Introduction

Wetlands harbour a wide range of biodiversity and play a crucial role in hydrological and biogeochemical cycles (Kingsford et al. 2016). For instance, wetlands improve water quality, attenuate floods, regulate stream flow, trap sediment, sequester carbon, control erosion, and serve as a marine and terrestrial species habitat. Thus, wetlands have the highest value per hectare among ecosystems (Xu et al. 2019). In South Africa, 48% of wetland ecosystems are classified as critically endangered, 12% as endangered and 5% as vulnerable (Job et al. 2018). Over half of the South African wetlands have already been degraded (Dumakude and Graham 2017). Moreover, fine-scale data, such as species occurrence and distribution, has not been captured for most inland South African wetlands (Van Deventer et al. 2018). Yet, identifying wetlands and their functions is essential to understanding, rehabilitating, and conserving wetlands and protecting the ecosystems that depend on them (Rebelo et al. 2009; Bonthuys 2020). The physical indicators for monitoring wetlands include extent and diversity, landscape pattern, hydroperiod and chemical contaminants. The biological indicators include vegetation species composition, greenness and percentage cover (Adamus 1992).

Traditional methods for mapping extent and diversity, such as floristic mapping, require rigorous fieldwork and manual estimation of proportional cover for each species of interest (Sharp and Keddy 1986; Wijana and Setiawan 2020). Additionally, the physical characteristics of wetlands make fieldwork very challenging, expensive and imprecise (Everitt et al. 2002; Hemati et al. 2023). Thus, only small areas can be mapped precisely, and extrapolation could be erroneous (Millar 1973; Harvey and Hill 2001; Morrison et al. 2020; Morrison 2021). Remote sensing technology offers a less intrusive (Rundquist et al. 2001; Lane et al. 2014, 2015) and more scalable approach (Rebelo et al. 2009; Adam et al. 2010; Yan et al. 2017; Moity et al. 2019) for mapping wetland plant species. Also, repeated coverage can facilitate the thorough detection of temporal changes in wetlands (Ramsey Elijah et al. 2009; Sica et al. 2016; Jia et al. 2020; Hasan et al. 2023), although the mapping capability of remote sensing technologies can be hampered by coarse spatial resolution (De Roeck et al. 2007).

Conversely, several characteristics of wetlands make them inherently difficult to monitor remotely. For example, wetlands are generally highly dynamic, and their spectral responses frequently change (Karabulut 2018; Montgomery et al. 2021). Studies have shown that the spectral responses of individual wetland species can vary significantly even within the same growing season (Gallant 2015). Moreover, steep environmental gradients within and around wetlands can result in narrow transition areas between ecological systems that are sometimes smaller than the spatial resolution of most sensors (Harvey and Hill 2001; Adam et al. 2010). Also, sensor resolution can limit understanding of the interactions between different ecological systems, making managing and protecting wetland ecosystems harder.

The advent of low-cost data collection platforms such as Unmanned Aerial Vehicles (UAVs) has made it affordable to gather high-resolution remote sensing data over specific areas at specified times (Tu et al. 2019; Wijesingha 2020; van Blerk et al. 2022). Recent developments in technology and data processing for UAVs have provided novel opportunities to resolve some of the impediments to wetland studies, such as the difficulty of field surveys, coarse satellite resolutions and high costs of piloted aerial photography (Dronova et al. 2021). In fact, UAVs have several advantages over satellites. For instance, UAVs are not affected by cloud cover. Also, UAV payloads are interchangeable, and end-users can control data acquisition parameters such as spatial resolution, frequency of data collection and view angles (Alvarez-Vanhard et al. 2021). On the other hand, UAV data acquisition is affected by wind and precipitation, requires a trained operator, accurate ground control, and it is highly regulated by legislation (Jeanneret and Rambaldi 2016; Stöcker et al. 2017; Assmann et al. 2019) Still, the use if UAVs in wetland studies is gaining traction. A review study found that several UAV-based studies were conducted in the United States, China and Europe with emphasis on riverine and floodplain, mangrove and peatland (Dronova et al. 2021). Only one of the 122 papers reviewed was a South African case study focused on wetland delineation (Boon et al. 2016).

