Refining Land-Cover Maps Based on Probabilistic Re-Classification in CCA Ordination Space

• Main Authors: Yue Wan, Jingxiong Zhang, Wenjing Yang, Yunwei Tang
• Format: Article
• Language: English
• Published: MDPI AG 2020-09-01
• Series: Remote Sensing
• Subjects: map refinement; land-cover; canonical correspondence analysis; class occurrence pattern indices; reference sample data; refinability
• Online Access: https://www.mdpi.com/2072-4292/12/18/2954

Due to spatial inhomogeneity of land-cover types and spectral confusions among them, land-cover maps suffer from misclassification errors. While much research has focused on improving image classification by re-processing source images with more advanced algorithms and/or using images of finer resolution, there is rarely any systematic work on re-processing existing maps to increase their accuracy. We propose refining existing maps to achieve accuracy gains by exploring and utilizing relationships between reference data, which are often already available or can be collected, and map data. For this, we make novel use of canonical correspondence analysis (CCA) to analyze reference-map class co-occurrences to facilitate probabilistic re-classification of map classes in CCA ordination space, a synthesized feature space constrained by map class occurrence patterns. Experiments using GlobeLand30 land-cover (2010) over Wuhan, China were carried out using reference sample data collected previously for accuracy assessment in the same area. Reference sample data were stratified by map classes and their spatial heterogeneity. To examine effects of model-training sample size on refinements, three subset samples (360, 720, and 1480 pixels) were selected from a pool of 3000 sample pixels (the full training sample). Logistic regression modeling was employed as a baseline method for comparisons. Performance evaluation was based on a test sample of 1020 pixels using a strict and relaxed definitions of agreement between reference classification and map classification, resulting in measures of types I and II, respectively. It was found that the CCA-based method is more accurate than logistic regression in general. With increasing sample sizes, refinements generally lead to greater accuracy gains. Heterogeneous sub-strata usually see greater accuracy gains than in homogeneous sub-strata. It was also revealed that accuracy gains in specific strata (map classes and sub-strata) are related to strata refinability. Regarding CCA-based refinements, a relatively small sample of 360 pixels achieved a 3% gain in both overall accuracy (OA) and F_0.01 score (II). By using a selective strategy in which only refinable strata of cultivated land and forest are included in refinement, accuracy gains are further increased, with 5–11% gains in users’ accuracies (UAs) (II) and 4–10% gains in F_0.01 scores (II). In conclusion, on condition of refinability, map refinement is well worth pursuing, as it increases accuracy of existing maps, extends utility of reference data, facilitates uncertainty-informed map representation, and enhances our understanding about relationships between reference data and map data and about their synthesis.