An Assessment of Wakkerstroom Wetland and Its Vegetation Communities from 1938 to 2019

There are few studies on the size and changes in species composition over time for wetlands in South Africa. Techniques such as remote sensing have become popular in assisting the development of management plans due to their spatio-temporal advantages and easily reproducible vegetation and land cover maps. The Wakkerstroom wetland was examined using aerial photography to examine possible changes in the extent and Landsat imaging was used to map its vegetation communities. To assess the distribution of vegetation types on Wakkerstroom wetland, in situ recording of vegetation types and their GPS coordinates was conducted and a Random Forest model was used to predict vegetation types from Landsat pixel spectra across the wetland extent. As calculated from aerial photographs, the Wakkerstroom wetland has increased in extent by 0.483 km from 1938 to 2009. The P. australis population density increased significantly over time (r = 0.89), whereas the T. capensis population density had a strong negative correlation over time (r = −0.70). A strong negative relationship between P. australis and T. capensis existed (r = −0.88). A need exists to introduce a management tool that will create a greater mosaic of vegetation communities thus ensuring a greater bird, reptile, and amphibian diversity.

Phragmites australis (Cav. Steud) and Typha capensis (Rohrb. N. E. Br.) [5] [6] [7] [8] [9]. Studies have voiced concern about the hyper competitiveness of P. australis which has achieved invasive weed status in several countries, including North America where huge monospecific swathes have outcompeted the native vegetation and have thus diminished biodiversity [10] [11] [12] [13]. Wetlands, due to their heterogeneous vegetation and landscape provide habitat for a highly diverse array of bird and invertebrate species [14]. Flood attenuation and water purification are key services provided by the Wakkerstroom wetland for users downstream. Grazing for cattle is supplemented by wetland vegetation especially in drier months where another grazing is inadequate or insufficient [15].
The need to conserve and manage remaining wetlands and the crucial ecosystem services they provide are essential for human wellbeing. Hall [16] proposed that it is not enough only to conserve the populations of plants and animals exploited by anthropogenic activities but that their health and sustainability depends on conserving all biodiversity included in the wetland. Bond [17] states that although climate sets the limit to all terrestrial and aquatic vegetation growth; fire and herbivory determines the pattern of vegetation distribution. A fire regime can be described as the combination of frequency, season, intensity and type of fire that occurs in an area. A fire regime results from a sequence of individual fire events. The response of ecosystems to fire depends on both the effects of that single fire and the behaviour inherited from previous fires [18]. Burning of wetlands has numerous potential positive consequences [15], which includes the maintenance of native fauna and flora; assisting in alien plant control; removing plant litter and improving grazing value. However, negative consequences exist too, susceptible animal species could asphyxiate or succumb to the direct effects of heat, very regular fires can increase rates of erosion (especially on heavily grazed wetlands), increased evaporation from the wetland combined with a decrease in organic matter as well as an increase in ash content in the soil affects productivity. P. australis can be considered a wetland management problem due to its rapidly colonising and dominating behaviours in disturbed soils  [19]. Kotze [20] found that when it was desirable to reduce the abundance of the dominant plant species to enhance habitat diversity both fire and herbicide were necessary. Unfortunately, successful conversion of dominated wetlands using only chemical control requires a diverse soil seed bank in addition to a nearby source of seeds for natural recruitment. Areas that do not possess these seed sources may need manual planting to speed up site recovery. Therefore, fire regimes and herbicide utilisation to increase plant diversity without harming the wetland are very site and species specific. Ailstock et al. [19] found that single herbicide application or herbicide application followed by burning can reduce the abundance of P. australis acutely in wetlands. Long term diversity maintenance required occasional herbicide applications to prevent regrowth of P. australis.
The Wakkerstroom wetland is classified as a national heritage site, but a pro-

Study Site
Wakkerstroom is in the southeast of Mpumalanga along the northern edge of the Thaka River which flows into the Zaaihoek Dam [6]. The wetland has an extensive sedge marsh belt containing Typha capensis which grades into sedge meadow (plants of the Cyperaceae family). The permanently flooded interior is covered by wet grassland-dominated by Phragmites australis [22].

Remote Sensing Data Collection
To determine if the boundary of Wakkerstroom wetland has changed over time aerial photographs collected in 1938,1953,1969,1979  for spectral extraction and modelling.

Modelling
The Landsat images were the spectral data sources used for the classification modelling and prediction mapping. The GPS points were used to match spectra from the Landsat images to their respective vegetation type using R [23] and the raster package. The spectra and associated vegetation classes were combined into a data frame for analysis. Random Forest was used to create models to predict vegetation classes across the entire wetland [24]. First, a training model was built using 70% of the total data by means of the bootstrap resampling method and the other 30% to validate and choose the best model for prediction mapping.
After the best training model was created two types of models were built. The

Analyses
Changes in vegetation structure and distribution were analysed using Pearson correlations. Correlations were performed between vegetation classes over time and the effect of one vegetation class on another.

Determining a Change in Extent of Wetland
To determine if the boundary of the Wakkerstroom wetland has changed over time, the wetland was studied from several aerial photographs that were collected in 1938,1953,1969,1979 and an orthophoto from 2009 ( Figure 3). The boundary for each year was determined as the distinctive visual border that separated wetland vegetation from surrounding grassland. Figure 4 suggests that the boundary of the wetland is relatively dynamic. Each image illustrates that while a general shape of the wetland exists, the edges of the wetland shift noticeably over time.

