TL;DR
This study evaluates the consistency of the widely used MODIS land cover product (MCD12Q1) against the finer-resolution Chinese GlobeLand30 dataset. By harmonizing classification schemes and employing multiple evaluation metrics, the research provides a crucial methodological reference for understanding the reliability of global land cover data at regional scales.
What Problem Does This Paper Address?
Accurate land cover data is fundamental for understanding Earth's systems, from climate change to ecosystem services. However, existing global land cover (GLC) products often suffer from coarse spatial resolutions (300m to 1km), leading to a lack of spatial detail and potential inaccuracies. While the MODIS MCD12Q1 product offers a more frequent update cycle and finer resolution (500m) than many predecessors, its consistency with other, even finer-resolution datasets, particularly at regional scales, remains an important area of investigation. This paper addresses this by comparing the MCD12Q1 product (with its five different classification schemes) against the 30m GlobeLand30 dataset, aiming to quantify their agreement and identify potential discrepancies.
Methods
The core of this study involves a multi-step approach to harmonize and compare the two datasets:
- Harmonization of Land Cover Classification: A significant challenge is that MCD12Q1 and GlobeLand30 use different classification schemes and categorical scales. To overcome this, the researchers developed a harmonization method based on a Knowledge Rule (KR) approach and the C4.5 decision tree (DT) classification algorithm. This process reclassified the land cover types from both datasets into five common categories: "woodland," "grassland," "cropland," "wetland," and "artificial surfaces."
- Regionalization: The study area, Anhui Province, China, was divided into three sub-regions (Wanbei, Wanzhong, and Wannan) to account for varying geographical features and potential spatial heterogeneities.
- Evaluation Metrics: Four key indicators were used to assess consistency:
- Area Consistency: Measured using Pearson's correlation coefficient (R) and percentage disagreement (PD) to quantify the agreement in area proportions of land cover types.
- Spatial Consistency: Assessed using a pixel-by-pixel comparison, focusing specifically on the "woodland" class, to evaluate how well the spatial distribution aligns.
- Classification Accuracy: Evaluated using producer accuracy (PA), user accuracy (PU), overall accuracy (OA), and the Kappa coefficient (K) derived from confusion matrices.
- Landscape Diversity: Analyzed using the modified Simpson's diversity index (MSIDI) and modified Simpson's evenness index (MSIEI) to understand how landscape composition and structure differ between the datasets.
Datasets
- MODIS Land Cover Product (MCD12Q1): This dataset, with a spatial resolution of 500m, was used as the verified data. It contains five different land cover classification schemes:
-
IGBP global vegetation classification
- University of Maryland (UMD) vegetation classification
- LAI/FPAR scheme
- NPP scheme
- Plant Functional Type (PFT) scheme The data used was for the year 2010.
-
GlobeLand30: This dataset, with a finer spatial resolution of 30m, was used as the reference data. It comprises ten land cover types and was also for the year 2010.
- Anhui Provincial Administrative Regions and China's State Sector Datasets: Used for masking and defining the study sub-regions.
Key Insights and Ideas
- Harmonization is Crucial: The study underscores the necessity of harmonizing classification schemes before comparing different GLC products, especially when they originate from different sensors, algorithms, and classification systems.
- Regional Scale Matters: Consistency can vary significantly across different sub-regions within a province, highlighting the importance of regional-scale validation rather than relying solely on global averages.
- LAI/FPAR and NPP Schemes Show Weaknesses: The LAI/FPAR and NPP schemes within MCD12Q1 demonstrated lower consistency, particularly with regard to "woodland" and "artificial surfaces," often due to misclassifications with "others" or "barren land."
- Landscape Heterogeneity Impacts Consistency: Areas with higher landscape heterogeneity (more diverse land cover types and fragmented distributions) tend to exhibit weaker consistency between the datasets.
- GlobeLand30's Finer Resolution Advantage: The 30m GlobeLand30 dataset, due to its finer resolution, generally shows higher landscape diversity indices, suggesting a more detailed representation of landscape patches and their proportions.
Key Results
- IGBP and UMD Show Strong Consistency: The IGBP and UMD classification schemes within MCD12Q1 demonstrated the best overall consistency with GlobeLand30, particularly in terms of area agreement (R values > 97%) and percentage disagreement (PD values being smaller).
- LAI/FPAR's Weak Spatial Consistency: The LAI/FPAR scheme exhibited the weakest spatial consistency for "woodland" (O = 58.17%), largely due to misclassification with "others" and "barren land."
- NPP's Variable Accuracy: The overall accuracy (OA) of the NPP scheme varied significantly across the sub-regions, indicating its inconsistency.
- "Artificial Surfaces" a Challenge: The accuracy of "artificial surfaces" was notably low in the LAI/FPAR scheme, contributing to its overall weak performance.
- Diversity Indices Reflect Heterogeneity: Landscape diversity indices (MSIDI and MSIEI) were consistently higher for GlobeLand30 compared to MCD12Q1, and the diversity indices generally increased from Wanbei to Wanzhong to Wannan, reflecting increasing landscape heterogeneity.
Limitations and Future Work
- Focus on "Woodland" for Spatial Consistency: The study focused on the spatial consistency of "woodland" only. Future work could extend this analysis to other land cover types.
- Harmonization of "Barren Land": The "barren land" class from GlobeLand30 was largely ignored or treated as "others" due to its small proportion in some sub-regions. Further investigation into its specific misclassifications could be beneficial.
- Methodological Reference: The paper explicitly states that its methodology can serve as a reference for evaluating other GLC products across various spatial scales, suggesting future applications of this approach.
- Understanding Drivers of Misclassification: While misclassifications are identified, a deeper dive into the specific drivers (e.g., spectral confusion, temporal differences, sensor limitations) for each scheme could further refine our understanding.
Summary updated