Assessing the accuracy of remote sensing data products: A multi-granular spatial sampling method

Abstract

Over the past two decades, remote sensing data have demonstrated significant potential across various fields. Accuracy assessment, as a crucial component, plays a vital role in ensuring the effective application of remote sensing data products. To assess the accuracy of remote sensing data products, this paper proposed a multi-granular spatial sampling method (MG-SSM). In the proposed MG-SSM, the assessed remote sensing data products were stratified into different layers based on spatial heterogeneity, quantified by calculating the aggregation index. Subsequently, multi-granular sampling units were defined for each layer based on spatial correlation, quantified using Moran’s I and Z-score. Finally, the accuracy of remote sensing data products was assessed by comparing it with the reference data. Using the classification results of Landsat 9 OLI/TIRS with a 30-m spatial resolution as assessed data, and the classification results of Sentinel-2 MSI with a 10-m spatial resolution as reference data, and employing Overall Accuracy (OA) and the Kappa coefficient as assessment metrics, the performance of MG-SSM was compared against other sampling methods, including SRS, ST, SY, SSS, and GLCM. The performances of SRS, ST, SY, SSS, GLCM, and MG-SSM were 74.41%, 75.39%, 76.57%, 78.47%, 76.77%, and 67.36% in terms of overall accuracy, and 0.6, 0.61, 0.64, 0.66, 0.64, and 0.46 in terms of the Kappa coefficient, respectively. The results indicated that MG-SSM exhibited the lowest values for overall accuracy and the Kappa coefficient, aligning with the objective of assessing the accuracy of remote sensing data products. This outcome is attributed to MG-SSM’s strategy of distributing more sample points in areas with lower aggregation of land-cover classes, which are more prone to classification errors. Collectively, the proposed MG-SSM enhanced the representativeness and reduced the information redundancy of the sample points, making it an efficient sampling method for assessing the accuracy of remote sensing data products.