% – 25% cover, 26% – 50% cover, 51% – 75% cover and 76% – 100% cover, and fractions cover errors were estimated accordingly at each category for each year, and for each fraction cover type.

Spatio-temporal feature selection

Feature selection is necessary to reduce noise and redundancy and can improve the model by only using the selected features in the model to reduce the model complexity (Theng & Bhoyar 2024; Zhou et al. 2018). Feature selection is also necessary for interpretation, increasing the accuracy and performance improvement.

To address the question of which multitemporal predictors are deemed important for separating between trees, shrubs and grass, all the annual spatio-temporal metrics per model were used to estimate the fractions, and an importance plot was generated, which shows the performance and variable ranking in estimating the fractions. Then RFR model in R by looking at the the %IncMSE from the predictors identifies the top metrics that contributed the most to predicting the fractions. After that, a brand-new RFR model is built this time, only allowed to see the top best predictors. The purpose is to create a more efficient model and tests if we can achieve similar (or even better) accuracy by removing the ‘noise’ of the less useful metrics. To identify the most relevant predictors, the algorithm iteratively evaluates feature importance through permutation. By randomly shuffling the values of a given feature, it measures the resulting drop in model accuracy or increase in prediction error (Mean Decrease Accuracy) (Farhana et al. 2023). Features that have not been recorded as a hit per number of iterations are rejected and removed. After that, only the spatio-temporal metrics features obtained from the most accurate model per fraction type and data model are shown to determine, which features are important when unmixing trees, shrubs or grass (Figure 6).

Ethical considerations

This article followed all ethical standards for research without direct contact with human or animal subjects.

Results

Shrubs achieved the lowest MAEs and RMSE compared to trees and grass across the three datasets (Figure 4 & Table 2). Shrub also exhibited the most MAE stability across the three data models (S1, S2 and S1+S2) and across the years (2017–2021) compared to trees and grass (Figure 4). Shrubs MAEs across the three datasets stayed in the range of 3.91% – 5.32% across the years. Compared to trees, shrubs S2 consistently produced the lowest MAEs for shrub than S1+S2 from 2019 to 2021, while S2 produced lower MAEs than S1+S2 in 2017–2018. However, differences between MAE for shrub from S2 and S1+S2 stayed less than 1% and just above 1% when comparing with S1 standalone (Table 3). The lowest MAE when estimating shrub is 3.91% obtained from both S2 and S1+S2 in both 2019 and 2017, respectively (Table 2). S1 slightly produced higher MAEs compared to S2 and S1+S2, with the highest MAE of 5.32% obtained from S1. Although shrub shows the best MAE and RMSE accuracies compared to other fractions, the correlation of shrubs was the lowest compared to trees and grass. Shrub R² values ranged between 0.61 and 0.66 for S2 and 0.59% – 0.68% for S1+S2 compared to trees R² value ranges of 0.65% – 0.72% for S2 and 0.67% – 0.76% for S1+S2. Despite shrub obtaining the most accurate MAEs compared to the two other fractions, the highest R² value achieved for shrubs is 0.68 from S1+S2 data in 2017 compared to the highest R² of 0.76% achieved from trees and 0.74% achieved from grass. The lowest correlation achieved from shrub is R² 0.50% with S1 data from 2019 to 2021. Shrub obtained positive bias values for all data models (Table 1), with S1+S2 and S2 showing the most stable positive bias values, all less than one, which shows that S2 and S1+S2 marginally only overestimated shrubs. S1 obtained positive bias values more than one, which indicates S1 moderately overestimated shrubs slightly higher than S2 and S1+S2.

FIGURE 4: Mean absolute errors for estimating trees, shrubs and grass fractions with: (a) Sentinel-2 (S2), (b) Sentinel-1 (S1) and (c) combination of Sentinel-1 + Sentinel-2 (S1+S2) models.