Fifteen papers reviewed by Dronova et al. (2021) focused on vegetation inventory, and most of them utilized Object-based Image Analysis (OBIA) using target characteristics such as canopy diameter, shape, height and distribution pattern. Unlike pixel-based classification methods, OBIA aggregates pixels into spectrally similar objects using segmentation algorithms prior to classifying the objects (Tian et al. 2020). The spatial and textural characteristics of the objects can complement spectral data and improve classification accuracy (Whiteside and Ahmad 2005; Liu and Xia 2010; Giglio et al. 2019; Du et al. 2021). The key drawback with OBIA is the difficulty of choosing segmentation parameters. During segmentation, individual pixels are grouped into segments (objects) based on specific criteria, including the uniformity within each segment, the capacity to distinguish them from neighboring elements (dissimilarity), and the consistency of their shapes (Veljanovski et al. 2011; Cheng and Han 2016). The accuracy of the classification process hinges on the quality of the segmentation process (Veljanovski et al. 2011; Cheng and Han 2016; Gibril et al. 2020). Determining the essential segmentation criteria often involves a process of trial and error, and these criteria can multiply and contradict one another, ultimately leading to either an excessive or insufficient level of segmentation (Liu and Xia 2010; Veljanovski et al. 2011). Moreover, feature selection can sometimes have an adverse effect on the classification process (Ma et al. 2017). Also, the consistency and repeatability of the segmentation and classification processes is still contentious (Veljanovski et al. 2011). In light of these challenges, pixel-based approaches are still more common (Nezami et al. 2020; Allen et al. 2021; Mirmazloumi et al. 2021; van Blerk et al. 2022; Zhu et al. 2022; Windle et al. 2023). Furthermore, some studies have explored both OBIA and pixel-based methods and recommended the latter (Giglio et al. 2019; Abeysinghe et al. 2019).

These include the emerging use of deep learning algorithms in wetland studies (Sun et al. 2021; Higgisson et al. 2021; Yang et al. 2022). Deep learning enables the creation of trainable models that can learn data representations with multiple levels of abstraction (Janiesch et al. 2021). Deep learning possesses a powerful capacity to comprehend intricate training samples and exhibit strong robustness when classifying complex features, like wetland landscapes in remote sensing (RS) images (Jafarzadeh et al. 2022). Some studies have found deep learning algorithms can outperform common shallow learning algorithms (Rezaee et al. 2018; DeLancey et al. 2019) and others have found contradicting results (Islam et al. 2023). However, deep learning is most pertinent when dealing with large, high dimensional data. Shallow machine learning algorithms can produce better results than deep learning algorithms when utilizing low-dimensional and low training data (Janiesch et al. 2021). The popular shallow learning non-parametric machine learning classifiers include Random Forest (RF), Support Vector Machines (SVM), and K Nearest Neighbor (KNN) (Belgiu and Drăguţ 2016; Chirici et al. 2016; Sheykhmousa et al. 2020). KNN is a non-parametric machine learning algorithm that classifies individual data points based on proximity to data points with known classes (Mucherino et al. 2009a). The algorithm compares a number (k) of the closest training data points in feature space to a new data point and then classifies the individual data point based on the most common class among the k-nearest neighbours. Support Vector Machines utilize an optimal line of separation, a hyperplane, to classify data points by maximizing the margin between the class boundaries (Mucherino et al. 2009b). SVMs are popular because of their capacity to control the trade-off between maximizing the margin and classification errors. Random Forest is an ensemble learning algorithm aggregating multiple decision trees to predict the most likely class of an individual point. Each decision tree in the Random Forest is trained on a randomly selected subset of the training data and a random subset of the features. This process, called bagging, helps to reduce overfitting by increasing the diversity of the trees in the forest (Breiman 1996).

This pilot study aimed to develop a methodology to map the significant plant species in a seep wetland using UAV aerial photography and establish a baseline to monitor changes. To our knowledge, this is the first study to map several seep wetland species in the Fynbos Biome simultaneously using UAV multispectral data.

Materials and methods

Study area

The study site is south of the upper Steenbras dam in the Steenbras Nature Reserve in Cape Town. The Steenbras Nature Reserve is located between Gordon’s Bay and Rooi-Els, within the greater Kogelberg Biosphere Reserve. The Kogelberg area is called ‘the heart of the fynbos’ because the reserve has several plant families and more than 1650 plant species (Wittridge 2011). Two-thirds of these species are endemic to this region. The Steenbras Catchment area was also identified as a pilot site for water abstraction from the Table Mountain Group Aquifer (TMGA)(Wiese et al. 2020). The yield from the first phase of the TMGA project is predicted to be 10 million liters per day, which will be channeled into the Steenbras Dam to complement the Cape’s bulk water supply and boost water resilience. However, boreholes should be drilled in ecologically acceptable areas and operated responsibly (Blake et al. 2021). This wetland site was chosen for monitoring because of its proximity to the City of Cape Town’s groundwater drilling sites in the Steenbras reserve. It is an inland seep wetland located within the Southern Folded Mountains ecoregion (Ollis et al. 2013) and has a variety of plant species spread over an area of approximately 1300 square meters (See Fig. 1). It is located in close proximity to a valley-bottom stream.

Fig. 1
figure 1

a Extent of study area. b Location of site in Western Cape Province. c Location of study area in Steenbras reserve. d Location of wetland area relative to the valley-bottom stream

Full size image

As previously mentioned, this pilot phase aimed to develop a methodology to map the significant plant species in the wetland using UAV aerial photography and establish a baseline to monitor any effects of water abstraction in the TMGA may have on the wetland plant species. The key plant families and species in the wetland are Proteaceae (Berzelia alopecuroides, Berzelia lanuginose), Iridaceae (Borbatia gladiata), Restionaceae (Elegia mucronata, Platycaulos compressus, Restio dispar), Ericaceae (Erica campanularis, Erica intervallaris and Erica serrata), Asteraceae (Grubbia rosmarinifolia), and Cyperaceae (Tetraria Thermalis).

