2.1.

Study Area

The Klipriviersberg Nature Reserve (KNR) located in Johannesburg, South Africa, was used for the study (Fig. 1). The reserve was declared a nature conservation area in 1984 and covers $\sim 651\mathrm{ha}$. In general, the vegetation types in the reserve include Andesite Mountain Bushveld and Clay Grassland, which are associated with a savanna environment.⁴⁵ The altitude of the area ranges between 1540 m in the south and 1790 m in the north, with a mean altitude of 1653 m. The mean annual rainfall around KNR ranges from 624 to 802 mm promoting foliage and canopy cover in wet periods. The wet season, which is largely associated with photosynthetically active plants, runs from November to March and the dry season associated with low foliage occurs between May and October. The mean annual temperature ranges between 17°C and 26°C in summer and 5°C and 7°C in winter.⁴⁶ The geology types found in the area, which lead to the floristic structure of the reserve, include quartzites, conglomerates, and dolomites.⁴⁷

Fig. 1

Klipriviersberg Nature Reserve and the distribution of sampling plots used in the study. The background image was derived from a WorldView-2 image in false color composite (red, band 6; blue, band 4; green, band 5).

2.2.

Field Data

In this study, 240 points distributed at $\sim 170\mathrm{m}$ intervals in the north–south and east–west directions were generated in ArcGIS (ESRI^® ArcGIS 10.6, Redlands, California). The point coverage was exported into a global positioning system (Garmin, GPSMAP^® 64, Kansas) and located in the field. Field surveys were conducted from May to June, 2017, representing the dry period in the study area.⁴⁶ A buffer with a 20-m radius was created around each point, making a plot; this size was specified to accommodate multiple pixels of the images used in the study (WorldView-2, SPOT-6, and Sentinel-2A). Circular plots were preferred over rectangular plots as they require only a single control point at the plot center.⁴⁸ Furthermore, circular plots were favored instead of angular shapes, since circular canopy shapes are more commonly witnessed in a natural vegetation environment. In each plot, plant species with height $\ge 2\mathrm{m}$ and land covers were recorded, with the record showing a minimum of one and a maximum of nine different species per plot. Overall, a total of 27 different species and three land cover types (grassland, bareland, and shrubs) were recorded in all plots. Additional structural attributes, such as species canopy size and species richness, were recorded in each plot. These attributes were used to confirm the assignment of a pixel to a class in an instance of mixed-pixel phenomenon. Accordingly, a pixel was allocated to species that had the dominant canopy size falling within that pixel. In the case of multiple plants with relatively small canopy sizes, the species with the most occurrence determined the classification of that pixel.

2.3.

Remote Sensing Data and Preprocessing

WorldView-2, SPOT-6, and Sentinel-2A were acquired on May 17, June 5, and June 10, respectively, coinciding with the time of the field surveys. WorldView-2 image (DigitalGlobe)⁴⁹ has eight multispectral bands in the 0.40 to $1.04\mu \mathrm{m}$ region and a panchromatic band in the 0.45 to $0.80\mu \mathrm{m}$ (Fig. 2) measured at 1.8 and 0.46 m spatial resolutions, respectively. SPOT-6 image was sourced from the South African National Space Agency (SANSA). The imagery has four multispectral bands in the 0.45 to $0.89\mu \mathrm{m}$ and a panchromatic band in the 0.45 to $0.75\mu \mathrm{m}$ (Fig. 2) measured at 6 and 1.5 m spatial resolutions, respectively. Sentinel-2A image was downloaded from the European Space Agency Data Hub.⁵⁰ Sentinel-2 has 13 multispectral bands (Fig. 2) with four bands (0.49 to $0.84\mu \mathrm{m}$) measured at 10 m spatial resolution and six bands measured at 20 m spatial resolutions. Prior to classification, the three images underwent atmospheric correction to ensure high signal-to-noise ratio. A comparison of atmospheric correction methods between dark object subtraction (DOS)⁵¹ and fast line-of-sight atmospheric analysis of hypercubes⁵² showed strong similarities between the two approaches (Pearson’s correlation, $r=0.95$). Therefore, we applied DOS to all individual bands of each imagery in ENVI 5.3 (©2015 Exelis Visual Information Solution Inc., Boulder, Colorado). Coastal bands were excluded due to the relative sensitivity of those bands to atmospheric interferences.⁵³ The remaining bands (7 for WorldView-2, 4 for SPOT-6, and 10 for Sentinel-2A, Table 1) were subsequently pansharpened. Notably this study utilized the Gram–Schmidt algorithm⁵⁴ which maximizes image sharpness and minimizes color distortions.

