Predictive Mapping of Blackberry in the Condamine Catchment Using Logistic Regression and Spatial Analysis

The development of control strategies for noxious weeds depends on reliable information about the location and extent of weed species. Consequently, there is a need to develop mapping and monitoring techniques that are accurate, costeffective and reliable. This paper investigated predictive modelling and mapping techniques for blackberry (Rubus fruticosus agg.) weed in the Condamine Catchment. In all, 19 bio-physical factors were assessed, of which a subset was analysed by logistic regression using SPSS. The model calculated the probability of a binary dependent variable (i.e. “presence of weed” vs. “absence of weed”) in response to the above independent (bio-physical) variables. The output model was brought into ArcGIS’s Spatial Analyst to produce the predictive map. The factors found to be significant in the model were a) distance from stream, b) foliage projective cover, c) elevation, and d) distance from NSW border. The use of logistic regression generated maps depicting the probability of blackberry occurrence with a model accuracy of greater than 90%. The predicted maps offer relevant information that could be useful to land planners and decision-makers on where to target or prioritise weed control strategies, or for other aspects of weed management. Introduction Weeds cause serious environmental problems, economic losses and social impacts. In Australia, the cost to the economy from the agricultural impacts of weeds is approximately $4 billion per annum (Sinden et al., 2005; Natural Resource Management Ministerial Council, 2006). The cost to nature conservation, tourism and landscape amenity, although not quantified by a formal study, is thought to be of similar magnitude. Thus, weeds and their impacts should be managed sensibly, and the prevention and control strategies must be formulated comprehensively with the aid of reliable information. Geographic information systems (GIS) offer useful tools to support weed control and management (e.g. Gillham et al., 2004; Main et al., 2004; Prather and Callihan, 1993). Because weeds grow in a given geographical location and their spread is influenced by environmental factors that vary spatially, the use of GIS for weed mapping, assessment and monitoring has increased significantly in the last 10 years. The various studies and applications may be categorised into four broad aspects: a) predicting weed invasion, b) mapping and monitoring, c) educating the public (e.g. through 3D visualisations), and d) development of management plans (USDA Forest Service RSAC, 2007). Despite the numerous studies on predictive mapping of weeds, it is apparent from the literature that no single model can be used to predict weed invasion for every occasion. This supports the notion that there is no “one-size-fits-all” model for predicting weed invasion. As there are many hundreds of weeds worldwide (e.g. 287 species are listed in the Australia’s weed identification tool at http://www.weeds.org.au/) which have diverse habitat characteristics, it will be prudent to adapt a model or develop a new one, for a particular weed species in a given area. In this context, this study focused on predictive mapping of blackberry using logistic regression analysis. Logistic regression can be used to describe the relationship of several independent variables to a dichotomous dependent variable such as “presence” and “absence” (Hosmer and Lemeshow, 2000), as is the case for most weeds. Although several studies pertaining to predictive weed mapping have been conducted in Australia (see for example, Crossman and Bass (2008); Dunlop et al., 2006; Chiou and Yu, 2005), no study has been reported on predictive mapping of blackberry (in contrast to detection and mapping using remote sensing, e.g. Dehaan, et al., 2007; Ullah, et al., 1989). While there is nothing particularly novel about using GIS and logistic regression for spatial analysis of weed species, their applications for blackberry mapping have never been exploited. Hence, the objectives of this study were to: a) to characterise the habitat of blackberry in a given area, and b) to develop a spatially explicit predictive model of blackberry using logistic regression and GIS. Queensland Spatial Conference 2008 – 17-19 July 2008, Gold Coast Global Warning: What’s Happening In Paradise? Paper No. 0043 2 Blackberry and Predictive Modelling Using Logistic Regression Blackberry (Rubus fruticosus agg.) is a perennial, semi-deciduous shrub with prickly stems (canes), often forming mounds (Figure 1) (Evans et al., 2007). First year canes are mostly reddish-purple where exposed, and green when not exposed. They may be with or without hairs. Its leaf mostly consists of five distinct leaflets. Flowers are white or pink, while the fruits are initially green and become red to black when ripe. Blackberry can reproduce by both seed and vegetative means. Birds, mammals and water easily disperse the fruits, while canes in contact with the ground, can root and produce new plants. Blackberry can be found throughout temperate Australia, in areas where the annual rainfall exceeds 700 mm. In Queensland, it occurs in the Stanthorpe, Warwick, Killarney and Toowoomba areas (NRW, 2006a). Figure 1. Sample photographs of blackberry collected in July 2007, Warwick and Cambooya area. In Australia, blackberry has been declared a weed of national significance (WONS) (Thorp and Lynch, 2000). It is ranked 3rd out of the 20 species included in the WONS’s inaugural list, based on its invasiveness, impacts, potential for spread, and socioeconomic and environmental impacts. The weed colonises pasture, orchards and forest lands and often forms impenetrable thickets and mounds that prevent access by animals and people for productive or recreational use. Thus, strategies for the control and prevention of blackberry should be a part of sound land management. Predicting weed invasion is an important task for weed management. The development of preventive and control strategies depends on information about the weed’s current location and potential invasion. Consequently, mapping and spatial modelling efforts to predict weed invasion (also called “susceptibility mapping” or “risk mapping”) have increased in recent years. A wide array of modelling techniques has been developed and applied to specific weeds and localities including: a) linear or weighted linear combination modelling (e.g. Gillham et al., 2004), b) use of genetic algorithm (e.g. Raimundo et al., 2007), c) ecological niche modeling (Peterson et al., 2003), d) rule-base fuzzy logic (e.g. Chiou and Yu, 2005), and e) logistic regression (e.g. Collingham et al., 2000). Logistic regression is a model that can be used to predict the probability of occurrence of an event as a function of the independent variables. It is useful when the observed outcome is restricted to two values, which usually represent the occurrence or non-occurrence (usually coded as 1 or 0, respectively) of an outcome event (Hosmer and Lemeshow, 2000). Logistic regression does not assume that the relationship between the independent variables and the dependent variable is linear, and nor does it assume that the dependent variable is normally distributed. Other modelling approaches for dichotomous dependent variables are also possible, but logistic regression is by far the most popular (Kleinbaum and Klein, 2002). This is because the model provides an estimate that must lie in the range between zero and one, and the S-shape of the logistic function is considered to be widely applicable in many multivariate analyses. Queensland Spatial Conference 2008 – 17-19 July 2008, Gold Coast Global Warning: What’s Happening In Paradise? Paper No. 0043 3 Methods The Study Area and Data Acquisition The study area covers the Condamine Catchment (Figure 2). It is located west of the Great Dividing Range in southern Queensland, covering an area of 24,434 km. The area has a highly variable subtropical climate, with average annual rainfall of 682-955-mm, and average temperatures ranging from 3°C to 30°C (NRW, 2006b). The vegetation in the basalt hills is dominated by mountain coolibah, narrow-leaved ironbark and silver leaf ironbark. In soils associated with sandstone areas, patches of brigalow/belah, poplar box, ironbark, bulloak and cypress pine are common. The extensive use of the area for agriculture has resulted in the loss of much of the original vegetation (<30% remains) (NRW, 2006b). Datasets relating to the location of blackberry were obtained from the Environmental Protection Agency’s (EPA) WildNet database (EPA, 2007) and the Department of Natural Resources and Water (NRW) computer-based pest management system, PestInfo (NRW, 2007). Additional data points were also collected during the field reconnaissance survey in the Warwick and Cambooya shires, where blackberry patches are known to occur. Using GIS, the spatial distribution of the weed was tested for patterns of clustering, dispersion, or randomness over the catchment study area. The Average Nearest Neighbour Distance tool in ArcGIS