Data collection

A Real-Time Kinematic (RTK) survey was undertaken using Global navigation satellite systems (GNSS) to establish coordinates of ground control points (GCPs). The GCPs were surveyed on the periphery of the block (Fig. 2). Studies have shown that GCPs placed around the periphery of the site can minimize planimetric errors (Martínez-Carricondo et al. 2018; Ulvi 2021). It is recommended to have a minimum of 4 to 5 Ground Control Points (GCPs) for every square kilometer (Ferrer-González et al. 2020). In total, six GCPs were surveyed from base station STEENBSE and used in the study. Then, a DJI Phantom 3 Professional (DJI-Innovations Inc., Shenzhen, China) was used with a Parrot Sequoia multispectral camera. The Phantom 3 had been modified such that the onboard camera had been removed and replaced with the Parrot Sequoia.

Fig. 2
figure 2

Spatial distribution of GCPs relative to the study area

Full size image

The Parrot Sequoia has four monochrome sensors with a global shutter focal length of 3.98 mm. The sensors correspond to four bands, namely Green (530–570 nm), Red (640–680 nm), Red Edge (730–740 nm) and Near Infrared (770–810 nm). The camera also has an RGB sensor with a rolling shutter and a focal length of 4.88 mm. The Sequoia has an irradiance (sunshine) sensor above the drone that is connected during data capture to facilitate the processing of at-sensor reflectance (Padró et al. 2019). A Micasense calibration reflectance panel was also used to process radiometrically corrected reflectance maps (Fig. 3d).

Fig. 3
figure 3

a Target placed over GCP; b View of target in a red edge photo; c DJI Phantom 3 with Parrot Sequoia payload and irradiance sensor above UAV (Padró et al. 2019); d A pre-flight photo of the radiometric calibration panel used for radiometric processing

Full size image

The aerial photographs were captured on 4th October 2018. Plastic ground control targets were placed over the GCPs such that the center of each target coincided with the GCP. The targets were black and white mats measuring 1 m by 1 m in dimension (Fig. 3a and b). The flight plan was designed in the Atlas Flight application. The flight parameters were a height above ground of 25 m, an overlap of 80%, and a flight speed of 5 m/s. In addition, all the sensors (RGB, Green, Red, Near Infrared and Red Edge) were activated, and the data collection started at midday. The drone was held over a Micasense calibration reflectance panel at the start and the end of the flight in order to take calibration photos in each band (See Fig. 3d).

Data processing

Initial processing of UAV aerial photography

All the RGB aerial photography processing was done in Pix4Dmapper (Pix4D SA, Lausanne, Switzerland). The first step involved computing 10,000 key points on the images to create a sparse point cloud. The coordinates of the surveyed GCPs were then added to the project, and the next step involved creating a densified point to create orthophotos and an orthomosaic with a resolution of 3 cm per pixel.

The Parrot Sequoia multispectral photos were processed similarly, except each band was radiometrically calibrated before the orthophotos were created. The calibration was done using the one-point calibration plus sunshine sensor method (Poncet et al. 2019). The calibration involved using average reflectance factor values provided by the camera manufacturers with the Micasense reflectance calibration panels. The average reflectance factors were recorded in Pix4D Mapper for each calibration photo based on the factors recommended for each band. The Pix4D software then calibrated the images during data processing based on the differences between the observed values and the actual reflectance values recorded at the reflectance target for each band in the camera. Several spectral indices were also calculated in the final step of the processing (See Table 1).

Table 1 Eighteen vegetation indices (VIs) were derived from multispectral images in this study
Full size table