Fig. 2

Spectral profiles extracted from satellite images used in the study.

Table 1

Comparison of spectral and spatial profiles of satellite images used in the study.

2.4.

Training and Classification of Remotely Sensed Data

Training of 27 unique woody plant species as well as grassland, shrubs, and bareland classes was performed on the three satellite images (WorldView-2, SPOT-6, and Sentinel-2A) separately. It should be noted that as the spatial resolution becomes coarser, individual pixels are less likely to capture small features resulting in mixed pixel phenomenon.^39,55,56 This study adopted the nearest-neighbor resampling technique on SPOT-6 and Sentinel-2A to resample the pixels to the size of the WorldView-2 image (0.5 m). Nearest-neighbor resampling technique was chosen because it does not alter values in the output raster data set and therefore appropriate for categorical data classification.⁵⁷ Resampling to 0.5 m ensured exact subdivision of Sentinel-2A and SPOT-6 images avoiding the mixing of information between neighboring pixels of the original resolutions. By superimposing the three images, it was also confirmed that the offset in the pixel locations never exceeded 0.03 m avoiding the spill effect of information into neighboring pixels. Similar sampling points were used for the three images to ensure direct comparability between results. A total of 8011 points representing 27 woody plant species, grassland, shrubs, and bareland were digitized inside the 240 plots on the three satellite images separately. Digitizing of points was guided by field surveys in which a local Cartesian coordinate system was used to locate the species. Finally, points were split into two portions of which 30% ($n=2408$) were allocated to training–classification and 70% ($n=5603$) to evaluate the accuracy of the classification. The spatial distribution of the training samples was taken into consideration when selecting the training samples. The species along with the proportions allocated to the training and testing of the classifications are given in Table 2.

Table 2

Illustration of the number of woody plant species used for training and evaluation of classification.

Two machine learning classification algorithms utilized in this study are Random Forest (RF) and Support Vector Machine (SVM). These classification algorithms were implemented using the Caret package⁵⁸ for R language.⁵⁹ The RF classifier is an ensemble machine learning approach, which utilizes bootstrap sampling to build multiple decision tree models.⁶⁰ The RF method was selected for this study due to the following reasons: (i) it can analyze large datasets, (ii) it is free from normal distribution assumptions, and (iii) it is powerful when dealing with outliers in the dataset.⁶¹ Internally, the RF uses two-thirds of the data (in-bag) for training the classification model and the remaining one-third, which is referred to as out-of-bag data, to evaluate the accuracy of the trained model.⁶¹ RF classifier utilizes ntrees (number of classifications trees) and mtry (a number of predicting variables) to generate a prediction model.⁶⁰ In this study, a 10-fold cross-validation analysis which was repeated 10 times was used to determine the optimal parameters. The explanatory power of the input variables (multispectral bands) was quantified to rank the importance of each band for the classification accuracy.

The SVM approach classifies features (reflectance of different bands) by identifying optimal decision (separation) boundary that maximizes the margin between two classes.⁶² The SVM, such as the RF, does not require the data to have a normal distribution,⁶³ and it performs well when using high dimensional and complex data. This study used a nonlinear SVM technique, i.e., radial basis function kernel that accommodates linear and nonlinear relationships between a response and a predictor⁶² customized for R.⁶⁴ The SVM classifier requires the specification of two parameters to balance the accuracy and reliability of the classification.⁶³ These parameters include cost factor ($C$) and gamma ($\gamma$). The $C$ factor relates to the penalty (cost) of misclassification error, and $\gamma$ determines the influence of a training sample to capture the complexity in the data.⁶² $C$ and $\gamma$ were also determined by running 10-fold cross-validation which was repeated 10 times similar to the approach applied for selecting optimal parameters of the RF. Training and classification of the images were performed on 30% ($n=2408$) of the data using each satellite image.