Trees best accuracies for all the years were achieved with S1+S2 data (Table 2). However, even though S1+S2 produced the best accuracies in the form of lower MAE and RMSE compared to S2, the differences in errors from the two datasets are not so significantly different from each other (Table 3). For example, S1+S2 produced the lowest MAE of 4.94% in 2020 for trees compared to S2 lowest MAE of 5.21% in 2019. The change differences between S1+S and S2 MAEs for estimating trees remained within < 1% for all the years (Table 3). The same trend is observed for RMSE between S1+S2 and S2. S1+S2 produced the lowest RMSE values compared to S2 for almost all the years except for 2019. S1+S2 produced the lowest RMSE of 12.33% in 2018 compared to S2 lowest RMSE of 13.14% in 2018. Once again, the difference between the RMSE for S2 and S1+S2 is marginally different with difference values staying below 1% (Table 3). The highest errors were obtained with S1 data. The highest MAE for estimating trees with S1 is 7.28% in 2017 compared to the highest MAE of 5.29% from S1+S2 and 5.73% from S2, while the highest RMSE from S1 is 17.93% compared to the highest RMSE of 13.84% and 14.34% from S1+S2 and S2, respectively (Table 2). The differences in trees MAE and RMSE from S1 and S1+S2 consistently stayed just well above 1% (Table 3). S1+S2 achieved the best correlation for trees. The highest was 0.76% from S1+S2 in 2018 and 2020. The lowest correlation (R² of 0.41%) was obtained with S1 data in 2017. Trees obtained negative bias values across the S1, S2, S1+S2 datasets, which indicate underestimation of trees. However, bias values for S1 only remained in a range of -0.02 to -0.75 (Table 2). This indicates very moderate underestimation, which shows that trees fractions estimated stayed stable when estimating with S1 compared to S2 and S1+S2. Bias values for both S2 and S1+S2 remained between -1.43 and -1.95 indicating a slight high underestimation of trees by the S2 and S1+S2 datasets compared to S1 data.

Generally, grass showed the highest MAE and RMSE errors compared to trees and shrubs; however, R² correlation values for grass remained comparable or similar to that of trees S2 and S1+S2 data (Table 1). Figure 3 shows that grass produced the highest MAEs across the 3 datasets compared to trees and shrubs, with the highest MAEs obtained from S1 data. Figure 3 further shows that MAE from grass only stayed consistent when estimating with S2 and S1+S2. Grass just like shrubs, grass best performing models were produced both by S2 and S1+S2 models. S2 produced better MAEs than S1+S2 (7.58% in 2017 and 7.21 in 2021, respectively), while S1+S2 produced better MAEs than S2 (7.76%, 7.49% and 7.60% in 2018, 2019 and 2020). Similarly, just like in trees and shrubs, the differences between the MAEs obtained from S2 and those obtained from S1+S2 remained very miniscule, less than 1%. However, S1 performance for grass was extremely poor and recorded the highest error across all the datasets. The difference between MAE from S1 compared to MAEs from S2 and S1+S2 was the highest compared to other fractions cover types, and ranged between differences in MAE of 1.88% – 5.95% for S2 and 1.42% – 6.13% for S1+S2 (Table 3). The lowest MAE for grass of 7.21% was obtained from S2, while the highest MAE of 13.71% was obtained from S1. This showed that the highest MAE obtained from grass is almost 50% higher than the highest MAEs obtained from shrubs or trees. The RMSE values also followed similar patterns, with the highest RMSE of 22.25% obtained from S1. As with shrubs, grass shows that, the addition of S1 to S2 did not always improve the results but seems to have deteriorated the errors for some years (Table 2). The highest correlation for grass was achieved with S2 data with an R² of 0.74% in 2021(Table 2). Correlation values for S2 and S1+S2 remained marginally different over the years 0.67% – 0.74% across the years. The highest R² correlation achieved for grass S1 was 0.62%, and the lowest correlation value was 0.43%. Estimating grass mostly showed minor and very marginal underestimation with bias values between (-0.03 to -0.94) across S1, S2 and S1+S2 datasets.