Feature selection

The processed orthomosaics were layer stacked along with the calculated vegetation indices described in the preceding section. Several studies reported improvements in classifying plant species in layer-stacked images after including spectral indices (Abeysinghe et al. 2019; Bhatnagar et al. 2020; Jin et al. 2016; Mudereri et al. 2020). During the survey of the GCPs, 211 points were surveyed within the wetland, and the description of the plant species at those points was simultaneously recorded. An additional 257 polygons were digitized in Quantum GIS (QGIS) based on the locations of the GNSS surveyed points to ensure sufficient samples for classification and accuracy assessment. A point-in-polygon algorithm in QGIS was used to randomly fit points in the polygons. The criteria were that up to 20 points could be fitted in a polygon depending on the polygon size but the point spacing had to be at least 5 cm to avoid having two points overlaying the same pixel. In total, 8466 points were created. A shapefile of the points was overlaid on the layer stack consisting of the multispectral bands and the vegetation indices. A shapefile is a data format of Geographic information systems (GIS)(Elliott 2014). Reflectance values were sampled where the GNSS surveyed positions intersected the layer stack, and those values were exported to a spreadsheet. The spreadsheet consisted of 8466 data rows, with one column containing the class (plant species) data and several columns corresponding to the sampled reflectance values for that class in the spectral bands and vegetation indices. The feature selection was implemented using Recursive Feature Elimination (RFE) in R based on the Caret and Random Forest libraries. RFE aims to identify the essential features in a dataset by iteratively removing less critical features and refitting the model until a desired number of features is obtained (Demarchi et al. 2020; Ramezan 2022). The importance of a variable can be determined by examining the rankings assigned to each variable during the RFE process. Variables that are consistently ranked high throughout the iterations are considered to be more important, while those that are consistently ranked low are considered to be less important (R.-C. Chen et al. 2020). In order to assess the efficiency of the feature selection, two image layer stacks were created prior to classification. One contained only the original bands and the other contained the key bands and indices selected by RFE.

Finetuning parameters

One of the crucial steps in classification with machine learning is tuning the optimum hyperparameters of the machine learning classifiers. These include, for instance, the best values for γ and C for the radial basis function kernel in Support Vector Machines, where γ denotes the free parameter of the radial basis function, and C is the parameter that allows a trade-off between the training error and the complexity of the model. For random forest classifiers, one must optimize the number of trees in the forest (best n_estimators) and the maximum number of features considered for splitting a node (best max_features). Tuning such hyperparameters involves testing different combinations of hyperparameter values and selecting the best performance (Kranjčić and Medak 2020; Thanh Noi and Kappas 2017).

The hyperparameters used in this study were determined statistically and used for classification. Fifteen-centimeter buffers were created around the GNSS shapefile points in QGIS in order to create polygons. The new polygon shapefile was merged with the digitized polygons to make a new shapefile containing the GNSS surveyed locations and the digitized polygons. Seventy per cent of the polygons in the new shapefile were used as classification samples, and the other 30% was retained for validation. The classification samples were used to determine the hyperparameters for Random Forest, Support Vector Machine and K Nearest Neighbour using stratified k-fold cross-validation. The Dzetsaka plugin (Karasiak 2016), a Quantum GIS (QGIS) plugin that uses Python Scikit-learn (Pedregosa et al. 2011), was used for the cross-validation. This approach randomly splits the observations into k groups, also called folds, of approximately equal size. The first fold is then treated as a validation dataset, and the classifier is trained on the remaining k − 1 folds. The Mean Square Error (MSE) is calculated on the validation fold. This process is repeated k times, and the MSE is averaged to get the cross-validation estimates. In the stratified cross-validation approach, the folds are made by preserving the same percentage of training samples for each information class.

In this study, the classification dataset was split into 5 groups (i.e., k = 5) and 80% of the samples from each class were used for training and the other 20% for validation. The best hyperparameters were identified by scripting a grid search in the Dzetsaka plugin.

Classification

All three machine learning classifiers, namely, Random Forest (RF), Support Vector Machines (SVM), and K Nearest Neighbour (KNN), were used in the classification. Two image layer stacks were classified: one with only the original bands and the other with the ten best bands and indices (See Fig. 4). In both instances, the classification was optimized with finetuned hyperparameters. The results are presented in the forthcoming section. The validation dataset comprising 30% of the samples was used to calculate the Overall Accuracy (OA) and the Kappa hat coefficient (k). The equations for calculating overall accuracy and kappa are shown below.

$$OA = {{number~of~~correctly~~classified~~pixels~} \over {total~number~~of~pixels}}$$
(1)
$$k = {{N\mathop \sum \nolimits_{i = 1}^r {x_{ii}} - \mathop \sum \nolimits_{i = 1}^r \left( {{x_{i + 1}} \times {x_{ + i}}} \right)} \over {{N^2} - \mathop \sum \nolimits_{i = 1}^r ({x_{i + }} \times {x_{ + i}})}}$$
(2)

where r = the number of rows and columns in the error matrix, xii = the number of observations in row i and column i, xi + = the marginal sum of row i, x+ i = the marginal sum of column i, and N = the total number of observations.

Fig. 4
figure 4

Graph showing classification model accuracy plotted against number of features

Full size image

User and producer accuracies were also calculated. Producer accuracies are the probability that a land cover class is classified as such (Story and Congalton 1986). The user accuracy represents the probability that the predicted class on the map is present on the ground or to what extent the other classes may have been misclassified as the class in question (Congalton et al. 1983; Patel and Kaushal 2010). Lastly, the degree of agreement or consensus between the classifiers was determined by comparing the spatial overlap of the class distributions in the classification images (Mas et al. 2022). The degree of agreement was calculated per classifier as well as per class.

