A fast tree-search algorithm for joint leak detection and localisation using surface-borne ultrasonic acoustic signals is developed through a wireless sensor network. Owing to environmental noise and multipath fading of ultrasonic signals, false sensor observations are frequent in the observation data. The problem is modelled as a Bayesian inference model and the maximum a posteriori solution is approximated through a tree-search structure. The algorithm initially divides the area into large cells and approximates the observation likelihood function over these large cells. In a tree structure, a large cell with high likelihood is divided into smaller cells and the tree is expanded until the required estimation precision is obtained. Simulation and experimental results reveal advantages of the proposed technique in terms of estimation error and convergence speed in comparison with other conventional Bayesian techniques such as particle filtering.