CAREER: Neural Fields for Efficient Representation and Exploratory Analysis of Heterogeneous Geospatial Signal Sets

Visualization and analysis of geospatial datasets play crucial roles in high-impact applications from agriculture to energy. However, these efforts are challenged by the large and heterogeneous nature of these datasets, which typically combine massive remote sensing signals (such as satellite images) with sparse sensors (such as ground temperature gauges), varying across sampling rates, data types, and noise profiles. Existing visualization software and standard artificial intelligence models struggle to efficiently process datasets with such challenging properties. This project will develop a novel artificial intelligence framework, with associated open-source software, designed to simplify and accelerate the representation, visualization, and analysis of large geospatial signal sets. Application domains of interest will include land surface temperature interpolation, soil moisture estimation, geophysical exploration, and aerial debris inference. Furthermore, the project will support education by creating interactive Python notebook demonstrations that allow high school, undergraduate, and graduate students to visually manipulate data science algorithms. These resources will reach thousands of learners through a free OpenStax university textbook, university courses, and a research experience program for Houston-area high school teachers. The project's technical aims are divided into three main thrusts to develop and evaluate a novel signal representation framework called the multivariate neural field (MVNF). The MVNF is a continuous function representation utilizing a shared neural field encoder and shallow per-signal multilayer perceptron (MLP) decoders to compactly represent multiple signals over a shared domain. First, the investigator will develop techniques enabling MVNFs to represent large, heterogeneous signal sets through mixed activation function layouts, multi-resolution hash encoders, and learnable parameter initializations to solve inverse problems. Second, the investigator will develop error prediction measures based on continuous piecewise affine spline and neural tangent kernel theories to rigorously analyze and control MVNF errors in applied settings. Third, the investigator will evaluate these methods across four geospatial applications and integrate the developments into an open-source visualization toolkit featuring real-time parallelized rendering. Finally, education activities will yield an open-source repository of interactive Python notebooks, enhancing instruction through OpenStax modules, university courses, and a Research Experience for Teachers program to co-design high school science and engineering curricula. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.