Results

Feature selection

The RFE results showed that using fifteen features instead of twenty-two would produce the best accuracy results (Fig. 4). However, the cross-validation accuracy assessment of the models developed using RFE when using ten, fifteen and all variables were very similar. The kappa values were 0.960, 0.961 and 0.959, respectively. Thus, only the top ten features, namely NGRDI, RG, Green, LogRE, NDRE, CIRE, GRVI, NDWI, GNDVI and Red were subsequently used to classify the plant species.

Figure 5 below shows the relative feature importance of the top fifteen features selected during the RFE feature selection.

Fig. 5
figure 5

Showing the variable importance of the bands and indices

Full size image

Classification maps

Figure 6 shows the spatial distributions of the plant species in the wetland. The plant species thrive in different parts of the wetland based on the degree of soil wetness. The three classifiers generally showed a similar distribution of the species.

Fig. 6
figure 6

Classified images a Random Forest, b SVM, c KNN for layer stack with indices

Full size image

Classification statistics

Figure 7 shows the accuracy statistics for each classifier for both layer stacks (one with original bands and the other with RFE selected bands and indices). The columns with the suffix ‘WI’ indicate the datasets after feature selection. The table after that (Table 2) shows classification accuracies per class.

Fig. 7
figure 7

Overall model training and classification accuracies for all classifiers

Full size image
Table 2 Classification accuracies per class for each classifier
Full size table

The best classifier was Random Forest, with an overall accuracy of 87% and kappa hat value of 0.85. The overall classification accuracy was 4% better than the results from Support Vector Machines and 2% better than K Nearest Number. The producer accuracies ranged between 84% and 96%. Figure 7 also shows that the overall training accuracies and kappa hat values were generally good for all the classifiers. Figures 8 and 9 are graphical representations of Table 2.

Fig. 8
figure 8

Graph showing producer and user accuracies per class

Full size image
Fig. 9
figure 9

Graph showing kappa hat values per class

Full size image

Classification accuracy per class

Classification accuracies differed across the plant species (Table 2). The descriptions of the acronyms in the table are as follows: B (Berzelia), BG (Borbotia Gladiata), DPC (Dry or Dead Platycaulos Compressus), EM (Elegia Mucronata), EC (Erica Campanularis), EI (Erica Intervallaris), ES (Erica Serrata), GS (Grey Soil), GR (Grubbia Rosmarinifolia), PC (Platycaulos Compressus), RD (Restio Dispar), S (Shadow), TT (Tetraria Thermalis) and WS represents White Soil. The accuracies vary with each class and classifier.

The Berzelias, Erica Serrata and Restio Dispar had the lowest classification accuracies with Kappa hat values of 0.63, 0.57 and 0.60, respectively. The Erica Serrata was classified better using the Support Vector Machines classifier with a Kappa hat accuracy of 0.73. The Berzelias were classified best using the K Nearest Neighbour classifier. Conversely, the taller, more dominant species, Grubbia Rosmarinifolia, Platycaulos Compressus and Elegia Mucronata were classified well with kappa hat values of 0.88, 0.96 and 0.92, respectively. The Borbotia Gladiata and Tetraria Thermalis clusters are less than half a meter in height but also had acceptable classification results with kappa hat values of 0.85 and 0.83, respectively. In contrast to the Erica Serrata, the Erica Campanularis and Erica Intervallaris showed good accuracies of 0.72 and 0.82, respectively. K Nearest Neighbor presented the best producer accuracies for Erica Serrata and Restio Dispar. K Nearest neighbour performed well classifying the Erica family of flowering plants and the small clusters of Borbotia Gladiata, Dead Platycaulos Compressus and Restio Dispar.

However, Random Forest generally had comparatively good user accuracy statistics. The user accuracy is a measure of reliability of the classification to the user (Rwanga and Ndambuki 2017). It is the likelihood that the classification result actually represents that category on the ground (Story and Congalton 1986). Though the Random Forest classifier wrongly classified some pixels as per the producer accuracy statistics (Fig. 8), it generally produced the most user-reliable classification of the individual classes. Nearest Neighbor presented the best producer accuracies for both the Erica Serrata and Restio Dispar. K Nearest neighbour performed well classifying the Erica family of flowering plants and the small clusters of Borbotia Gladiata, Dead Platycaulos Compressus and Restio Dispar. However, Random Forest generally had comparatively good user accuracy statistics. Thus, although the Random Forest classifier wrongly classified some pixels as per the producer accuracy statistics (Fig. 8), it generally produced the most user-reliable classification of the individual classes.

Degree of agreement between classifications

Table 3 shows the degree of similarity between the different classification maps. The descriptions of the acronyms in the table are as follows: B (Berzelia), BG (Borbotia Gladiata), DPC (Dry or Dead Platycaulos Compressus), EM (Elegia Mucronata), EC (Erica Campanularis), EI (Erica Intervallaris), ES (Erica Serrata), GS (Grey Soil), GR (Grubbia Rosmarinifolia), PC (Platycaulos Compressus), RD (Restio Dispar), S (Shadow), TT (Tetraria Thermalis) and WS represents White Soil.

