Herbivory and misidentification of target habitat constrain region-wide restoration success of spekboom (Portulacaria afra) in South African subtropical succulent thicket

Restoration of degraded subtropical succulent thicket, via the planting of Portulacaria afra (spekboom) truncheons, is the focus of a public works programme funded by the South African government. The goals of the programme, which started in 2004, are to create jobs, sequester carbon, restore biodiversity, reduce erosion, improve soil water holding capacity and catalyse private sector investment for upscaling of restoration. Here we report on a region-wide experiment to identify factors that can improve project success. Measures of success were survivorship and annual aboveground biomass carbon sequestration (ABCsr) of spekboom truncheons some 33–57 months after planting—starting in March 2008—into 173 fenced plots (0.25 ha) located throughout the global extent of spekboom thicket vegetation. We also collected data for 18 explanatory variables under the control of managers, and an additional 39 variables reflecting soil physical and chemical characteristics and rainfall patterns post restoration, all beyond the influence of managers. Since the latter covariates were available for only 83 plots, we analysed the two data sets separately. We used a prediction rule ensemble to determine the most important predictors of restoration success. There was great variation in percentage survivorship (median = 24, range = 0–100%) and ABCsr (median = 0.009, range = 0–0.38 t C ha−1 yr−1). The model using management variables explained less variance (53%) in survivorship than the model incorporating additional soil and rainfall covariates (62%). ABCsr models were better fits (78 and 88% variance explained, respectively). All model configurations identified browse intensity as a highly influential predictor of restoration success. Predicted success was highest for plots located in target habitat; however, only 45% were thus located, suggesting the need for expert input and habitat modelling for improving target habitat identification. Frost exposure was another important predictor influencing all models but was likely a consequence of locating sites off target habitat. Sites planted on equatorward slopes during the warm season showed reduced carbon sequestration, possibly due to elevated soil moisture stress associated with high radiation loads. Physiographic factors associated with improved restoration success were location on sloping ground (reduced frost exposure), increasing longitude (more warm-season rainfall) and increasing latitude (less frost coastwards). Few trends were evident among post-restoration climatic factors beyond the control of managers. Higher rainfall during the year post restoration had a negative impact on carbon sequestration while higher rain during the early months post restoration had a positive effect on both carbon sequestration and survivorship. Soil factors showed little importance for the survivorship model, whereas silt content, % K and Mg CEC emerged as predictors of carbon sequestration. Our results have direct relevance for improving the success of landscape-scale restoration projects envisioned for the ca. 8,930 km2 of degraded spekboom thicket.


Introduction
Both reviewers questioned the absence of a number of variables in the models presented in our original manuscript. Reviewer 2 suggested the use of a range of topographic variables derived from digital elevation models, while Reviewer 1 questioned why rainfall variables of 1 and 3 months preceding the planting were not used in the original models. Here we show the results of a rerun of our models with these suggested variables included.
The list of variables added (see Table 1) were added to the original variables (see Table 1 in the submitted manuscript) in this additional analysis. numerical TRI Terrain Ruggedness Index -is the mean of the absolute differences between the value of a cell and the value of its 8 surrounding cells (5 ). numerical roughness the difference between the maximum and the minimum value of a cell and its 8 surrounding cells (5 ).
numerical hillS hill shade from slope and aspect layers (both in radians) (2 ). numerical aspect2 Calculated from ASTER DEM with the Horn (2 ) algorithm, considered best for rough surfaces (radians) numerical slope2 Calculated from ASTER DEM with the Horn (2 ) algorithm, considered best for rough surfaces (radians).

numerical planform
The planform curvature is the second derivative(s) of the elevation surface (slope of the slope) and perpendicular to the direction of the maximum slope (6 ). numerical

McNab
McNab's variant of the surface curvature (concavity/convexity) index (3 , 4 ). numerical mmtpre1|3 Cumulative rainfall 1|3 months before plot was planted, including month planted numerical The new models presented here (see Table 1, original model names appended with "b"), replicated from those presented in our manuscript. We added the newly generated variables outlined in Table 1 to those outlined in Table S1 in the original manuscript. To our original models originally fitted with variables related to topographic and management factors, an additional nine topographic variables were added for both survival and carbon sequestration as response variables.
To the orignal models fitted with all available variables, these newly derived topographic variables were also added, together with the two additional rainfall variables (i.e. total rainfall one and three months before planting respectively).

Results
A comparison of the variance explained between the original and new models with extra variables ( Table 1 in the original manuscript vs Table 2 Table 2). Since there is little improvement in the variance explained between the models fitted in the original manuscript and those presented here with the added suggested predictors, the new models are not more accurate than those originally fitted.
The derived topographic variables added as suggested by Reviewer 1 did not significantly change the accuracy of the models, although it may have identified new variables that could be used interchangeably with some those selected originally. Two of the new variables (flow direction and planform curvature) could be useful for future potential mapping of target habitat before planting, but more research is needed in this regard. These variables are certainly useful in technical habitat modeling, but requires digital processing from satelite derived data.
Since the aim of our paper was to identify variables easily identified in the field by restoration practitioners, we did not include the new variables in our models as they do not provide better results than our field-derived variables.

Conclusion
The fitting of additonal new derived variables as suggested by the reviewers did not produce better models or better predictor variables for either survival or carbon sequestration than those identified in the original models and manuscript.