calculated a nearest neighbour index based on the average distance from each feature to its nearest neighbouring feature. For thematic layers representing the habitat of blackberry or the relevant site factors of the area, several digital maps were acquired from different sources (Table 1). The selection of these site factors was based from the literature (e.g. Evans et al., 2007; CRC for Australian Weed Management, 2003) and interviews with local pest management officers (e.g. Dight, personal communication, 20 July 2007). Table 1. Spatial datasets acquired for the study Map of Site Factors Primary Data Source Key Processing to Generate Dataset 1. Distance from road Road Map (PSMA) Buffering in raster (Euclidean distance) 2. Distance from stream Drainage Map (GA) Buffering in raster (Euclidean distance) 3. Distance from water body (dams and lakes) Water Body (NRW) Buffering in raster (Euclidean distance) 4. Distance from NSW border Outline Map of Australia (GA) Select feature; buffering in raster (Euclidean distance) 5. Foliage Projective Cover (FPC) Foliage Projective Cover (NRW) None 6. Distance from FPC Value of >= 40% Foliage Projective Cover (NRW) Select feature; buffering in raster (Euclidean distance) 7. Distance from FPC Value of <= 12% Foliage Projective Cover (NRW) Select feature; buffering in raster (Euclidean distance) 8. Elevation DEM (NRW) Fill “sink” cells 9. Slope DEM (NRW) Calculate slope in percent 10. Aspect DEM (NRW) Calculate aspect 11. Landform DEM (NRW) Calculate using a third-party ArcView extension 12. Soils Soil Map (CSIRO) None 13. Rainfall Rainfall (BoM) Reclassify in raster 14. Geology Geology (GA) None 15. Land cover Land cover (NRW) None 16. Land use (Level 1) Land use (NRW) Select feature 17. Land use (Level 2) Land use (NRW) Select feature 18. Tenure DCDB (NRW) Select feature 19. Regional ecosystems Regional ecosystems (EPA) Select feature PSMA–Public Sector Mapping Agencies, GA–Geoscience Australia, NRW–Queensland Department of Natural Resources and Water, EPA–Queensland Environmental Protection Agency, CSIRO–Commonwealth Scientific and Industrial Research Organisation, DEM– Digital Elevation Model, DCDB–Digital Cadastral Data Base, BoM–Bureau of Meteorology Data Pre-processing and Analysis Since many datasets were received from disparate sources, in various formats and extent, the pre-processing of the data was essential. The pre-processing tasks included projection and coordinate transformation, clipping, vector-to-raster conversion, masking, feature selection, buffering, grid resampling, reclassification, etc. The datasets were processed within ArcGIS 9.2 using the Spatial Analyst extension. The pixel size was set to 25m x 25m, with 1:100,000 as the approximate mapping scale. Most of the thematic maps (e.g. FPC, land cover, land use, DEM, etc.) used in the predictive mapping were captured at this map scale. Queensland Spatial Conference 2008 – 17-19 July 2008, Gold Coast Global Warning: What’s Happening In Paradise? Paper No. 0043 4 Figure 2. Location map of the study area for blackberry predictive mapping. The spatial datasets and field data were “intersected” (using Spatial Analyst’s “sample” operation) to produce a table that showed the grid cell value of each attribute for each sample weed location (Figure 3). This produced a data array of 911 samples x 19 variables or attributes. Frequency tables for each variable were generated to enable the analysis of the habitat characteristics of blackberry. A total of 2,277 randomly selected sample points from the “no weed” and “weed” pool were selected. Out of this, 1,592 samples (70% of the total) were used as a training set for the analysis. The remaining 685 samples (30% of the total) were used as a test or validation set. For the training set samples, the attributes of independent variables were compiled to create a data array in Microsoft Excel. The number of independent variables was reduced to 10, based on a) results of the frequency table analysis, b) interpretation of relevant literature, and c) constraint in data processing due to scale of measurement (i.e. categorical data for land use, geology, etc). SPSS software had the ability to handle categorical data for logistic regression, but the number of classes for categorical variables was too large to warrant a sensible approach. Queensland Spatial Conference 2008 – 17-19 July 2008, Gold Coast Global Warning: What’s Happening In Paradise? Paper No. 0043 5 Field Data of Blackberry (point vector layer) Predictive Map of Blackberry Intersect Spatial Datasets (19 thematic raster layers)