Table 3 Agreement between the classification results of the different classifiers
Full size table

The best agreement across all classifiers was for Borbotia Gladiata, Dead Platycaulos Compressus, Elegia Mucronata, Erica Campanularis and Tetraria Thermalis. There was 70.48% overall agreement (and kappa of 0.66) between the Random Forest and K nearest neighbor results. Among the plants, the highest agreement was achieved when mapping Tetraria Thermalis, Restio Dispar, Elegia Mucronata and Borbotia Gladiata respectively. Despite an overall agreement of almost 65.77% (and kappa of 0.60), there was poor agreement between the classification outputs of Random Forest and Support Vector machines for Platycaulos Compressus, Erica Serrata and Restio Dispar.

Lastly, there was fairly good agreement between K Nearest neighbor and Support vector Machines (70.51% and kappa of 0.66). The highest consensus for the two classifiers was found for Elegia Mucronata, Grubbia Rosmarinifolia and Berzelia and poor agreement was found for the classification of Erica Serrata and Restio Dispar.

Discussion

This study sought to ascertain the UAV multispectral aerial photography capacity to remotely map several seep wetland plant species in the Fynbos Biome. The key plant families and species explored were Proteaceae (Berzelia alopecuroides, Berzelia lanuginose), Iridaceae (Borbatia gladiata), Restionaceae (Elegia mucronata, Platycaulos compressus, Restio dispar), Ericaceae (Erica campanularis, Erica intervallaris and Erica serrata), Asteraceae (Grubbia rosmarinifolia), and Cyperaceae (Tetraria Thermalis). The study used three machine learning algorithms, namely, Random Forest (RF), Support Vector Machines (SVM), and K Nearest Neighbor (KNN). The hyperparameters for all three classifiers were finetuned to optimize the classification (Duncan et al. 2023; Kuradusenge et al. 2023). The classification was done on a dataset of critical spectral bands and indices selected based on their capacity to optimize the classification process. The findings indicated that by utilizing the RFE-chosen variables, accuracy of the classification training model increased by as much as 3.7%, while the overall accuracy of the classification improved by up to 1.6%. The classification maps showed the spatial distribution of the wetland species in the study area.

Feature selection

Studies have shown that vegetation indices (VIs) can highlight phenological differences and improve the potential for classification (Doughty and Cavanaugh 2019; Ma et al. 2019; Zhuo et al. 2022). Conversely, incorporating VIs before classification increases the dimensionality of multispectral data. Having many features can cause a learning model to overfit and increase the computational cost of data processing. This problem can be mitigated by either feature extraction or feature selection (Jovic et al. 2015; Tang et al. 2014). In feature extraction, the original features are transformed into a new set that retains the more meaningful information from the original collection (Jovic et al. 2015; Tang et al. 2014). Remote Sensing studies (Arun 2022; Avola et al. 2019; Nikolakopoulos et al. 2004) have used feature extraction techniques such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Canonical Correlation Analysis (CCA) to reduce data dimensionality. The alternative approach, feature selection, extracts a small subset of features from the original set of features without any transformation (Jovic et al. 2015; Tang et al. 2014). Features are ranked from strongly relevant to redundant, and feature selection aims to capitalize on relevance and diminish redundancy (Jovic et al. 2015). The feature selection methods are broadly categorized as filter, wrapper, embedded, and hybrid. Filter methods select features by assessing their performance independently of data modelling algorithms. Wrapper methods perform better than filter methods because the feature subsets are evaluated by how well they improve the performance of a modelling algorithm (Jovic et al. 2015).

This study used Recursive Feature Elimination (RFE) for feature selection. RFE is a wrapper feature selection method frequently used with random forest and support vector machines (Demarchi et al. 2020; Poona et al. 2016). RFE starts by testing the complete feature set and computing each component’s importance score. Then, the least important features are iteratively removed as the model is reassembled, and an importance score is recalculated until the user-defined number of subsets is reached. This study used the RFE to create ten subsets to identify the best fifteen features for classification. The process was repeated five times. Eighteen vegetation indices were assessed along with the four multispectral bands (n = 22).

The results showed that the model’s accuracy peaked at 15 features out of 22. However, the accuracies when using ten, fifteen and all features were very similar, with kappa hat values of 0.960, 0.961 and 0.959, respectively. The features added after the first ten did not significantly improve the accuracy of the classification model. Consequently, 60% of the features were discarded before classification at an insignificant cost to classification accuracy. The retained indices were NGRDI, RG, LogRE, NDRE, CIRE, GRVI, NDWI and GNDV. The NGRDI had a significantly higher importance score than the other features.