2.5.

Accuracy Assessment

Classification results derived from the remotely sensed data (WorldView-2, SPOT-6, and Senetinel-2A) were assessed on 70% ($n=5603$) of the data. Although the RF has an internal evaluation system, we believe that the use of such a large independent sample dataset provides a more convincing evaluation of the classification. An error matrix was used in the study that uses overall accuracy, producer’s accuracy, and user’s accuracies statistics.⁶⁵ The user’s accuracy indicates the probability that classified woody species and land cover types on the map represent the same category on the ground. This accuracy is thus calculated as the number of true observations of a class divided by the number of predicted observations. The producer’s accuracy indicates the probability of a reference class being classified correctly, and it is calculated as the number of true observations of a class divided by the number of true reference observations of that class. Kappa coefficient statistic [Eq. (1)] was applied to evaluate the quality of classified imagery.⁶⁶ The kappa statistic is used to control the cases which might have been correctly classified by chance:

McNemar’s test, which is a nonparametric and standardized normal test based on confusion matrices in a $2\times 2$ dimension,⁶⁷ was also applied in this study. The McNemar’s test computed using Eq. (2) determines the binary distinction between correct and incorrect class designation by different images (Worldview-2, SPOT-6, and Sentinel-2A) using RF and SVM classifiers:

3.1.

Classification Accuracies Derived from WorldView-2, SPOT-6, and Sentinel-2A

Figure 3 shows visual comparisons of woody species and coexisting land cover types using Worldview-2, SPOT-6, and Sentinel-2A images classified using RF and SVM classifiers. Overall classification accuracies of species and coexisting land covers derived from the three images showed the best overall accuracy of 65% for WorldView-2 image, followed by Sentinel-2A and then SPOT-6 using the RF classification (Fig. 4). Similarly, the SVM returned classification accuracies derived from the three images in the same ranking order as the RF, although the SVM had a lower accuracy for each image. Kappa coefficient statistics derived from the same three images showed that WorldView-2 had the highest kappa coefficient value of 0.63 followed by Sentinel-2A and then SPOT-6 using the RF classifier (Fig. 4). The kappa values of the three images using the SVM classifier had the same ranking order as the RF classifier (Fig. 4).

Fig. 3

Visual illustration on the performances of RF classification using (a) WorldView-2, (b) SPOT-6, and (c) Sentinel-2A; SVM classification using (d) WorldView-2, (e) SPOT-6, and (f) Sentinel-2A images in classifying woody plant species and coexisting land cover types.

Fig. 4

Overall accuracies and kappa coefficient statistics for WorldView-2, Sentinel-2A, and SPOT-6 using RF and SVM classifiers.

Results from the McNemar’s test revealed statistically significant difference between the accuracies of WorldView-2 (${\chi }^2=9.15$; $p=0.002$) and SPOT-6 (${\chi }^2=6.34$; $p=0.007$) when RF classification was used. The difference between Sentinel-2A (${\chi }^2=4.48$; $p=0.023$) and WorldView-2 images (${\chi }^2=5.36$; $p=0.034$) also were significant using the RF classifier. Such statistically significant difference was also observed between Sentinel-2A (${\chi }^2=4.75$; $p=0.042$) and SPOT-6 (${\chi }^2=4.97$; $p=0.048$). The SVM-based classification resulted in statistically significant difference between the accuracies of WorldView-2 (${\chi }^2=7.15$; $p=0.03$) and SPOT-6 (${\chi }^2=8.34$; $p=0.04$). Similarly, the differences between Sentinel-2A (${\chi }^2=4.48$; $p=0.023$) and WorldView-2 images (${\chi }^2=5.36$; $p=0.034$) were statistically significant using the SVM classifier. However, the differences between Sentinel-2A (${\chi }^2=0.75$; $p=0.142$) and SPOT-6 (${\chi }^2=1.97$; $p=0.248$) were statistically not significant using SVM.

