Site selection for subtropical thicket restoration: mapping cold-air pooling in the South African sub-escarpment lowlands

Region

The terrain of the coastal lowlands is particularly suited for producing CAP due to the characteristic valley morphologies. The west–east trending Cape Folded Belt consists of erosion resistant sandstones that form long parallel ranges with shales and small patches of Dwyka tillite persisting on the lower slopes and valley floors; this gives rise to steep convex slopes with wide flat valley bottoms (similar to the glacially carved U-shaped valleys in the Lundquist, Pepin & Rochford (2008) study). The long parallel ranges are each fairly narrow (rarely more than 10 km wide), separated by valleys that range in width (but usually not wider than 30 km). The ranges are cut through by very narrow defiles by rivers flowing from the Great Escarpment to the sea (i.e. roughly from north to south). The rugged ranges, open valley floors, with valley constrictions through mountains increase the potential for CAP in this landscape.

Our approach, in brief, was to calibrate the CAP model (Lundquist, Pepin & Rochford, 2008) using two adjacent valleys in the sub-escarpment lowlands where a range of CAP-related data have been collected (e.g. during and after frost events, including temperature, transplant experiments and field observations; Duker et al., 2015a, 2015b). Next, we tested the model predictions at six different localities—selected from satellite imagery (via Google Earth Pro)—using the visually distinctive treeline between the thicket and Karoo shrubland. Finally, we tested our CAP model by using productivity data of spekboom planted in the biome-wide thicket restoration experiment. We describe each of these steps in detail below.

Cold air pooling model calibration

Two valleys (S33°15′25.00″ E25°25′20.10″ and S33°15′10.34″ E25°24′13.34″; Figs. 2 and 3) were used for model calibration and frost-risk mapping. These sites were used because of the availability of 5 years of temperature data (from July 2012 to October 2017), and many hours of field observations during and after frost events (over 100 days spent on site spread across four consecutive winters from 2013 to 2016). These ‘calibration valleys’ are located in the Zuurberg mountain range in the north-eastern portion of the Cape Fold Mountains. Temperature loggers and field observations have confirmed that radiative frost (0 °C at ground level) occured on valley floors up to 15 times each winter as a result of CAP on still and clear nights (detailed in Duker et al. (2015a, 2015b)).

The first valley (Klipfontein) is ~7.5 km long, ~3 km wide and ~300 m deep, whereas the second valley (Buffels Nek) is ~12 km long, ~2.5 km wide, and ~500 m deep (Figs. 3B and 3C). The vegetation patterns in these valleys consist of grassland and fynbos occupying the higher elevation sandstone ridges, and thicket dominating the shale-derived soils on the lower valley slopes, and Karoo shrubland occupying valley floors (which is a combination of shale- and alluvially-derived soils). In these valleys, soil conditions (i.e. particle size, depth, infiltration rate, sodium and electrical conductivity) are consistent across the valley floors and shale-derived sections of the valley slopes (Becker et al., 2015). Nonetheless, Karoo shrubland replaces thicket on the valley floor—this is largely driven by frost (Duker et al., 2015a, 2015b). In these valleys, some of the thicket-shrubland treelines are intact and clearly defined due to historically low levels of livestock pressure (Ian Ritchie, 2016, personal communication); this boundary has become blurred in places as livestock has removed thicket, and Karoo shrubland species have subsequently invaded. In such cases, frost-intolerant, but browser-resistant, indicator species (e.g. Pappea capensis) were used to identify the position of the treeline.

As outlined in the introduction, radiative frost on the southern African sub-escarpment lowlands predominantly occurs on valley floors and in depressions, and is associated with CAP. Following the Lundquist, Pepin & Rochford (2008) model, and using the ALOS 30 m (Takaku, Tadono & Tsutsui, 2014) digital elevation model (DEM), we generated the range of topographic characteristics—slope, height relative to surrounding cells and curvature—required to identify areas prone to CAP and radiative frost (Fig. 4). Specifically, the Lundquist, Pepin & Rochford (2008) model applies thresholds to maps of slope, rank elevation and curvature and combines these to generate a summary map predicting where CAP should occur. Slope was calculated using the Horn (1981) algorithm in the raster package (Hijmans & Van Etten, 2012; Fig. 4). Rank elevation for each DEM grid cell (i.e. percentile elevation relative to surrounding terrain) was determined by calculating it’s topographic position within a given radius in the valley landscape—the radius was manually calculated for each valley and is defined as half the distance between the two ridges that define the valley in which the DEM grid cell is situated (Fig. 4). Specifically, we calculated the ‘rank elevation of each DEM grid cell relative to the elevation of surrounding DEM grid cells within the square with the specified radius in each cardinal direction from the DEM grid cell’ (Lundquist, Pepin & Rochford, 2008). Calculating rank elevation in this way places each DEM grid cell within the context of the broader landscape (i.e. valley-scale), rather than the conventional approach that only uses the immediate neighbouring cells for rank calculations.