The wetland vegetation properties

This study found that NGRDI, RG, Green, LogRE, NDRE, CIRE, GRVI, NDWI, GNDVI and Red were essential for classifying wetland vegetation. Of these variables, the Normalized Green–Red Difference Index (NGRDI) was significantly more important than the rest. NGRDI leverages differences in the reflectance of the Green and Red bands (Gitelson et al. 2002). Studies have also shown that leaf pigments (chlorophyll, carotenoids and anthocyanins) influence the interaction of vegetation with the visible portion of electromagnetic radiation (Gausman 1977; Govender et al. 2009). The absorbance of the red band is based primarily on chlorophyll content, whilst the absorbance of the green band is based on both chlorophylls and anthocyanins (Gitelson 2011).

Consequently, NGRDI can leverage differences in reflectance of the red portion of the electromagnetic spectrum and highlight vitality in vegetation (Song and Park 2020). Notably, the Red Green Vegetation Index (RG) tested in this study also exploited the reflectance in the Green and Red bands and was the second most crucial variable. Lastly, the Green-Red Vegetation Index (GRVI) is highly sensitive to chlorophyll (Duncan et al. 2023; Yang et al. 2017) and the Green Ratio Vegetation Index (GRVI), which is mathematically similar to the modified anthocyanin index (mACI) is sensitive to anthocyanins(Gitelson 2011; Motohka et al. 2010).

NGRDI has also been correlated with biomass and nitrogen content (Choudhary et al. 2021; Elazab et al. 2016; Hunt et al. 2005; Jannoura et al. 2015; Li et al. 2016; Smigaj et al. 2019) Biomass, chlorophyll concentration and leaf water content are the most significant biophysical and biochemical properties that characterize wetland vegetation (Adam et al. 2010; Mishra 2020). Vegetation biomass is typically a proxy for local carbon storage, wetland health, and vulnerability to human activity and environmental or climate change (Doughty and Cavanaugh 2019; Han et al. 2019; Klemas 2013; Sun et al. 2021). Nitrogen is an essential component of chlorophyll (Bassi et al. 2018; Wang et al. 2014). Thus, the results suggest that there could be significant differences between the chlorophyll content and biomass in the study area to warrant a high performance of NGRDI.

Leaf water content is frequently estimated using the near-infrared to shortwave portions of the electromagnetic spectrum (Govender et al. 2009). However, there is a correlation between leaf water and chlorophyll activity and reflectance of the red-edge part of the spectrum (Ndlovu et al. 2021) Thus, other than NDWI, red-edge indices such as NDRE, GNDVI and CIRE indices have been proposed as water-sensitive vegetation indices (Ndlovu et al. 2021; Yang et al. 2017) Indices can help highlight different vegetation features across various plants, plant concentrations and stages of growth (Boiarskii and Hasegawa 2019).

Classification statistics

Three classifiers were used to classify two datasets. One dataset contained only the original spectral bands, and the other included the bands and spectral indices selected during the RFE process. The results showed that the use of the chosen variables improved the out-of-bag accuracy of the classification training model by up to 3.7% for Random Forest (RF), 1.8% for Support Vector Machines (SVM) and 3.5% for K nearest Neighbour (KNN). In addition, the overall accuracy of the classifications increased by 0.4% (RF), 1.6% (SVM) and 1.2% for K Nearest Neighbour (KNN). Overall, RF had the best classification statistics with an overall accuracy of 87.4% and kappa accuracy of 0.85. KNN had an overall accuracy of 85.3% and kappa accuracy of 0.83; SVM had an overall accuracy of 83.6%; and kappa accuracy of 0.81.

The RF classification determined that Platycaulos Compressus was the best classified plant species, followed by the Elegia Mucronata and the Dead Platycaulos Compressus. All three classes had kappa accuracies of more than 0.9 and less than 10% commission error. The Grubbia Rosmarinifolia, Borbotia Gladiata, Erica Intervallaris and Tetraria Thermalis had kappa accuracies of 0.88, 0.85, 0.82 and 0.83. Of these four species, Grubbia Rosmarinifolia had the highest producer accuracy of 74.78%, indicating the omission of 25.22% of the species in the final map. In contrast, the user accuracy was 90.21%, suggesting that only 9.79% of the other species were misclassified as Grubbia Rosmarinifolia.

The poorly classified species included Berzelia (kappa of 0.63), Restio Dispar (kappa of 0.56) and Erica Serrata (kappa of 0.57). They were generally classified well, with a producer accuracy of 85.05% and low omission error (14.95%). However, the user accuracy of 66.73% suggests the species was overestimated by 33.27%. The KNN classification of the same species only overestimated the species by 27.78%. The Erica Serrata, and Restio Dispar had the poorest classification statistics. The classification results of Restio Dispar indicate that more than 50% of the species were omitted and 39.29% of the classified pixels belonged to other classes. However, it must be noted that Restio Dispar, which grew in individual tufts, was also one of the least represented species in the wetland. The Erica Serrata was classified with a producer accuracy of 35.36% and a user accuracy of 57.50%. This result means that more than half of the class was omitted, and 42.5% of the class is erroneous. Like Restio Dispar, the Erica Serrata also occurred in patches and small clusters west of the wetland. The SVM and KNN classifiers performed better than RF in classifying Erica Serrata, with 27% and 31% commission errors, respectively. Studies suggest small patch size and plant density can significantly impact classification accuracy (Adam et al. 2010; Duncan et al. 2023). In addition, machine learning classifiers generally require many samples for a good classification (Duncan et al. 2023). The classification accuracies of Berzelia, Restio Dispar and Erica Serrata could be improved by collecting the data in a different season since the spectral signature of wetland species can be affected by different seasons and illumination (Adam et al. 2010; Gallant 2015).