Producer’s and user’s accuracies of individual plant species are shown in Fig. 5. The producer’s accuracies ranged between 27% (Cordyline australis) and 83% (grassland) for different species across the three images when using the RF classifier [Fig. 5(a)]. When using WorldView-2 image and RF, 20 species and coexisting land covers had producer’s accuracy exceeding 60% with 10 of them having 70% or higher accuracies. Sentinel-2A and RF combination yielded producer’s accuracies of $>60\%$ for 15 species and coexisting land covers, while the accuracy exceeded 70% for seven species. Significantly, fewer species and coexisting land covers had good accuracies when SPOT-6 was used with only four species having $>60\%$ accuracies. The SVM classifier yielded producer’s accuracies varying between 14% (Heteromorpha arborescens) and 94% (Melia azedarach) across the three images [Fig. 5(b)]. Specifically, WorldView-2 image returned producer’s accuracies exceeding 60% for 16 species and exceeding 70% for seven species. The combination of Sentinel-2A and SVM resulted in quite low producer’s accuracy with only eight species estimated at $>60\%$ accuracy while only two species scoring $>70\%$. SPOT-6 and SVM combination fared better than SPOT-6 and RF combination but only marginally with six and three species having $>60\%$ and 70% accuracies, respectively. The user’s accuracies ranged between 31% (grassland) and 95% (Acacia caffra) across the three images using RF [Fig. 5(c)] and from 11% (H. arborescens) to 92% (A. caffra) using SVM classifier [Fig. 5(d)]. Seventeen species and coexisting land covers had user’s accuracy $>60\%$ and seven of those with $>70\%$ using WorldView-2 and RF combination [Fig. 5(c)]. For Sentinel-2A and RF combination, 13 species had user’s accuracies exceeding 60% with five of those species having $>70\%$ accuracies. Using the SVM classifier, the numbers of species identified at accuracies $>60\%$ were significantly lower for the three images than what were obtained using the RF classifier; with the lowest performance observed for Sentinel-2A image (Fig. 5).

Fig. 5

Accuracies of identifying individual species using WorldView-2, Sentinel-2A, and SPOT-6; (a) RF producer’s accuracy, (b) SVM producer’s accuracy, (c) RF user’s accuracy, and (d) SVM user’s accuracy. Species names represented by the two-or three-letter codes are given in Table 2.

Using the producer’s accuracy, which evaluates classification quality against the reference (truth), we can compare the relative performance of each image against the other two in identifying species (Fig. 6). The WorldView-2 was better than Sentinel-2A for 16 species and coexisting land covers [Fig. 6(a)] with the relative improvement in producer’s accuracy ranging between 1% and 38% for RF and 1% and 60% for SVM. The improvements exceeded 10% for 13 species using both classifiers. WorldView-2’s improvement over SPOT-6 was even more for 26 and 21 species and coexisting land covers using RF and SVM, respectively [Fig. 6b], with the improvement exceeding 10% for 15 or 14 species and coexisting land covers for the two classifiers. Notably, Sentinel-2A performed better than WorldView-2 for 13 (RF) and 11 (SVM) species with the improvement exceeding 10% for six species [Fig. 6(c)]. Sentinel-2A’s better performance over SPOT-6 was observed for 18 and 14 species using the RF and SVM classifiers, respectively [Fig. 6(d)]. The relative performance of SPOT-6 over WorldView-2 was noted to be better in four or nine species depending on the classifier [Fig. 6(e)]. SPOT-6 was advantageous over Sentinel-2A for 12 or more species and coexisting land covers; this was the case more using the SVM than the RF classifier [Fig. 6(f)].

Fig. 6

Relative producer’s accuracy of an image over the other images in identifying species types. Worldview-2 (WV), SPOT, and Sentinel-2A multispectral image (S-MSI) in the $y$ axis represent satellite images. Species names represented by the two-or three-letter codes are given in Table 2.

