Neural process for uncertainty-aware geospatial modeling

Uncertainty-aware geospatial modeling has been a long standing challenge in geographic information science. Conventional geostatistical methods, such as the kriging family of methods, are often used for uncertainty-aware geospatial analysis and modeling but tend to fall short in capturing complex spatial patterns. Spatially explicit deep-learning models have been shown to excel at capturing complex patterns in high-dimensional geospatial datasets, but they often require large amounts of training data, which can be difficult to obtain in practice. In this study, we discuss a meta learning-based approach, namely Neural Processes (NP), for uncertainty-aware spatial analysis of small geospatial datasets. By integrating the concepts of deep neural networks, meta-learning, and stochastic processes, NP can capture complex patterns in small datasets while effectively accounting for uncertainty. We adopt a recent development in NP family, the latent bottlenecked attentive neural process (LBANP), and apply it to groundwater availability mapping as a case study. We highlight the performance of the method with a comprehensive comparison with two commonly used deep learning-based Gaussian process regression methods, deep kernel learning and deep Gaussian process.