We propose a modified particle swarm optimization (PSO) algorithm to identify multiple optima (top-k peaks) in descending order, rather than just a single optimum value. With advances in computer technology and robotics, autonomous machines are used in applications such as search and rescue after a disaster. Survivors typically have only a limited amount of time to live, so finding them and directing rescuers to them are both time-critical. One way of rescuing more survivors more quickly is to deploy a number of aerial drones in advance and, after the disaster, use them to scan the area and identify the locations where survivors are most likely to be found. Thus, we first model such a situation by using mixture bivariate normal distributions with randomized means and identify individual drones as particle agents. Then, we propose top-k PSO, which an extension of the conventional Clerk-Kennedy PSO algorithm, to locate the top k peaks efficiently with high probability by remembering a list of global optima and introducing a strategy to increase the diversity in swarms to improve exploration. We conducted extensive experiments to evaluate top-k PSO by comparing its results with those produced by the baseline methods, canonical PSO, Clerk-Kennedy PSO, and NichePSO. Our experimental results indicate that the proposed PSO can find multiple peaks with higher probabilities than the baseline methods in various environments.