Overall, all sets of data S1, S2 and S1+S2 underestimated trees. Underestimation of trees is by a high margin with S2 and S1+S2 but very marginally/moderately underestimated with S1. Shrubs were over-estimated. Overestimation of shrubs is by high margin with S1 but very marginal with S1 and S1+S2. Grass is marginally underestimated by all datasets (S1, S2 and S1+S2). Highest accuracy is achieved with shrubs from both S2 and S1+S2, followed by trees with the highest accuracies for trees obtained with S1+S2. Grass obtained the highest errors for the three datasets (S1, S2, S1+S2) compared to shrub and trees. S1 always produced the highest errors compared to S2 and S1+S2 for all fractions. Correlation stayed fairly high especially for S2 and S1+S2 trees and grass. Generally, shrub albeit obtaining the lowest errors, produced much lower correlation values compared to shrubs and trees.

Impact of fractional cover proportions on the prediction accuracy

For grass cover, a consistent inverse relationship is observed between cover percentage and prediction error as shown in Figure 5. Impact of fractional cover proportions on the prediction accuracy for grass cover, a consistent inverse relationship is observed between cover percentage and prediction error (Figure 5). The highest MAE values are consistently found at 0% – 25% grass cover, as it is represented by high residuals at that grass proportion in Figure 5, with average MAE values of 7.75% for Sentinel-2, 8.61% for Sentinel-1, and 7.72% for the combined S1 + S2 model. As grass cover increases, the MAE values progressively decrease, reaching the lowest average RMSE errors of 4.38% (S2), 4.32% (S1), and 4.01% (S1+S2) errors in the 76% – 100% grass cover category (Table 2). The decreasing error at high grass proportions is also shown by the low residuals at high fraction values. S1 alone produced the highest errors albeit with insignificant differences.

FIGURE 5: Average mean absolute errors in percentage per fractions percentage categories predicted with Sentinel-2 (S2), Sentinel-1 (S1) and Sentinel-1 + Sentinel-2 (S1+S2) over the 5 years (2017–2021) for: (a) trees, (b) shrub and (c) grass.

In contrast, the shrub and trees cover prediction shows that the lowest MAE values are generally observed at low cover of 0% – 25% cover category (Figure 5). Predicting with S2 shows that lowest MAE values are produced from pixels with shrubs and trees cover percentage between 0% and 50%. Shrub shows the lowest average MAE at 0% – 25% with MAE below 5% and at 25% – 50% with MAE between 5.43% and 5.62%, with the highest average MAEs observed for pixels with shrub cover of 50% – 100% showing an MAE of above 6%. This is also shown by the low residual errors at low and intermediate shrub proportions and high residual errors at high proportions of shrubs (Figure 5).

Trees also show a similar trend as the shrub. The lowest average MAE values are generally observed at the lower tree percentage cover categories (0% – 25% with S2 at 5.45%, S1 at 5.23%, and S1+S2 at 4.68; and 26% – 50% with S2 at 5.41%, S1 at 5.52%, and S1+S2 at 5.36%), and an increase in MAE (from 4.68% – 5.62% to 6.48% – 6.83%) is observed at 51% – 75% and 76% – 100% fraction categories (Table 2). This means that estimation errors tend to increase as shrub and trees cover also increase in the pixel, peaking around the 51% – 75% category and maintaining comparable similar MAE values from there until at 76% – 100% covers (Figure 5). Trees also exhibit low residuals errors at low and intermediate proportions and high residual errors at high cover proportions (Figure 5).

Multi-temporal metric features selection

Trees’ primary predictors for S2 datasets are represented by mainly visible bands, specifically green median and green mean, followed by blue metrics (Figure 6). When estimating with S1 alone, the most significant predictors are VH median, VH mean and coherence maximum. Green median, vh median and SAVI maximum were the most dominant metrics when estimating trees with the combined S1+S2 metrics. Median green reflectance and VH backscatter were the most prevalent predictors that emerged consistently across the three data models.

