Invasive plants pose serious threats to economic, social and environmental interests throughout the world. Developing strategies for their management requires a range of information that is often impractical to collect from ground based surveys. In other cases, such as retrospective analyses of historical invasion rates and patterns, data is rarely, if ever, available from such surveys. Instead, historical archives of remotely sensed imagery provides one of the only existing records, and are used in this research to determine invasion rates and reconstruct invasion patterns of a ca 70 year old exotic mesquite population (Leguminoseae: Prosopis spp.) in the Pilbara Region of Western Australia, thereby helping to identify ways to reduce spread and infill. A model was then developed using this, and other, information to predict which parts of the Pilbara are most a risk. This information can assist in identifying areas requiring the most vigilant intervention and pre-emptive measures. Precise information of the location and areal extent of an invasive species is also crucial for land managers and policy makers for crafting management strategies aimed at control, confinement or eradication of some or all of the population. Therefore, the third component of this research was to develop and test high spectral and spatial resolution airborne imagery as a potential monitoring tool for tracking changes at various intervals and quantifying the effectiveness of management strategies adopted. To this end, high spatial resolution digital multispectral imagery (4 channels, 1 m spatial resolution) and hyperspectral imagery (126 channels, 3 m spatial resolution) was acquired and compared for its potential for distinguishing mesquite from coexisting species and land covers. / These three modules of research are summarised hereafter. To examine the rates and patterns of mesquite invasion through space and time, canopies were extracted from a temporal series of panchromatic aerial photography over an area of 450 ha using unsupervised classification. Non-mesquite trees and shrubs were not discernible from mesquite using this imagery (or technique) and so were masked out using an image acquired prior to invasion. The accuracy of the mesquite extractions were corroborated in the field and found to be high (R2 = 0.98, P<0.001); however, accuracy varied between classes (R2 = 0.55 to 0.95). Additional sampling may be required in some of the wider class intervals, particularly the moderate density class (30 to 90%) as sampling frequency was poor within the range of 60 to 90%. This is a direct result of there being relatively few quadrats available to be randomly selected in this class. That is, quadrats with between 60-90% cover were only evident in 4% of the test area. A more robust approach would, therefore, be to split this class into two (e.g. 30-60% and 60-90%) and select an additional 15 quadrats in the 60-90% range. The resolution of the imagery (1.4 m) precluded mapping shrubs smaller than 3 m2. Rates and patterns were compared to mesquite invasions in its native range. / It was determined that: (i) the shift from grass to mesquite domination had been rapid, with rates of increase in canopy cover comparable to invasive populations where it is native; (ii) rate of patch recruitment was high in all land types (stony flats, red-loamy soils and the riparian zone), but patch expansion and coalescence primarily occurred over the riparian zone and redloamy soils; (iii) mesquite had been spread by sheep and macropods and the recent switch to cattle is likely to exacerbate spread as it is a far more effective dispersal vector; and (iv) early successional patterns, such as high patch initiation followed by coalescence of existing stands are similar to where mesquite is native, but patch mortality did not occur. A knowledge based model was used to predict which parts of the Pilbara region are most at risk. Several limitations of models often employed in predicting suitability ranges of invasive plants were identified and include: (i) an inability to incorporate the notion that within a suitability range there is likely to be a scale of favourability; (ii) an inability to assign greater importance to evidence that is likely to have more importance in defining the areas suitable for invasion; and (iii) an inability to control the level of conservatism in the final results. These three shortcomings were mitigated through the use of: (i) fuzzy membership functions to derive a range of favourability from poor to best; (ii) pairwise comparison to derive higher weights for layers perceived to be more important and vice versa; and (iii) the use of ordered weighted averaging to directly control the level of conservatism (or risk) inherent in the models produced. / Based on the outcomes of the historical reconstruction of spatial rates and patterns, data sources included land types, land use, and the derivation of a steady state wetness index from spot height data. Model outputs were evaluated using two methods: the area under the curves (AUC) produced from relative operating characteristic (ROC) plots and by the maximum Kappa procedure. Both techniques agreed that the model most representative of the validation data was the one assuming the most risk. To create a Boolean output representing areas suitable/not suitable for invasion, optimal cut-points were derived using the point closest to the top left hand corner of the ROC plot and by the maximum Kappa method. Both methods obtained identical cut-points, but it is argued that the coefficient produced by the maximum Kappa method is more easily interpreted. The highest AUC was found to be 0.87 and, based on the maximum Kappa method, can be described as good to very good agreement with the validation records used. Digital multispectral imagery (DMSI), acquired in the visible and near infrared portions of the spectrum (3 visible bands, 1 near infrared) with a spatial resolution of 1 m and hyperspectral imagery (126 bands, 3 m spatial resolution) was acquired to assess the potential of developing a reliable and repeatable mapping tool to facilitate the monitoring of spread and the effects of control efforts. Woody vegetation was extracted from the images using unsupervised classification and grouped into patches based on contiguity. Various statistics (e.g. maximum, minimum, median, mean, standard deviation, majority and variety) were assigned to these patches to garner more information for species separation. / These statistics were explored for their ability to separate mesquite from coexisting species using Tukey’s Honestly Significantly Different (HSD) test and, to reduce redundancy, followed by linear discriminant analysis. Two approaches were taken to select the patch statistics offering the best discrimination. The first approach selected patch statistics that best discriminated all species (named “overall separation”). This was compared to a second approach, which selected the best patch statistics that separated each species from mesquite on a pairwise basis (named “pairwise separation”). The statistics offering the best discrimination were used as input in an Artificial Neural Network (ANN) to assign class labels. An incremental cover evaluation, whereby producer’s accuracy was computed from mesquite patches grouped into various size-classes, showed that identification of mesquite patches smaller than 36 m2 was relatively low (43-51%) regardless of the method used for choosing between the patch statistics or image type. Accuracy improved for patches >36 m2 (66-94%) with both approaches and image types. However, both approaches used on the hyperspectral imagery were more reliable at capturing patches >36 m2 than the DMSI using either approach. The lowest omission and commission rates were obtained using pairwise separation on the hyperspectral imagery, which was significantly more accurate than DMSI using an overall separation approach (Z=2.78, P<0.05), but no significant differences were found between pairwise separation used on either media. / Consequently, all methods and imagery types, except for DMSI processed using overall separation, are capable of accurately mapping mesquite patches >36 m2. However, hyperspectral imagery processed using pairwise separation appears to be superior, even though not statistically different to hyperspectral imagery processed using overall separation or DMSI processed using pairwise separation at the 95% confidence level. Mapping smaller patches may require the use of very high spatial resolution imagery, such as that achievable from unmanned airborne vehicles, coupled with a hyperspectral instrument. Alternatively, management may continue to rely on visual airborne surveys flown at low altitude and speed, which have proven to be capable at mapping small and isolated mesquite shrubs in the study area used in this research.

