Deep Spatial Prediction via Heterogeneous Multi-Source Self-Supervision

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