Answers are grounded in this paper and cite their sources.
Automated and high-precision land cover classification (LCC) using synthetic aperture radar (SAR) is crucial for environmental monitoring, while its advancement has been constrained by the time-consuming and labor-intensive acquisition of sufficient and high-quality training samples. Most existing automated SAR LCC sample extraction methods suffer from reliance on prior information of LCC products, which introduces classification and co-registration errors and fails to adequately leverage the intrinsic scattering characteristics of SAR imagery. To address this challenge, a multiscale superpixel-based automatic SAR LCC sample extraction and purification (MS-ASLPEP) method was proposed that integrates multiscale spatial structures with scattering characteristic constraints. First, the multitemporal European Space Agency WorldCover land cover products were selected to identify the spatiotemporally stable regions and then multiscale superpixel segmentation and hierarchical consistency voting mechanism were combined to perform coarse sample extraction. Subsequently, in order to refine the coarse samples, a three-stage progressive statistical purification method was designed, which was achieved in the SAR feature space through local homogeneity testing based on Euclidean distance and global outlier removal using Mahalanobis distance. The accuracy and applicability of the proposed automatic sample extraction method were analyzed in two study areas with distinct topographies, namely Kaifeng and Beijing in China, and the results indicate that: purified training samples exhibited enhanced intraclass compactness and interclass separability, as indicated by t-SNE analysis results andthe overall accuracy of the six classic and deep learning algorithms in both study areas exceeded 83.65%, with the highest reaching 89.05%, using MS-ASLPEP. This article indicates that the combination of geostatistical methods with SAR multiscale superpixel feature analysis can effectively utilize historical LCC products to generate high-confidence SAR training samples, offering a novel approach for improving the automation level and accuracy of large-scale automatic SAR LCC.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: