Spatial statistics often rely on Gaussian processes (GPs) to capture dependencies across locations. However, their computational cost increases rapidly with the number of locations, potentially needing multiple hours even for moderate sample sizes. To address this, we propose using Semi-Implicit Variational Inference (SIVI), a highly flexible Bayesian approximation method, for scalable Bayesian spatial interpolation. We evaluated SIVI with a GP prior and a Nearest-Neighbour Gaussian Process (NNGP) prior compared to Automatic Differentiation Variational Inference (ADVI), Pathfinder, and Hamiltonian Monte Carlo (HMC), the reference method in spatial statistics. Methods were compared based on their predictive ability measured by the CRPS, the interval score, and the negative log-predictive density across 50 replicates for both Gaussian and Poisson outcomes. SIVI-based methods achieved similar results to HMC, while being drastically faster. On average, for the Poisson scenario with 500 training locations, SIVI reduced the computational time from roughly 6 hours for HMC to 130 seconds. Furthermore, SIVI-NNGP analyzed a simulated land surface temperature dataset of 150,000 locations while estimating all unknown model parameters in under two minutes. These results highlight the potential of SIVI as a flexible and scalable inference technique in spatial statistics.
翻译:暂无翻译