Quantifying the performance of silvopastoralism for landslide erosion and sediment control in New Zealand’s hill country : a dissertation presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Physical Geography at Massey University, Palmerston North, New Zealand

Chapters 3, 4 and 5 have been republished with the permission of the publisher, Elsevier.Landslide erosion results in loss of productive soils and pasture. Moreover, sediment delivered to streams from landslides can contribute to the degradation of freshwater and marine receiving environments by smothering benthic habitats and increasing turbidity, light attenuation, and sediment-bound contaminants. Silvopastoralism is an important land management practice used to combat landslide erosion and improve the health of downstream aquatic ecosystems. Yet, the effectiveness of widely spaced trees in reducing shallow landslide erosion and sediment delivery at hillslope to catchment scales remains largely unknown. Previous studies have been limited by scale (e.g., hillslope) or method (e.g., univariate analyses). This research aims to develop spatially explicit modelling to assess the impact of differing tree species, planting densities, and individual tree location, on rainfall-triggered landslides and sediment delivery while accounting for varying environmental conditions, such as slope gradient, lithology, or soil type. As such, this thesis combines geospatial methods and statistical models to address key challenges related to erosion and sediment control in New Zealand’s pastoral hill country. First, using a study area in the Wairarapa, located in the southeast of the North Island, New Zealand (840 km2), a method was developed using open-source remote sensing products to generate high-resolution individual tree influence models for the dominant tree species. The objective was to generate a spatial explicit representation of individual trees for landscape-scaled statistical slope stability modelling. The combined hydrological and mechanical influence of trees on slopes was inferred through the spatial relationship between trees and landslide erosion. These spatial distribution models for individual trees of different vegetation types represent the average contribution to slope stability as a function of distance from tree at 1-m spatial resolution. The normalised models (0-1) largely agree with the shape and distribution of force from existing physical root reinforcement models. Of exotic tree species that were planted for erosion and sediment control, poplars (Populus spp.) and willows (Salix spp.) make up 51% (109,000) of trees located on hillslopes at a mean density of 3 trees/ha. In line with previous studies, poplars and willows have the greatest contribution to slope stability with an average maximum effective distance of 20 m. Yet, native kānuka (Kunzea spp.) is the most abundant woody vegetation species on hillslopes within the study area, with an average of 24 trees/ha, providing an important soil conservation function. A large proportion (56% or 212.5 km2) of erosion-prone terrain in the study area remains untreated. In a world-first, this allowed the influence of individual trees to be included in a statistical landslide susceptibility model using binary logistic regression to quantify the effectiveness of silvopastoral systems at reducing landslide erosion and to support targeted erosion mitigation. Models were trained and tested using a landslide inventory consisting of 43,000 landslide scars mapped across the study area. Model performance was very good, with a median Area Under the Receiver Operating Characteristic curve (AUROC) of 0.95 in the final model used for predictions, which equates to an accuracy of 88.7% using a cut-off of 0.5. The effect of highly skewed continuous tree influence models on the maximum likelihood estimator was tested using different sampling strategies aimed at reducing positive skewness. With an adequate sample size, highly skewed continuous predictor variables do not result in an inflation of effect size. Application of the landslide susceptibility model was illustrated using two farms from within the study area (Site 1: 1,700-ha; Site 2: 462-ha) by quantifying the reduction in shallow landslide erosion due to trees. Compared to a pasture only baseline, landslide erosion was reduced by 17% at Site 1 and 43% at Site 2 due to all existing vegetation. The effectiveness of individual trees in reducing landslide erosion was shown to be less a function of species than that of targeting highly susceptible areas with adequate plant densities. The excellent model performance means spatial predictions are precise, which has implications for land management as the maps provide greater certainty and spatial refinement to inform landslide mitigation. The terrain occupied by the “high” susceptible class – defined as the terrain where 80% of mapped landslides were triggered in the past – occupies only 12% of Site 1 and 7% of Site 2. This suggests there is great potential for improved targeting of erosion mitigation to these areas of the farms where landsliding may be expected in the future. To enable biological mitigation to be targeted to critical source areas of sediment, determinants of sediment connectivity were investigated for a landslide-triggering storm event in 1977. In a first of its kind, a morphometric landslide connectivity model was developed using lasso logistic regression to predict the likelihood of sediment delivery to streams following landslide initiation. An experiment was undertaken to explore a range of connectivity scenarios by defining a set of sinks and simulating varying rates of sediment generation during runoff events of increasing magnitude. Sediment delivery ratios for the 1977 event ranged from 0.21 to 0.29, equating to an event sediment yield of 3548 t km-2 to 9033 t km-2. The likelihood of sediment delivery was greatly enhanced where debris tails coalesce. Besides scar size variables, overland flow distance and vertical distance to sink were the most important morphometric predictors of connectivity. When scar size variables were removed from the connectivity model, median AUROC was reduced from 0.88 to 0.75. By coupling landslide susceptibility and connectivity predictions in a modular form, we quantified the cost effectiveness of targeted versus non-targeted approaches to shallow landslide mitigation. Targeted mitigation of landslide-derived sediment was found to be approximately an order of magnitude more cost-effective than a non-targeted approach. Compared with a pasture-only baseline, a 34% reduction in sediment delivery can be achieved by increasing slope stability through spaced tree planting on 6.5% of the pastoral land. In contrast, the maximum reduction achievable through comprehensive coverage of widely spaced planting is 56%. The coupled landslide susceptibility and connectivity predictions (maps) provide an objective basis to not only target mitigation to areas where future shallow landslides are likely to occur, but – perhaps more importantly – target future tree planting to locations that are likely to be future sources of fine sediment. In this way, the research presented in this thesis is both methodologically novel and has immediate application to support land management decisions aimed at creating a more sustainable socio-ecological landscape