3.2.

Comparison of Images Based on Confusions

It is useful to evaluate the level of confusion of a species against other species in a localized area with a diverse vegetation composition. Logically, the image that yields the smallest amount of confusion among coexisting species is considered desirable. Detailed confusion matrices using RF and SVM classification types are given in Table 3 in the Appendix A. Figure 7 provides the count of other species and coexisting land covers against which a species is confused. The RF classification clearly showed that the WorldView-2 identified nearly each of the species with the least number of confusions with other species. Using this image, a species is confused on average with 11 other species or land cover types, while 12 species were confused with $<10$ species and coexisting land covers [Fig. 7(a)]. In contrast, Sentinel-2A confused a species with an average of 18 species and coexisting land covers [Fig. 7(a)]. Despite an overall weaker performance of Sentinel-2A, it had comparable confusion level with WorldView-2 for certain species (e.g., Acacia karro and Dispyros natalensis) and fared better than WorldView-2 for three species (Afrocanthium mundianum, C. australis, and Zanthoxylum capense). SPOT-6 created considerable confusion in identifying each species at an average of 24 species and coexisting land covers confused with each species. The classification using SVM reduced the discrepancy among the three images, compared with the RF classifier [Fig. 7(b)]. In particular, the average number of species and coexisting land covers confused with any species was equal for WorldView-2 and Sentinel-2A at 16 while SPOT-6 confused a species with an average number of 21 species, which is an improvement from 24 species and coexisting land covers using the RF classifier. Notably, Sentinel-2A had a lower confusion rate than WorldView-2 for 16 species [Fig. 7(b)]—an improvement from only three species using the RF classifier [Fig. 7(a)].

Fig. 7

Number of species confused against a given species for (a) RF and (b) SVM classifiers.

While the above comparison focused on confusion based on the number of different species, it is important to compare images using the number of samples of each species that contributed significantly to inaccuracies. We illustrate this using a select species for both RF and SVM classifiers, and exhaustive confusions are given in Table 4 in the Appendix B. To balance the comparison across accuracies, we selected species from three producer’s accuracy categories including $<60\%$, 60% to 70%, and $>75\%$. Since the intention was to compare images, the selected species in each category needed to satisfy the criterion using at least two images. The selection was made based on the results presented in Figs. 8(a) and 8(b), for the RF- and SVM-based producer’s accuracies, respectively. Figure 8 illustrates three selected species including A. caffra ($<60\%$), A. karoo (60% to 70%), and Dombeya rotundifolia (60% to 70%), and how their accuracies were affected largely by few species. Notable similarities were observed in terms of species type that contributed to inaccuracies of identifying each species; for instance, A. caffra was confused mainly with A. karro, Afrocanthium mundianum, C. australis, and grassland when using RF [Fig. 8(a)] and SVM [Fig. 8(b)]. It is also noteworthy to mention the agreement between the two classifiers in identifying the images with comparable confusions. For example, confusion of A. caffra with A. karro and A. mundianum was noted when Sentinel-2A and SPOT-6 were used exploiting RF and SVM classifiers. Although comparable images did not match consistently, certain similar observations can be seen in the identification of species. This can be noted for A. karoo whose inaccuracy was compromised by A. caffra, Celtis australis, and Ehretia rigida for both classifiers [Figs. 8(c) and 8(d)] while D. rotundifolia was confused mainly with A. caffra and Ehretia rigida for both classifiers [Figs. 8(e) and 8(f)].

Fig. 8

Comparison of images based on number of samples contributing to confusions of three selected species: (a), (b) A. caffra, (c), (d) A. karro, and (e), (f) D. rotundifolia and using the RF and SVM classifiers.

3.3.

