SINKHOLE PROBABILITY MAP USING BINARY LOGISTIC REGRESSION COUPLED WITH GIS: A CASE STUDY IN SOUTHWEST MISSOURI

Southwest Missouri has a karst region and several sinkhole collapses had been reported in the past decade. The objective of this study was to create a sinkhole probability map using logistic regression coupled with GIS to identify areas susceptible to sinkhole collapse for efficient land management and decision making for urban planners. Logistic regression model is one of the most common regression models used to predict latent probability of a binary qualitative outcome as a function of a number of explanatory variables. In this study, logistic regression was used to establish a functional relationship between dichotomously observed sinkhole locations (absence or presence of sinkholes) and various factors (predictors) which are expected to play a role in the development or collapse of sinkholes.

The sample for this study consisted of 998 random data points around Nixa, southwest Missouri, 217 of which were observed sinkhole points and 781 were non-sinkhole points. Among various factors potentially responsible for sinkhole development and collapse, distance to road, slope of the surface, distance to a spring, distance to lineaments, and static water levels (SWL) were considered in the logistic model (Log Likelihood χ²=114.9, p<.001). All distance and SWL measures were in 100 ft units. Preliminary findings suggested that, whereas SWL and distance to springs were both positive correlates to the odds of sinkhole occurrence, only SWL significantly increased the odds (OR=1.688, p<.01). In addition to improving the Pseudo R² by 2 percentage points, including SWL in the model demonstrates that the seemingly positive correlation between distance to a spring and sinkhole occurrence might be in effect spurious. Distance to road (OR=0.946, p<.01), slope of the surface (OR=0.888, p<.001), and distance to lineaments (OR=0.973, p<.001) all negatively contributed to the odds of sinkhole occurrence. Using the estimates of logistic regression, predicted probabilities of sinkhole occurrence were obtained and mapped.

Further calibration of the model that incorporates additional factors such as distance to nearest sinkhole as well as higher-order terms using data for extended sample regions is being tested to ensure the robustness of the estimates and improved accuracy of predictions.