Invasive Prosopis trees pose significant threats to biodiversity and ecosystem services in the Northern Cape Province of South Africa. Several estimates have been made of the spatial extent of alien plant invasion in South Africa. The South African Plant Invaders Atlas (SAPIA) suggested that about 10 million hectares of South Africa has been invaded. However, the rate and spatial extent of Prosopis invasion has never been accurately quantified. The objective of the study is to use Remote Sensing and Geographic Information System (GIS) techniques to: (i) reveal areas susceptible to future invasion, (ii) describe the current extent and densities of Prosopis, (iii) to reveal the spatial dynamics and (iv) establish the extent of fragmentation of the natural vegetation in the Northern Cape Province. Image classification products were generated using spectral analysis of seasonal profiles, various resolution image inputs, spectral indices and ancillary data. Classification approaches varied by scene and spatial resolution as well as application of the data. Coarse resolution imagery and field data were used to create a probability map estimating the area vulnerable to Prosopis invasion using relationships between actual Prosopis occurrence, spectral response, soils and terrain unit. Multi-temporal Landsat images and a 500m x 500m point grid enabled vector analysis and statistical data to quantify the change in distribution and density as well as the spatial dynamics of Prosopis since 1974. Fragmentation and change of natural vegetation was quantified using a combined cover density class, calculating patch density per unit (ha) for each biome The extent of Prosopis cover in the Northern Cape Province reached 1.473 million hectare or 4% of the total land area during 2007. The ability of the above mentioned Remote Sensing and GIS techniques to map the extent and densities of Prosopis in the Northern Cape Province of South Africa demonstrated a high degree of accuracy (72%). While neither the image classification nor the probability map can be considered as 100% accurate representations of Prosopis density and distribution, the products provide use full information on Prosopis distribution and are a first step towards generating more accurate products. For primary invasion management, these products and the association of a small area on a map with Prosopis plants and patches, mean that the management effort and resources are efficiently focused. Further studies using hyper-spectral image analysis are recommended to improve the classification accuracy of the spatial extent and density classes obtained in this study. / Thesis (M. Environmental Science)--North-West University, Potchefstroom Campus, 2010.

Invasive Prosopis trees pose significant threats to biodiversity and ecosystem services in the Northern Cape Province of South Africa. Several estimates have been made of the spatial extent of alien plant invasion in South Africa. The South African Plant Invaders Atlas (SAPIA) suggested that about 10 million hectares of South Africa has been invaded. However, the rate and spatial extent of Prosopis invasion has never been accurately quantified. The objective of the study is to use Remote Sensing and Geographic Information System (GIS) techniques to: (i) reveal areas susceptible to future invasion, (ii) describe the current extent and densities of Prosopis, (iii) to reveal the spatial dynamics and (iv) establish the extent of fragmentation of the natural vegetation in the Northern Cape Province. Image classification products were generated using spectral analysis of seasonal profiles, various resolution image inputs, spectral indices and ancillary data. Classification approaches varied by scene and spatial resolution as well as application of the data. Coarse resolution imagery and field data were used to create a probability map estimating the area vulnerable to Prosopis invasion using relationships between actual Prosopis occurrence, spectral response, soils and terrain unit. Multi-temporal Landsat images and a 500m x 500m point grid enabled vector analysis and statistical data to quantify the change in distribution and density as well as the spatial dynamics of Prosopis since 1974. Fragmentation and change of natural vegetation was quantified using a combined cover density class, calculating patch density per unit (ha) for each biome The extent of Prosopis cover in the Northern Cape Province reached 1.473 million hectare or 4% of the total land area during 2007. The ability of the above mentioned Remote Sensing and GIS techniques to map the extent and densities of Prosopis in the Northern Cape Province of South Africa demonstrated a high degree of accuracy (72%). While neither the image classification nor the probability map can be considered as 100% accurate representations of Prosopis density and distribution, the products provide use full information on Prosopis distribution and are a first step towards generating more accurate products. For primary invasion management, these products and the association of a small area on a map with Prosopis plants and patches, mean that the management effort and resources are efficiently focused. Further studies using hyper-spectral image analysis are recommended to improve the classification accuracy of the spatial extent and density classes obtained in this study. / Thesis (M. Environmental Science)--North-West University, Potchefstroom Campus, 2010.