FIGURE 6: Random forest feature importance for the best performing model per S1, S1+S2 and S2 datasets. Metrics with high feature importance were the most important predictor: (a) S2 Trees 2019 (b) S1 Trees 2019 (c) S1+S2 Trees 2020 (d) S2 Shrubs 2019 (e) S1 Shrubs 2018 (f) S1+S2 Shrubs 2018 (g) S2 Grass 2021 (h) S1 Grass 2019 (i) S1+S2 Grass 2021.

For shrubs, S2 blue median, maximum and mean were highly prominent when S2 model was used (Figure 6). When S1 standalone was used for shrub estimation, VV minimum, Coherence mean and VH metrics stood out. SAR-based metrics VH median, VH and Coherence mean were the top most three predictors. It shows that blue band was significant in optical data, while SAR backscatter VH and coherence metrics were the most prevalent for identifying the shrub fraction in the integrated model. For grass S2 data, blue median and green median were the most important features. When S1 was used for grass, VH mean, minimum and median ranked the highest. The integrated model for grass blue and green medians were the most top performers followed by red median and VH minimum. For grass fractions, median values for blue and green bands are the most important even when integrated with SAR data.

Vegetation fractions maps

The maps show fraction covers derived from the S1+S2 model, the model that generally produced the highest accuracies (Figure 7). Trees are represented in blue, shrubs in red and grass in green, while the dark black colours represent the masked-out soil (Figure 7). To indicate how well the fractions captured the vegetation patterns of the study area, a zoomed in section of the study area is shown. The zoomed area shows a section of the savanna biome with scattered trees species (Vachellia erioloba and Vachellia tortillas) and very few shrubs and a continuous grassland (Figure 5). The pattern for fractions derived from the S1+S2 temporal metrics match very well with the vegetation pattern observed in the very-high resolution Google Earth Pro^© images (Figure 7). On the other hand, some areas such as with rocky outcrops or dark soils and rocks or grass areas with rough soil structures in the Karoo and Grassland biomes are sometimes confused for trees and shrubs, and therefore could lead to overestimations of trees or shrubs sometimes. The fractions may show variations per year because of variability in precipitation patterns, droughts and fires.

FIGURE 7: Fractions maps for trees, shrubs and grass vegetation fraction covers derived from S1+S2 model in Benfontein Nature Reserve.

Discussion

Our study evaluated if integrating Sentinel-1 (S1) radar and Sentinel-2 (S2) improves the estimation of shrubs, grass and trees fractions in a semi-arid savanna. While the results demonstrate that integrated S1+S2 data consistently produced the highest accuracies for trees across all years, the benefits for shrubs and grass were inconsistent. Our results further showed that differences between results obtained from the integrated model S1+S2 and optical S2 remained consistently highly insignificant. In various instances (e.g. 2019–2021 for shrubs, and grass in 2017, 2021), model performances deteriorated when S1 data were added compared to the S2 standalone model. While the integration of radar and optical shows potential, the overall gains stayed marginal (typically < 1% MAE), which suggests that S2 multispectral temporal metrics do not perform too differently from S1+S2 and may also perform adequately for monitoring semi-arid savannas.

