This paper focus on predicting the number of injured people in the aftermath of an earthquake, applying Gaussian process (also known as Kriging), which is a well-known method for machine learning. Different from neural networks, Gaussian process estimates also the variance of the predictor. The predicted injuries and their variances are used in a chance-constrained optimization problem to determine shelter locations from a given set of candidate locations. The resulting problem is a mixed-integer nonlinear programming optimization, which can be solved through commercial software in a reasonable amount of time for small scale problems.