The broad scale curvature was calculated using the Liston & Elder (2006) snow model formula: $$cv={\displaystyle \frac{1}{4}\left\{{\displaystyle \frac{1}{2r}\left(Z-{\displaystyle \frac{Zw+Ze}{2}+Z-{\displaystyle \frac{Zn+Zs}{2}}}\right)+{\displaystyle \frac{1}{2\sqrt{2r}}\left(Z-{\displaystyle \frac{Zsw+Zne}{2}+Z-{\displaystyle \frac{Znw+Zse}{2}}}\right)}}\right\}}$$where cv is the curvature at any particular DEM grid cell, Z is the elevation of that DEM grid cell, r is the relevant radius for the landscape under investigation, Z_w/e/n/s is the elevation a distance r to the west/east/north/south of the DEM grid cell, and Z_sw/ne/nw/se is the elevation a distance r to the south-west/north-east/north-west/south-east away from the DEM grid cell. Again, the curvature calculated here does not use the cells immediately surrounding each DEM grid cell, but is a derivative of the specific radius set by the user for the particular valley. Thus, this determines whether a DEM grid cell is on a ridge or within a valley within the broader surrounding landscape (Fig. 4). The curvature and rank elevation calculations used the same user-defined valley radius.

The Lundquist, Pepin & Rochford (2008) CAP model was developed in the northern hemisphere (beyond 37°N)—here we applied this same approach in the southern hemisphere, where continental temperatures are ameliorated by the larger ocean area and in a region that is at least 3° closer to the equator. The model thresholds used by Lundquist, Pepin & Rochford (2008) had poor results when applied to the sub-escarpment coastal plains in South Africa. Specifically, the model had a high degree of false positives—that is it predicted CAP far upslope beyond the thicket-shrubland tree-line. This is likely because the temperature gradients that form along these southern coastal lowlands are not as steep as those experienced at the high elevation (1,000–3,000 m) and latitude (37–40° N) study areas of Lundquist, Pepin & Rochford (2008). Therefore, we modified the threshold criteria based on observations and data of known frost events in the calibration valleys. Thus, our calibrated thresholds are to predict CAP are (i) slope less than 10° (originally: <30° in Lundquist, Pepin & Rochford (2008)), (ii) height relative to surrounding cells less than 30% (i.e. approximately the lower 1/3 of valleys; originally: <50%) and (iii) cells with curvature values less than 0 (unchanged). These modifications improved predictions, qualitatively assessed (i.e. a visual match up), of the thicket–shrubland treeline at the two calibration valleys. Most importantly, it was observed that frost only occurred in areas on the bottom of valleys, and in particular those with slope values less than 10° (the uppermost extent of the most extreme frosts). Visual matchups were conducted by generating surface maps of frost risk (a binary no risk or high risk), exporting these as a .kml file (kml package in R; Genolini et al., 2015) to Google Earth Pro, and comparing the CAP predictions with existing vegetation boundaries evident in the imagery (Fig. 5).

Testing the calibrated CAP model

We used two approaches to test the regionally-adjusted CAP model. First, we assessed its accuracy in predicting the location of frost exposed areas in relation to thicket–shrubland treeline in other valleys. Second, we tested the model predictions for a wide range of valleys against the productivity of the frost-sensitive spekboom cuttings planted in the TWP experiment (described below).

Vegetation boundaries in test valleys in the Eastern Cape

This regionally calibrated model correctly identified the thicket-shrubland treeline at five of the six test valleys (Figs. S1–S6). Where valleys were deeper relative to their width, and had a steeper change in gradient from the valley floors to valley slopes (i.e. Figs. S1–S6, excluding S3), the CAP (and thus frost presence) prediction aligned closely with the thicket–shrubland treeline. However, where the valley was shallower (relative to its width, that is Fig. S3), the model predicted a larger area covered by the CAP, and thus frost prone area, which extended ~50 m further upslope than the observed treeline between livestock degraded thicket and Karoo shrubland. Nonetheless, in such cases this minor overprediction (false positives) will still assist in landscape restoration planning to avoid planting in frost exposed areas. This predictive mapping of frost-risk (via cold-air pool mapping) may be considered somewhat simplistic as it is based on topographic features alone. Nonetheless, this approach does capture the areas in the landscape that would likely be prone to frost, especially in more mountainous or valley areas such as the Zuurberg mountain range. In addition, physics-based weather modelling has not been shown to have greater accuracy at predicting frost events (Pagès, Pepin & Miró, 2017).

In geomorphologically simpler terrain, such as the wide flatland areas found between the Great Escarpment and the Cape Fold Mountains, the CAP model as it is used here (in terms of the thresholds for slope and broad landscape position) would not be suitable for predicting the occurrence of frost. In these areas, where spekboom can be locally dominant, weather station data and field observations would need to be collected to establish the areas within the landscape that frost occurs, and its intensity under different weather conditions.

The TWP experiment

Of the 70 TWPs selected for the analysis, 30 plots occurred in frost-prone areas and 40 were in frost-free areas. Within this subset, spekboom planted in frost-free areas had significantly higher levels (72.2%) of mean aboveground phytomass production relative to those situated in the frost-prone areas (U = 720, p = 0.046; Fig. 6).