The S1+S2 model showed consistent success for trees estimation. This is likely because of the C-band SAR known sensitivity to vegetation structure, volume and texture of the woody vegetation (Urbazaev et al. 2015). Optical sensors saturate or capture the greenness of top of the canopy, while radar backscatter responds to the physical structure of the trees (Li & Guo 2016). Shrubs and grass vegetation are small in stature and can be dominated by radar speckle noise and by fluctuations in background soil moisture, which can affect their vegetation signal. In areas of low biomass, SAR is affected by contributions of signals from the ground other than the from actual vegetation, such as surface roughness, soil moisture effects (Anjitha et al. 2024; Hill et al. 2005; Mitchard et al. 2011). The contribution of the background soil factors such as soil roughness and moisture content are reported to affect the backscatter signal response in semi-arid regions with sparse vegetation cover (Eisfelder, Kuenzer & Dech 2012; Forkuor et al. 2020). Our study site exhibits mostly a scattered woody vegetation and open grassland communities with different structural traits that differ by terrain and soil properties and sometimes with the ground trampled by wildlife. Therefore, the exposure of soil and surface roughness by wildlife trampling and leftover grass tufts, which may imitate the woody signal has possibly led to the introduction of highly variable SAR signal and most likely contributed to more uncertainty in the estimation of grass.

Year-to-year variations in the integrated model performance for shrubs and trees could be explained by the annual rainfall patterns and their effect on vegetation moisture content availability. Savannas are phenologically highly variable, with wetter years exhibiting high biomass and moisture. Increases in biomass and moisture have significant effect on the SAR backscatter, which could potentially improve or complicate the backscatter signals depending on the contrasts between the vegetation and soil. In drier years where there is a weak relationship between the training data and the SAR backscatter because of interference of background soil and surface roughness, optical data may dominate more significantly than the SAR data, but wetter years may also mean that SAR is saturated or affected by the presence of moisture from the ground (Anjitha et al. 2024; Mitchard et al. 2011). Savanna’s high temporal variations are vulnerable to bias effects from pixels selected in different phenological stages (Hüttich et al. 2011b; Müller et al. 2015). South Africa experienced droughts from 2015 to 2018, which also impacted the Northern Cape and Free State, with 2017 reported as the driest year in 84 years at the time (Masupha, Moeletsi & Tsubo 2025; Theron et al. 2022). In years with close to no rainfall, the soil is completely bare most of the time; and we therefore expect the signal returns to be highly affected by soil back ground in a semi-arid area such as BNR.

In addition, the variations in availability of the time series data, which were used to calculate the annual temporal metrics could also explain the variations in the results. The number of optical data predictors already outnumbers the number of predictors available for SAR data. The integrated model used 85 predictors, while optical standalone data used 65 compared to 25 predictors used in the SAR standalone data to estimate the fractions. Times series data aggregation such as the spectral-temporal metrics are affected by variations in data availability within and between years, which may lead to the observed variations in prediction accuracies (Frantz et al. 2023; Pham et al. 2024).

Overall, our findings align with studies that have mapped vegetation in similar semi-arid environments. Studies in similar environments have confirmed potential for integrating SAR and radar to mapping vegetation classes in the semi-arid savannas, however with insignificant gains in accuracy when compared with optical standalone data (Borges et al. 2020; Higginbottom et al. 2018; Karakizi et al. 2023; Lopes et al. 2020). Higginbottom et al. (2018) integrated multiseasonal SAR and optical data to map fractional cover in South African semi-arid savannas, and only found a relative improvement of 1%. Lopes et al. (2020) mapped vegetation in West African savannas and also reported an insignificant gain in the combination of radar and optical compared to just the optical data. Overall, accuracies achieved by the S2 and S1+S2 are comparable to those of other radar-optical integration studies. Higginbottom et al. (2018) achieved R² of 0.77 and 0.74 using Landsat based multi-seasonal and single season images composites from the dry and wet seasons at multiple spatial scales (30 m – 120 m), while Urbazaev et al. (2015) study based on multiple and single-season SAR PALSAR images of 50 m resolution achieved R² values of 0.71 and 0.66 respectively. Both these results agree with our S2 and S1+S2 R² results for trees and grass. Naidoo et al. (2016) achieved a much higher accuracy of 0.80 and 0.81 with L-band ALOS PALSAR radar but used a much longer wavelength L-band and a larger spatial scale of 105 m.

For our study, the question of whether the integration of SAR (Sentinel-1) with optical (Sentinel-2) benefits the estimation of a semi-arid savanna environment is not straightforward. The benefits are only consistently observed for trees and with variations for shrubs and grass. The benefits are also very marginally beneficial (< 1%) when compared to optical data alone. Therefore, given the data processing cost, optical data alone could be sufficient in mapping vegetation fractions in semi-savannas.

Influence of percentage of vegetation cover in a pixel on the prediction error

When we generalised according to the vegetation per cent cover categories, the model’s prediction accuracy consistently exhibited a notable relationship with per cent cover. Grass accuracy increased with the increase of cover in the pixel, while the woody covers (shrub and trees) exhibited an inverse relationship with the proportions of fractions cover in the pixel. The highest improvements in MAE were observed for homogenous or near-homogenous grass cover. Grass prediction accuracy generally improved with increasing cover, indicated by a progressive decrease in MAE by more than 50% in some scenarios. This major increase in accuracy for near-pure pixels for grass models is likely attributed to the increasing dominance and homogeneity of the grass signal at higher cover percentages within the BNR, a predominantly grass-dominated savanna (Mogonong et al. 2023). Research in grassland-dominated environments have similarly suggested that simplifying vegetation into broader groups can enhance estimation, particularly for heterogeneous herbaceous fractions, using both combined and standalone optical and radar data (Hill et al. 2005). This aligns with established remote sensing principles: more uniform, dense and structurally distinct vegetation classes are typically easier to identify than diverse or ambiguous mixtures (Hill et al. 2005), as such features are better represented in estimation components.

Trees and shrubs fractional covers exhibited higher estimation error values at higher cover values. Various studies have reported poor accuracy for estimating dense woody cover in the savannas (Bucini et al. 2010; Naidoo et al. 2016). Higginbottom et al. (2018) reported a deteriorating performance for fractional cover as the woody cover approached the highest values larger than 70%, further stating that it was partly because of the rare occurrence of the class, therefore affecting the performance of the regression analysis. Given the manual nature of our reference data creation, it may be that our training data could have under-represented some of the classes. Benfontein Nature Reserve is a very sparsely vegetated reserve, where woody vegetation is very sparse and isolated, which could have led to some woody classes being rare in the training data. The sparse nature of the study site may have made it difficult to find the majority of pixels with a high per cent cover of trees and shrubs, therefore leading to the high estimation errors observed for those woody classes.

Spatio-temporal feature selection

Green band and VH were the top most features for estimating trees, which highlights the advantages of photosynthetic and structural characteristics of trees. Green band ranked high in estimation of tree fractions. Green is known for its sensitivity to chlorophyll content and tree canopy and has been used to monitor seasonal changes in chlorophyll, and often reveal temporal variations and phenological transitions in vegetation from wet to dry, and monitoring vegetation phenology (Yin et al. 2022). Healthy trees in the wet season are expected to exhibit high vegetation vigour and to reflect more green energy (Gupta & Pandey 2021; Nguy-Robertson et al. 2014). Semi-arid savannas such as our study site have a strong seasonal factor where trees are green during most of the season except during the dry season from May/June–September. Therefore, it is expected that the green band, which is sensitive to chlorophyll, is the most determining factor for estimating trees as trees are expected to stand out from grass and shrubs because of their high content of chlorophyll. The cross-polarised VH bands are sensitive to the volume characteristics of the trees and canopies and shrubs structures (Anjitha et al. 2024; Urbazaev et al. 2015).

Blue spectral bands and SAR coherence ranked the most for mapping shrubs. Coherence can assist in distinguishing between woody vegetation and surrounding vegetation such as grass. Shrubs, for example have structural properties compared to grass. Coherence may especially be useful to differentiate woody shrubs especially from non-vegetated areas or leaf-on and leaf-off seasonal changes (Khalil & Saad-ul-Haque 2018; Thijssen et al. 2025). Blue band and green ranked high for mapping grass because blue is absorbed in healthy grass with chlorophyll. Low reflectance of the blue band often indicates robust photosynthetic activity, while high reflectance suggests plant stress or the presence of bare soil or dead grass with no pigments (Tumendemberel et al. 2026). The two bands in semi-arid savannas may be essential to separate between grass species and grass from shrubs and background soils (Almalki et al. 2022; Bantelmann et al. 2024). Grass is one of the fractions in our study site affected by the soil signal, as most of the grass classes have a presence of exposed soil.

