Towards Seamless Validation of Land Cover Data

Ekaterina Chuprikova, Lukas Liebel, Liqiu Meng

Abstract

This article demonstrates the ability of Bayesian Network analysis for the recognition of uncertainty patterns associated with the fusion of various land cover datasets, including GlobeLand30, CORINE (CLC2006, Germany), and land cover data derived from Volunteered Geographic Information (VGI) such as OpenStreetMap (OSM). The results are expressed as probability and uncertainty maps, which can be regarded as a by-product of the GlobeLand30 data. This uncertainty information may guide quality improvements in GlobeLand30 by incorporating ground truth data, superior quality information, expert know-how, and crowd intelligence. Such an endeavor aims to pave the way towards seamless validation of global land cover data and targeted knowledge discovery in areas with higher uncertainty values.

Introduction

Global Land Cover (GLC) maps are indispensable tools for environmental monitoring and climate modeling. However, their accuracy is often inconsistent, with reported values ranging from 46% to 83% depending on the region and terrain complexity. Validating these maps on a global scale is a massive undertaking, primarily because high-quality "ground truth" data is expensive and difficult to obtain.

This research proposes a framework for "seamless validation." Instead of treating validation as a binary pass/fail assessment, the authors treat spatial uncertainty as an indicator of conflict between different data sources. The goal is to create a system that automatically highlights geographic "hotspots" of disagreement, signaling exactly where human experts or citizen scientists should focus their attention to improve map quality.

Methodology

The core of this research is the application of Bayesian Networks (BN) for data fusion. BNs are probabilistic graphical models that represent variables and their conditional dependencies via a Directed Acyclic Graph (DAG). The study focused on the Upper Bavaria region of Germany, chosen for its diverse land classes and data availability.

Data Harmonization

The authors unified three distinct datasets—GlobeLand30, CORINE, and OSM—into a consistent 30m resolution grid using a 10-class classification schema. A custom pipeline was developed to transform OpenStreetMap (OSM) vector data into a rasterized land cover map, mapping complex OSM tags (e.g., "meadow" or "pond") to standard GLC30 classes.

Workflow for converting raw OSM data into a rasterized land cover map.

Probabilistic Modeling

Using the bnspatial package in R, the team modeled the dependencies between the datasets. The BN calculates the posterior probability—the likelihood of a specific land cover class existing at a pixel given the evidence from all input maps.

Results and Discussion

The primary output of the study is a series of high-resolution probability and uncertainty maps. Uncertainty is quantified using the Shannon Index (entropy). High entropy indicates that the input datasets are in high conflict, while low entropy suggests a consensus.

Uncertainty map of Upper Bavaria based on the Shannon Index.

Key Insights

  • Uncertainty as a Functional By-product: Uncertainty is treated as a primary output that guides data refinement rather than just a statistical footnote.
  • VGI as a Validation Engine: While OSM is not a perfect "ground truth," its high resolution and crowd-sourced nature make it an excellent auxiliary source for cross-checking official satellite products.
  • Non-linear Reasoning: Bayesian Networks allow for the incorporation of expert "priors" (pre-existing knowledge about class likelihoods), improving fusion results beyond simple map algebra.

Conclusion

The study successfully produced high-resolution uncertainty maps for Upper Bavaria, effectively capturing discrepancies across multiple datasets. These maps act as a visual guide for targeted validation, allowing researchers to focus manual efforts on areas where data sources disagree most. Future work will focus on developing interactive applications to integrate these uncertainty maps into geoportals, facilitating real-time, crowd-sourced validation of global land cover data.

Summary updated

Abstract

This article demonstrates the ability of the Bayesian Network analysis for the recognition of uncertainty patterns associated with the fusion of various land cover data sets including GlobeLand30, CORINE (CLC2006, Germany) and land cover data derived from Volunteered Geographic Information (VGI) such as Open Street Map (OSM). The results of recognition are expressed as probability and uncertainty maps which can be regarded as a by-product of the GlobeLand30 data. The uncertainty information may guide the quality improvement of GlobeLand30 by involving the ground truth data, information with superior quality, the know-how of experts and the crowd intelligence. Such an endeavor aims to pave a way towards a seamless validation of global land cover data on the one hand and a targeted knowledge discovery in areas with higher uncertainty values on the other hand.
BibTeX
@inproceedings{chuprikova2018towards,
  author = {Ekaterina Chuprikova and Lukas Liebel and Liqiu Meng},
  title = {{Towards Seamless Validation of Land Cover Data}},
  booktitle = {Proceedings of the ICA},
  year = {2018},
  doi = {10.5194/ica-proc-1-23-2018},
}
Figure