Table 3 shows the degree of agreement between the classifiers. The assessment of classifier agreement revealed strong consensus among classifiers when it came to classifying Borbotia Gladiata, Dead Platycaulos Compressus, Elegia Mucronata, Erica Campanularis, Tetraria Thermalis, and bareground. However, there was notably less agreement when classifying Restio Dispar and Erica Serrata, particularly because these species exhibited patchy distribution within the wetland. Remarkably, the Random Forest classifier and K Nearest Neighbour exhibited the highest similarity in classifying these two challenging classes. This observation suggests that these classifiers may demonstrate greater robustness in classifying vegetation types when confronted with limited sample data.

The spatial distribution of the plant species

This paper aimed to map several seep wetland plant species remotely. The study findings showed that the distributions of the dominant plant species in the wetland can be depicted. The plants were found to be clustered in different areas of the wetland. Grubbia Rosmarinifolia and Platycaulos Compressus were the most prevalent species in the northern portion of the wetland. That area is also the wettest portion of the wetland.

Erica Campanularis and Erica Intervallaris thrived in both the wettest portions of the wetland to the North and the drier parts of the wetland to the West and South. Erica Serrata was spread around the drier portions of the wetland in small clusters. Elegia Mucronata was found to coexist with Berzelia Lanuginosa and Berzelia Alopecuroides, particularly in less water-logged soils around the transect line at the centre of the wetland. That portion of the wetland is seasonally wet. Borbotia Gladiata, Restio Dispar and Tetraria Thermalis were tiny clusters in the dry parts of the wetland adjacent to the Southernmost transect line. Tetraria Thermalis clusters were all located west of the wetland. Of all the species, the Platycaulos Compressus was the most dominant in the wetland and surrounding portions of the Steenbras reserve.

Conclusions

This study presented a methodology for using multispectral aerial photography to discriminate several wetland plant families and species. The study found that the Normalized Green–Red Difference Index (NGRDI) and Red Green Vegetation Index (RG) were the most critical indices for the discrimination of the different wetland plant species at the start of summer. The other chlorophyll and water-sensitive indices were also essential for classifying the plant species. It was also found that classifying a subset of indices and bands produced overall accuracies of between 87.4% and 83.6% and kappa hat accuracies of 0.85 and 0.81. The accuracy of the classification models improved by up to 3.7% after combining selected vegetation indices and band data, and the overall classification accuracy of all three classifications by between 0.4% and 1.6%.

Of the three classifiers, Random Forest performed best, with an overall accuracy of 87% and kappa hat of 0.85. It was followed by Support Vector Machines and then K Nearest Number. However, K Nearest Neighbour performed well when classifying the Erica family of flowering plants and the small plant clusters of Borbotia Gladiata, Dead Platycaulos Compressus and Restio Dispar.Grubbia Rosmarinifolia and Platycaulos Compressus were the most notable species in the northern and wettest portion of the wetland. The Ericaceae species were spread around the drier parts of the wetland. Elegia Mucronata, Berzelia Lanuginosa, Berzelia Alopecuroides and Tetraria Thermalis thrived in the wetland’s moist soils at the centre and western regions. Borbotia Gladiata and Restio Dispar occurred in tiny clusters in the wetland on dry patches.

This study used UAV multispectral aerial photography to classify plant species in a wetland in the Fynbos Biome. To our knowledge, no study has tested the viability of using remote sensing data to map wetland species in the Proteaceae, Iridaceae, Restionaceae, Ericaceae, Asteraceae and Cyperaceae families. The results show that the methodology used in this study could be replicated in other Fynbos wetlands. Future studies should explore classifications of the same species in different seasons to assess the best time of year to classify them. In addition, future studies should explore the use of Geographic Object-Based Image Analysis and deep learning classifiers.

There are significant benefits and implications of this pioneering study. Firstly, the study contributes to the spectral characterisation of wetland plant species in the Fynbos Biome. It provides the foundation for future Fynbos remote sensing research and practical implications for conservation and land management. Moreover, understanding the distribution and spatial preferences of key wetland species can assist conservation managers in making more informed decisions about land use and protection strategies in the Cape Floristic Region. Thus, this knowledge will contribute to preserving these vital wetland ecosystems and their ecological functions.