Intelligent Geospatial Science and Technology Multi-Dimensional Cognition of Temporal-Spatial Scenes
CHEN Fei, ZHANG Zhiwei
Land cover is a comprehensive term that encompasses various coverings and their characteristics on the Earth's surface, reflecting the distribution and evolution of material types. It primarily includes surface vegetation, glaciers, lakes, marsh wetlands, and various buildings, serving as essential foundational data in fields such as food security, greenhouse gas emissions, ecosystem structure and function, water resource circulation and redistribution, and biogeochemical cycles. However, due to the influence of production technology, land cover data may encounter issues of misclassification and omission, which directly impact the effectiveness of the data in subsequent applications. To ensure the reliability of future applications, the accuracy assessment of land cover data is an indispensable component of the production process and has become a key research focus in the production and application of remote sensing data both domestically and internationally.
Current research on accuracy assessment methods for land cover data encompasses all aspects of the evaluation process, including sampling methods, sample judgment and checking, and a variety of evaluation indicators. This paper aims to clarify the necessity, main technical components, and practical cases of land cover data accuracy assessment methods in recent years, providing a systematic reference for land cover accuracy assessment.
The main method for accuracy assessment of land cover data is the confusion matrix, from which several indicators, such as overall accuracy, Kappa coefficient, producer's accuracy, and user's accuracy, are derived. To address the challenge of unbalanced sampled probabilities in stratified sampling accuracy assessment, a weighted confusion matrix has been developed, allowing for the calculation of weighted overall accuracy, weighted producer's accuracy, and weighted user's accuracy. Additionally, metrics such as quantitative disagreement and allocation disagreement between classified maps and reference data samples, as well as the F1 score, have also been introduced.
Sampling for verification involves the selection of representative samples within the verification area based on statistical sampling principles. This serves as the first step in land cover validation and significantly influences its accuracy and objectivity. Key principles guiding this process include the probability principle, feasibility principle, emphasis on rare classes, and considerations of spatial heterogeneity and spatial correlation. The main tasks include calculating sample sizes and designing spatial layouts.
The collection of reference data for sample checking is the process of determining the true ground type at sample locations based on reference data to evaluate the correctness of classifications. This stage is also the most time-consuming and costly part of the accuracy evaluation process. Reference data selection should adhere to principles such as temporal proximity, seasonal consistency, high spatial resolution, high geometric accuracy, compatibility of classification system, and authoritative sources. Methods such as multi-scale sample checking, multi-semantic sample checking, and credibility grading can be employed.
Finally, some cases of the accuracy assessment processes of key land cover datasets at a 30 m resolution, such as GlobeLand30, GLC_FCS30, and FROM-GLC, are analyzed. Current accuracy assessment methods exhibit two notable characteristics. First, the design of sampling methods has diversified. Many validation cases for medium-resolution land cover datasets take into account spatial heterogeneity and spatial correlation, involving related index calculations, quantitative analyses, and geographic stratification designs, leading to increasingly varied sampling approaches. Second, the assessment indicators have expanded beyond traditional measures such as overall accuracy, user's accuracy, producer's accuracy, and the Kappa coefficient, now incorporating a wider range of metrics.
In conclusion, the core mission of land cover accuracy assessment focuses on achieving a deep integration of standardization, ensuring that land cover data products reliably support major decision-making in areas such as ecological security, territorial governance, and climate change response.