Band Importance of WorldView-2, SPOT-6, and Sentinel-2A Images

Variable importance, which measures the percentage that the prediction error increases when a predictor variable is removed, was computed for each image using both classifiers in an attempt to see the general trend in the contribution of individual bands [Figs. 9(a)–9(c)]. In general, there was a strong similarity among the three images in terms of the important regions of the electromagnetic spectrum for RF and SVM classifiers. Infrared range bands had the most contributions to the classifications for the three images ($>50\%$) using both RF and SVM classifiers. For example, the near-infrared (NIR) band alone contributed to 65% or greater accuracy in all the images [Figs. 9(a)–9(c)]. There were also similarities among the three images in the contributions made by the blue and green bands each of which contributed to $<40\%$ accuracy. Further similarities can be noted between WorldView-2 and Sentinel-2A that have more spectral bands in the infrared wavelength regions. For example, each of NIR2 and red edge bands of WorldView-2 and VRE and SWIR bands of Sentinel-2A contributed $>65\%$ of the classifications. The SWIR1 and SWR2 available only in Sentinel-2A also made significant contributions to the classifications using both classifiers [Fig. 9(b)].

Fig. 9

The relative importance of the multispectral bands in classifying woody plant species using RF and SVM derived from (a) WorldView-2, (b) Sentinel-2A, and (c) SPOT-6 images; NIR, near-infrared; VRE, vegetation red edge; and SWIR, shortwave infrared band.

4.1.

Performances of WorldView-2, SPOT-6, and Sentinel-2A

Numerous studies have applied remotely sensed data with relatively low spectral and spatial characteristics failed to capture the true extent of localized plant species diversity.^34,41,68 Therefore, this study utilized WorldView-2, SPOT-6, and Sentinel-2A satellite images with improved spatial or spectral characteristics to capture several woody plant species and generic land cover types ($n=30$) in a localized savanna environment during a dry season. WorldView-2 image yielded the highest overall classification accuracy (65%) using the RF classifier compared to other images (Fig. 4). One advantage of WorldView-2 data is the availability of a significant number of spectral bands present within a narrow spectral range.⁶⁹ Such property allows for improved discrimination of subtle differences among species in dry periods associated with low foliage, which may obscure characteristic differences between species.⁴⁰ The combination of improved spatial and spectral characteristics offered by WorldView-2 image reduces the mixed pixel problem inherent in coarser spatial resolution and low spectral resolution images.^53,55 Variability in overall classification results of the three images might also have been influenced by different acquisition times of the three images. Although the images were all collected in dry period, even minor variations in weather conditions on different dates and times can result in variability of reflected radiation captured by the satellite images⁷⁰ thus contributing to a level of confusion in species identification.

Sentinel-2A image returned the second-best accuracy of discriminating the woody plant species and coexisting land covers in this study (overall accuracy = 59%). This accuracy is encouraging, given the relatively coarse spatial resolution of Sentinel-2A ($\ge 10\mathrm{m}$) compared with WorldView-2 (65%) and considering the large number of species targeted in the study. Our study moderately agrees with Ref. 71 which used Sentinel-2A image to map 24 woody plant species and reported accuracies $>65\%$. A key advantage of Sentinel-2 over WorldView-2 is that it has spectral bands in the SWIR region of the electromagnetic spectrum. Therefore, this advantage has compensated for the loss in the spatial resolution of the image. SPOT-6 imagery achieved the lowest accuracy (52%) of the three images used in the study. Although SPOT-6 imagery has relatively better spatial information compared with Sentinel-2A, it lacks detailed spectral bands particularly in the infrared regions that are suitable for differentiating plant species.⁷² A similar inferior performance of SPOT image was reported by Ref. 34, which compared the accuracies of Landsat (50%) and SPOT-5 (30%) in the classification of plants in a savanna region in Australia. The producer’s and user’s accuracies generally corresponded with the overall accuracies in showing the superiority of WorldView-2 image over the others (Fig. 5). It is important to note that the above observations were somewhat similar for the RF and SVM classifiers, indicating the reliability of the findings in ranking the three images irrespective of a classification approach.

