An entropogram-based Random Field model for categorical geospatial data prediction

Abstract

Categorical geospatial data underpin applications from biodiversity monitoring to land-use planning, yet existing approaches often fail to recover rare classes while preserving realistic patch structures. We introduced an Entropogram-based Random Field (ERF) model that integrates intrinsic randomness from local class probabilities with entropogram-derived spatial dependence, balancing local class proportions with global neighborhood associations. Using a 10-class, 1-km land-cover map of Northern Ireland, we compared ERF against Indicator Kriging (IK), multi-phase Indicator Kriging (MIK), Compositional Data Analysis (CoDA) and a spatial multinomial logistic (SMLM) model. ERF matches IK and MIK in overall accuracy but achieves higher recall and F1 scores for minority classes, reducing the loss of small, coherent patches. While CoDA ensures compositional validity, it underperforms on rare classes and increases spatial aggregation; MIK improves rare-class recovery but still favors dominant types. SMLM performs comparably to ERF but with far higher computational demand. Landscape metrics showed that ERF and SMLM best preserved patch diversity and realistic geometry, whereas IK and CoDA produced more aggregated patterns. Together, these results highlight ERF as a computationally efficient, scalable and balanced solution for categorical mapping, particularly in applications where minority-class recovery and spatial realism are critical for biodiversity monitoring, habitat connectivity and land-use planning.

More information

Identifiers

ORCID for Wen-Bin Zhang: orcid.org/0000-0002-9295-1019

ORCID for Shengjie Lai: orcid.org/0000-0001-9781-8148

ORCID for Peter M. Atkinson: orcid.org/0000-0002-5489-6880

Catalogue record

Export record

Altmetrics