Multi-Point Proximity Encoding For Vector-Mode Geospatial Machine Learning

TL;DR

This paper introduces MultiPoint Proximity encoding, a novel method for representing vector geospatial data as scaled distances to reference points, improving shape differentiation and spatial relationship capture for machine learning.

Contribution

The paper presents MPP encoding, a new shape encoding technique that enhances geometric feature representation and outperforms rasterization-based methods in geospatial ML tasks.

Findings

MPP encoding is shape-centric and continuous.

It effectively differentiates spatial objects based on geometry.

It captures pairwise spatial relationships with high accuracy.

Abstract

Vector-mode geospatial data -- points, lines, and polygons -- must be encoded into an appropriate form in order to be used with traditional machine learning and artificial intelligence models. Encoding methods attempt to represent a given shape as a vector that captures its essential geometric properties. This paper presents an encoding method based on scaled distances from a shape to a set of reference points within a region of interest. The method, MultiPoint Proximity (MPP) encoding, can be applied to any type of shape, enabling the parameterization of machine learning models with encoded representations of vector-mode geospatial features. We show that MPP encoding possesses the desirable properties of shape-centricity and continuity, can be used to differentiate spatial objects based on their geometric features, and can capture pairwise spatial relationships with high precision. In all cases, MPP encoding is shown to perform better than an alternative method based on rasterization.