Sensitivity analysis in simulating oil slick using CA model

Abstract

Cellular automata (CA) model is one of effective models in geographic simulation due to its self-organization and "bottom-up" approach. In this paper, logistic regression method is used to obtain parameter values. And logistic regression CA model is constructed for simulating oil slick. After that, parameter values are calibrated using sampling ratios, spatial scales and neighborhood structures. Finally, the model is applied to the simulation of oil slick in DeepSpill project. Experiments showed that (1) higher sampling ratio will help to obtain better parameter values. However, when sampling ratio exceeds 10%, improvement is small in overall accuracy. In addition, suitable proportion of oil areas in training sample can help to get better results. It can be seen from parameter values that distance is the most important factor, followed by currents, wind, salinity and temperature. (2) Overall accuracy of simulation results has fluctuation characteristics in different spatial scales. In experiments of extended neighborhood, accuracy first increases and then decreases with increase of neighborhood size. This model yields the best simulation result in 7 × 7 Moore neighborhood, which can reach 97.40%. (3) Two simulation results are obtained by using 3 × 3 Moore neighborhood in spatial resolution of 2m (Result A) and 7 × 7 Moore neighborhood in spatial resolution of 6m (Result B). After comparison, both results have characteristics of diffusion and drift. Center point of verification image drifted 23.152m toward 92.8967°. In Result A, center point only drifted 7.087m toward 86.3767°. It has obvious deviation in the east and west because of short drift distance. However, center point drifted 18.599m toward 97.2276° in Result B. It has better performance in shape and indices (overall accuracy, Kappa and FoM).