AbstractInferring seafloor topography from gravity anomaly currently is the dominant method to obtain a global view of the oceans. Standard techniques rely on an approximate, linear relationship between topography and gravity, which is valid only if the local topography is smooth compared with the regional topography, so the estimation accuracy in the very rugged areas is low. Current methods can be improved by removing the linear approximation and estimating the topography through simulated annealing and by using gravity gradiometry that is more sensitive to topography at short wavelengths than the gravity anomaly. Simulated annealing is a global optimization technique that can process nonlinear inverse problems. It is developed to estimate the seafloor depths by minimizing the difference between the observed and forward‐computed vertical gravity gradients. The method is tested on altimetry‐derived gravity gradients in a 2∘×2∘ area of rugged seafloor topography in the West Pacific Ocean and results in estimates with a root‐mean‐square error of ±236 m. Compared to estimates from an existing model obtained by standard techniques this represents an accuracy improvement of 22%.