Autonomous underwater vehicles (AUVs) are robotic platforms that are commonly used to map the sea floor, for example for benthic surveys or for naval mine countermeasures (MCM) operations. AUVs create an acoustic image of the survey area, such that objects on the seabed can be identified and, in the case of MCM, mines can be found and disposed of. The common method for creating such seabed maps is to run a lawnmower survey, which is a standard method in coverage path planning. We are interested in exploring alternate techniques for surveying areas of interest, in order to reduce mission time or assess feasible actions, such as finding a safe path through a hazardous region. In this paper, we use Gaussian Process regression to build models of seabed complexity data, obtained through lawnmower surveys. We evaluate several commonly used kernels to assess their modeling performance, which includes modeling discontinuities in the data. Our results show that an additive Mat\'ern kernel is most suitable for modeling seabed complexity data. On top of the GP model, we use adaptations of two standard path planning methods, A* and RRT*, to find safe paths for marine vessels through the modeled areas. We evaluate the planned paths and also run a vehicle dynamics simulator to assess potential performance by a marine vessel.