Conclusion

Our thorough analysis of the estimation errors by harnessing annual spatio-temporal data from different earth observation sensors is necessary in providing detailed information, which are needed for accurate estimation of vegetation in semi-arid savannas for evidence-based conservation efforts in protected areas.

Our analysis revealed that, in an arid savanna ecosystem, integration of radar and optical does offer benefits; however, they are marginal and not significant when compared with optical data. Consistent improvements were notable for tree estimation, and varied for shrubs and grass estimation.

Feature selection has revealed the importance of bands when modelling with S2, while VH seems to be slightly preferred in comparison to VV when estimating vegetation in a semi-arid vegetation.

Future research should optimise our results with multiresolution, multitemporal, multisensors (Landsat, Sentinel-1 and Sentinel-2) and coupled with recent synthetic training data methods to create dense time-series spanning over many years to monitor changes of vegetation dynamics in protected savanna areas.

Acknowledgements

This article is based on research originally conducted as part of Hilma S. Nghiyalwa’s doctoral thesis titled ‘Spatio-temporal mixed pixel analysis of the South African Savanna Ecosystems’, to be submitted to the Faculty of Chemistry and Geosciences, Friedrich Schiller University of Jena in 2026. The thesis is currently unpublished and not publicly available. The thesis is being supervised by Christiane Schmullius. The manuscript is being revised, and some parts are being adapted for journal publication. The authors confirm that the content has not been previously published or disseminated and complies with ethical standards for original publication.

The authors reported that they received funding from the Federal Ministry of Education and Research (BMBF) under the SPACES2 Joint Project: South Africa Land Degradation Monitor (SALDi), which may be affected by the research reported in the enclosed publication. The authors have disclosed those interests fully and have implemented an approved plan for managing any potential conflicts arising from their involvement. The terms of these funding arrangements have been reviewed and approved by the affiliated university in accordance with its policy on objectivity in research.

Hilma S. Nghiyalwa: Conceptualisation, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualisation, Writing – original draft, Writing – review & editing. Eliakim Hamunyela: Conceptualisation, Investigation, Methodology, Supervision, Writing – review & editing. Tshililo Ramaswiela: Writing – review & editing. Jussi Baade: Funding acquisition, Project administration, Resources, Writing – review & editing. Christiane Schmullius: Conceptualisation, Funding acquisition, Investigation, Project administration, Resources, Supervision, Writing – review & editing. All authors reviewed the article, contributed to the discussion of results, approved the final version for submission and publication, and take responsibility for the integrity of its findings.

This work was supported by the Federal Ministry of Education and Research (BMBF) under the SPACES2 Joint Project: South Africa Land Degradation Monitor (SALDi) (BMBF grant no.: 01LL1701 A–D). HSN PhD scholarship is made possible by Deutscher Akademischer Austauschdienst: DAAD Ref No. SPACES II.2 CaBuDe 57531823. The University of Namibia Staff Development (SDF) programme made it possible for HSN to undertake a Doctorate in Germany.

The Southern Africa Environmental Observation Network (SAEON) office in Kimberley is acknowledged for providing the biomes data, and in assisting with the logistics for fieldwork campaign conducted between March 2022 and April 2022.

The data from the study cannot be made available, because ethical clearance and participant consent for the study did not include consent to share raw data with third parties.

The views and opinions expressed in this article are those of the authors and are the product of professional research. They do not necessarily reflect the official policy or position of any affiliated institution, funder, agency or that of the publisher. The authors are responsible for this article’s results, findings and content.

References