Abstract:
|
Spatial processes observed in many applications, such as climate and environmental science, are often large-scale and exhibit spatial nonstationarity. Gaussian processes are widely used in spatial statistics to model such nonstationarity by specifying a nonstationary covariance function, such as the nonstationary Matern covariance. In literature, existing work relies on spatial region partitions to estimate the spatially varying parameters in the covariance function. Although the choice of partitions is a key factor, it is typically subjective and not data-driven. In this work, we exploit the capabilities of the Convolutional Neural Networks (CNNs) to perform dynamic splitting to the nonstationary spatial region. This dynamic splitting can identify the nonstationary subregions after the first split and recursively resplit them until all the spatial subregions behave close to stationary. We also provide a parallel high-performance implementation of the nonstationary modeling and predictions on most recent hardware architectures, including shared memory, GPUs, and distributed memory systems. Our proposed approach shows better accuracy and performance than the traditional methods.
|