A comparison among the performance of the three images using relative improvement of producer’s accuracies clearly showed the superiority of WorldView-2 image followed by Sentinel-2A and SPOT-6 as shown in Fig. 6. Specifically, the improvement of WorldView-2 over the other images in terms of producer’s accuracy was evident for many species and coexisting land covers. The advantage of WorldView-2 over Sentinel-2A was also reported by Ref. 73, which estimated deciduous small spiral thin leaf oak plants ($n=13$). It is noteworthy to mention the preference of Sentinel-2A over WorldView-2 for many species; this could be attributed to the image’s spectral superiority. The vegetation type in the study area is dominated by narrow-leaved plant species (Table 2) and limited chlorophyll content due to the dry season conditions under which the data were collected. The combination of these factors necessitates the use of imagery that has multiple and narrow bands such as those found in Sentinel-2A.³² The advantage of WorldView-2 over SPOT-6 agrees with Ref. 74 which reported better performance of WorldView-2 compared with SPOT-5 image in classifying plant species in a savanna vegetation environment, although they targeted fewer species ($n=5$). Comparatively, SPOT-6 performed better than the other images for quite a few species [Figs. 6(e) and 6(f)], understandably due to the lower spectral qualities it possesses compared with the other images.

4.2.

Confusion Levels of Images

In most cases, the classification using WorldView-2 (as opposed to the other images) resulted in each species confused with the least number of different species particularly when the RF classifier was used [Fig. 6(a)]. The weakness of Sentinel-2A in this regard can be attributed to the physiological characteristics of the vegetation in the study area of which $>90\%$ of species type are narrow-leaved plants such as H. arborescens, A. karro, C. australis, etc. Narrow-leaved plant species are difficult to discriminate using images with coarse spatial resolution images^34,71 such as the Sentinel-2A (10 to 60 m, spatial resolution) used in this study. Although SPOT-6 has a relatively high spatial resolution, it still resulted in confusion among several species as well as coexisting land cover types. It is important to note that deciduous species identified in the study area lose their leaves during the winter period when the data for this study were collected.⁴⁵ The limited spectral capability offered by SPOT-6 is unlikely to differentiate species accurately. Although SPOT-6 imagery has better spatial information than Sentinel-2A, the results show that spatial information alone is insufficient to discriminate woody plant species accurately.

It is worth noting that the SVM classifier applied to Sentinel-2A resulted in lower misclassification of each species with other species when compared with the RF classifier. In fact, the improvement in general showed similarity of the image with WorldView-2. This can be attributed to the spectral advantages of Sentinel-2A which divides the red-edge region into four bands.⁷⁵ The performance of Sentinel-2A comparable to WorldView-2’s is encouraging since the former is widely and publicly available at no cost.

In identifying a species in a highly diverse savanna ecosystem, it is critical to pinpoint species that have major contributions to misclassifications of a target species. The findings of this study showed that few species created most of the confusions in the classification of the 27 woody plant species, and these confusions are evidenced by at least two of the three images (Fig. 8). Such confusions can be attributed to the similarity in foliage (leaf) characteristics of the species inducing somewhat similar spectral responses captured by the three sensors in the study.⁷³ This suggests, among others, the need for advanced remote sensing systems with spectral and spatial characteristics better than those used in this study.

4.3.

Comparison of Images Based on Band Importance

Comparison of variable importance findings showed that the highest contributions from all the three images in the analysis (WorldView-2, SPOT-6, and Sentinel-2A) and classifiers (RF and SVM) were made by the infrared bands (Fig. 9). This finding is in agreement with Refs. 76 to 79 who discriminated narrow-leaved plant species in a savanna environment. The yellow band of WorldView-2 image performed also performed well comparatively to infrared bands. This is expected since the band is useful in detecting woody plant species with gray to yellow coloring often present during dry periods in savanna environments.⁶⁸ The significant contribution of the red-edge band available in WorldView-2 and Sentinel-2A is related to the sensitivity of the band to chlorophyll variations of even small narrow-leaved plants.⁸⁰ The availability and significant contributions of vegetation red edge bands in Sentinel-2A clearly places it at advantage over Landsat imagery, which largely shares similar spatial and spectral characteristics with Sentinel-2A.