Creating a Random Forest Model for Vegetation Classifications
The in situ vegetation recorded at the GPS points were used to construct Random Forest models with the respective spectra from the Landsat images. Two approaches to Random Forest models were tested. The first model was built on the assumption that for each image the vegetation type at those GPS points would remain the same as 2019 in situ classification. The second type of Journal of Water Resource and Protection model was built on the assumption that the spectral signature of each vegetation type would remain the same. A model was derived for the 2019 image and then applied to prior years. Neither of these assumptions is strictly accurate in a natural environment but using images at similar times of year the supposition was that the reflectance value as a product of photosynthetic action of each vegetation community would be similar and therefore adequate to be used for modelling purposes. The second type of modelling using the reflectance of each vegetation community created models with the same error rates and modelling values as the first model and so model one was used hereafter (Table 1).
Each model was trained with 70% of the total data. As the model was calibrated, each decision tree component of the random forest model was tested by the samples not used in building that tree. This is known as the out of bag error estimate as it is an internal error estimate of a random forest model as it is being constructed. The final models were tested with the remaining 30% of the  Class error (Table 2) represents the likelihood of misclassification and so a higher percentage result correlates to a higher likelihood of misclassifying the vegetation classification. Table 2 illustrates that the model had the most difficulty classifying grass (42.9% classification error) and P. australis (58.8% classification error). The validation model predicted a 70% total accuracy when using the out of bag data points to create a vegetation classification and so 70% of the time a data point will be classified correctly to its in situ classification. ror rate, the model had a 53% accuracy value when using the validation data to predict vegetation classes from Landsat spectra when compared to in situ classification. Therefore 53% of the time a data point will be classified correctly to its in situ classification.
Similarly in Table 2, the 1997 classification model struggles to classify grass (57.1% likelihood of misclassification) and P. australis (70.6% likelihood of misclassification) but additionally, this model struggles to classify P. australis/T. capensis mix (60% likelihood of misclassification).
This study used six reflectance bands from Landsat 5 for modelling and    Table 4).

Examining the Change in Wetland Extent Using Aerial Photographs
Tuominen and Pekkarinen [25] discussed how aerial photography has assisted in vegetation and landscape mapping since the 1930s. Aerial photographs provide a general "big picture" view of landscapes and the surrounding terrain that can be used to delineate site boundaries including wetland borders [26]. A consideration put forward by Davis and Wang [30] is that each image should be geo-referenced to the same base map using the same reference coordinate The development of the open water area east of the Paul Kruger Bridge, which is likely due to the building of the causeway in 1979, is a distinctively visual example of human induced change to the wetland extent. The continual increase in wetland extent from 1938 to 2009 could possibly be attributed to the competitive and rapid growth pattern of P. australis [9]. The dense rooting network created by homogenous stands of P. australis slows water flow velocity, increases diffuse flow across the wetland and encourages sediment deposition which prolongs periods of flooding. These conditions may possibly have led to an increase in lateral extension of the wetland.

Mapping and Quantifying the Distribution of Each Vegetation Type Using Remote Sensing
To gather the longest historical record of data, images from both Landsat 5 (TM) and Landsat 8 (OLI) satellites were used. Both satellites have the same image spatial resolution and although these satellites have different spectral resolutions, the presence of infrared sensors on each satellite allowed vegetation identification possible. Supervised classifications can be used to classify vegetation classes automatically using computer algorithms to determine the probability that a pixel belongs to a certain class. One such method is Random Forest which is a non-parametric extension of decision tree modelling where unclassified vegetation classes can be categorised using in situ training samples [24].  [35]. The aim to map the distribution of vegetation types present on the wetland did not provide significant results that demonstrated the model could classify vegetation with a high level of certainty but these results did allow for an improved understanding of wetland vegetation dynamics. Factors that might be responsible for these results could include the spatial and spectral resolution of the data interpreted from each satellite. Repeating the modelling using other satellites that have a finer spectral or spatial resolution might be useful for this type of study. Photosynthetic vegetation which contains chlorophyll a and b tend to absorb light of the wavelengths ~400 -700 nm and has a much higher reflectance in the near infrared region between ~700 -1400 nm. The results showed that the spectral reflectance using the Landsat narrow bands for each vegetation type was very similar and contained no discernible contrast with which to make distinc-Journal of Water Resource and Protection tive classifications. This could be attributed to the vegetation types having similar concentrations of photosynthetic pigments and similar photosynthetic activity. Mahdavi et al. [36] and Wang et al. [37] attributed the difficulty they encountered when using satellite imagery to differentiate between vegetation types to wetlands having a high intra-species and low inter-species variability in reflectance.

Conclusion
The conservation of the Wakkerstroom wetland is critical due to its importance in providing a habitat for numerous bird and invertebrate species. It is necessary to understand the dynamics of the size of the wetland as well as changes in vege- . This is worrying as P. australis is extremely aggressive and limits the types of habitats available for the birds and invertebrates. A carefully governed fire regime with occasional herbicide application would be the only management tool that would help limit the spread of P. australis. The fire should be used to introduce a mosaic of vegetation species and herbicide application to limit the regrowth of P. australis.