Research was undertaken in the indigenous tussock grasslands of South Island of New Zealand in order to quantify past rates of conversion to agricultural land use and to develop vulnerability models to predict future conversion spatially and temporally. The study area was delineated using the median spectral reflectance of indigenous grasslands and included the largest extent of unprotected contiguous grasslands concentrated in the central South Island. Conversion from indigenous grasslands to a non-indigenous cover was quantified by comparative mapping over three intervals (1840-1990, 1990-2001, and 2001-2008). The basic premise in using satellite imagery to detect changes in land-use/cover is that these are revealed by changes in spectral signature. However, New Zealand’s indigenous and non-indigenous grasslands have overlapping spectral trajectories and high inter-annual variability, therefore contextual information was needed in order accurately map conversion from indigenous grassland cover to exotic pasture. Within the study area around the time of European settlement (1840) there were approximately 3.3 million hectares of indigenous grasslands. Between 1840 and 1990 around 1 million hectares of indigenous grasslands were converted to a non-indigenous cover. The extent of conversion during the preceding time period (1990-2008) was approximately 71,261 ha, of which 72% was converted to pasture and cropland and the remaining 28% to mining, urban settlements and exotic forestry. Although the overall rate of grassland conversion decreased relative to the period of European settlement and 1990, the proportion of remaining indigenous grasslands converted each year increased. Almost two-thirds of post-1990 conversion has occurred in environments with less than 30% indigenous cover remaining, and much is in land classified as non-arable with moderate to extreme limitations to crop, pasture and forestry growth. To assess the relative vulnerability of remaining areas of indigenous grassland to intensive land use (mainly intensive pasture production but also exotic conifer plantations, urban use and mining), spatial predictions using Generalized Additive Models (GAMs) were used to establish relationships between two different types of dependent (response) variables (presence or absence of conversion) and potential environmental and proxy socio-economic explanatory variables. The chosen predictors for the final model were used to map conversion probabilities in geographic space. The selected GAMs showed the mean probability of conversion in remaining indigenous grasslands was 0.15 and the mean area of conversion was 116 ha. Habitat that was most vulnerable to conversion was at moderate elevations and on medium slopes, and had previously been classified as being of low suitability for production. To interpret the regression models, plots of the partial response curves resulting from the model, and overall contributions of variables to the model, were used. The most important explanatory variables for predicting the probability of conversion in order of ‘alone contribution (the potential for each variable alone to explain conversion) was slope, rainfall, land tenure, distance to roads, proximity to existing agricultural, regional council, and mean annual temperature. Interpretation of the GAMs showed that conversion was negatively related to: slope, rainfall and distance roads; positively related to mean annual temperature; higher in the Otago and Canterbury regions and on privately owned or recently privatized lands, and peaked at intermediate proximity to roads. The prediction of the probability of conversion model was cross-validated both spatially and temporally. Temporal cross-validation compared predicted probabilities of conversion against reference maps of observed ‘current’ conversion. Spatial cross-validation evaluated model discrimination between ‘converted’ and ‘not converted’. Temporal and spatial performance was measured using the Receiver Operating Characteristic (ROC), a graphical plot of the true positive rate (sensitivity) as a function of the false positive (1-specificity) for different probability thresholds. For temporal cross-validation there was high correlation between ‘predicted’ and ‘observed’ (ROC = 0.913), and for spatial validation the relationship between the fitted and observed was also high (ROC= 0.921), indicating there was good discrimination between ‘converted’ and ‘not converted’. Integrating validated estimates of the probability of conversion (vulnerability) into conservation planning tools is an important component of conservation planning. Comparison of conservation prioritisation outputs with validated estimates of vulnerability of New Zealand’s indigenous grasslands showed variable effectiveness of vulnerability surrogates; one surrogate performed most poorly where vulnerability of grasslands to conversion was greatest and realized probability of protection was lowest. Furthermore, estimates of vulnerability using surrogates underestimated vulnerability on flat land that was closer to roads and overestimated areas on steeper land that was topographically invulnerable to conversion. There is an increased disparity between patterns of protection and patterns of conversion indicating that existing conservation planning tools are not effectively targeting the most vulnerable areas of remaining indigenous grasslands. An up-to-date validated vulnerability assessment offered a practical and a responsive technical bridge for the gap between science and implementation. This approach can be applied more widely to provide national models of vulnerability from representative samples of conversion.