Abstract:Spatial prediction is to predict the values of the targeted variable, such as PM2.5 values and temperature, at arbitrary locations based on the collected geospatial data. It greatly affects the key research topics in geoscience in terms of obtaining heterogeneous spatial information (e.g., soil conditions, precipitation rates, wheat yields) for geographic modeling and decision-making at local, regional, and global scales. In-situ data, collected by ground-level in-situ sensors, and remote sensing data, collected by satellite or aircraft, are two important data sources for this task. In-situ data are relatively accurate while sparse and unevenly distributed. Remote sensing data cover large spatial areas but are coarse with low spatiotemporal resolution and prone to interference. How to synergize the complementary strength of these two data types is still a grand challenge. Moreover, it is difficult to model the unknown spatial predictive mapping while handling the trade-off between spatial autocorrelation and heterogeneity. Third, representing spatial relations without substantial information loss is also a critical issue. To address these challenges, we propose a novel Heterogeneous Self-supervised Spatial Prediction (HSSP) framework that synergizes multi-source data by minimizing the inconsistency between in-situ and remote sensing observations. We propose a new deep geometric spatial interpolation model as the prediction backbone that automatically interpolates the values of the targeted variable at unknown locations based on existing observations by taking into account both distance and orientation information. Our proposed interpolator is proven to both be the general form of popular interpolation methods and preserve spatial information. The spatial prediction is enhanced by a novel error-compensation framework to capture the prediction inconsistency due to spatial heterogeneity. Extensive experiments have been conducted on real-world datasets and demonstrated our model's superiority in performance over state-of-the-art models.

On Cokriging, Neural Networks, and Spatial Blind Source Separation for Multivariate Spatial Prediction

Nonlinear blind source separation exploiting spatial nonstationarity

Bayesian Blind Source Separation for Data with Network Structure

Visual Parameter Selection for Spatial Blind Source Separation

Sparsity and adaptivity for the blind separation of partially correlated sources

An Adaptive Approach to Blind Source Separation Using a Self-Organzing Map and a Neural Gas

A novel Bayesian blind source separation approach for extracting non-stationary and discontinuous components from structural health monitoring data

Estimating the prediction performance of spatial models via spatial k-fold cross validation

Comparison of Kriging and artificial neural network models for the prediction of spatial data

A Generalized Spatial Autoregressive Neural Network Method for Three-Dimensional Spatial Interpolation

Importance of spatial predictor variable selection in machine learning applications -- Moving from data reproduction to spatial prediction

Underdetermined Joint Blind Source Separation of Multiple Datasets

The Spatial Kernel Predictor based on Huge Observation Sets

k-nearest neighbors prediction and classification for spatial data

Deep Spatial Prediction via Heterogeneous Multi-Source Self-Supervision

Modelling multivariate spatio-temporal data with identifiable variational autoencoders

Nonparametric spatial autoregressive model using deep neural networks

Semiparametric Estimation of Cross-covariance Functions for Multivariate Random Fields

Spatial Regression Graph Convolutional Neural Networks: A Deep Learning Paradigm for Spatial Multivariate Distributions.

Semi-Blind Source Separation with Learned Constraints

A multivariate Gaussian random field prior against spatial confounding