Land cover harmonization using Latent Dirichlet Allocation

Zhan Li,Joanne C White,Michael A Wulder,Txomin Hermosilla,Andrew M Davidson,Alexis J Comber

doi:10.1080/13658816.2020.1796131

Abstract

ABSTRACT Large-area land cover maps are produced to satisfy different information needs. Land cover maps having partial or complete spatial and/or temporal overlap, different legends, and varying accuracies for similar classes, are increasingly common. To address these concerns and combine two 30-m resolution land cover products, we implemented a harmonization procedure using a Latent Dirichlet Allocation (LDA) model. The LDA model used regionalized class co-occurrences from multiple maps to generate a harmonized class label for each pixel by statistically characterizing land attributes from the class co-occurrences. We evaluated multiple harmonization approaches: using the LDA model alone and in combination with more commonly used information sources for harmonization (i.e. error matrices and semantic affinity scores). The results were compared with the benchmark maps generated using simple legend crosswalks and showed that using LDA outputs with error matrices performed better and increased harmonized map overall accuracy by 6–19% for areas of disagreement between the source maps. Our results revealed the importance of error matrices to harmonization, since excluding error matrices reduced overall accuracy by 4–20%. The LDA-based harmonization approach demonstrated in this paper is quantitative, transparent, portable, and efficient at leveraging the strengths of multiple land cover maps over large areas.

Highlights

Mapping land cover is critical to estimating, understanding and modeling continuous and dynamic exchanges of energy and matter between the atmosphere and the land surface (Sellers et al 1997, Running 2008)
We found the source map (VLCE and ACI) land cover classes agreed with each other for more than 75% of the total map area (Figure 3, combinations in agreement are marked by black dots in the co-occurrence matrix)
We introduced the Latent Dirichlet Allocation (LDA) model to the harmonization of two national land cover products over southern Canada respectively focused on forest (VLCE) and agricul tural (ACI) environments

Summary

Introduction

Mapping land cover is critical to estimating, understanding and modeling continuous and dynamic exchanges of energy and matter between the atmosphere and the land surface (Sellers et al 1997, Running 2008). Land cover mapping over large areas has evolved from manual compilations of land-related source data of varying qualities (Matthews 1983) to automated approaches that rely on Earth observation data (Townshend 1992, Wulder et al 2018). Such advances have improved the efficiency, consistency, and transparency of land cover mapping over large areas, facilitating the production of land cover maps at national, continental, and global scales by various agencies (Franklin and Wulder 2002). The resulting maps may have discrepant land cover labels assigned to the same land areas, as reported by a range of studies comparing land cover maps (Comber et al 2004a, 2004b; Giri et al 2005, McCallum et al 2006, Fritz and See 2008, Herold et al 2008, Fritz et al 2011, Kaptué Tchuenté et al 2011, Pflugmacher et al 2011, Pérez-Hoyos et al 2012a, 2017)

Objectives

Methods

Results

Discussion

Conclusion