Global perspectives on sand dune patterns: Scale-adaptable classification using Landsat imagery and deep learning strategies

Sand dune patterns (SDPs) are spatial aggregations of dunes and interdunes, exhibiting distinct morphologies and spatial structures. Recognizing global SDPs is crucial for understanding the development processes, contributing factors, and self-organization characteristics of aeolian systems. However, the diversity, complexity, and multiscale nature of global SDPs poses significant technical challenges in the classification scheme, sample collection, feature representation, and classification method. This study addresses these challenges by developing a novel global SDP classification approach based on an advanced deep-learning network. Firstly, we established a globally applicable SDP classification scheme that accommodates the diversity nature of SDPs. Secondly, we developed an SDP semantic segmentation sample dataset, which encompassed a wide array of SDP representations. Thirdly, we deployed the SegFormer network to automatically capture detailed dune structures and developed a weighted voting strategy to ensure scale adaptability. Experiments utilizing Landsat-8 imagery yielded a commendable overall accuracy (OA) of 85.43 %. Notably, most SDP types exhibited high classification accuracies, such as star dunes (97.43 %) and simple linear dunes (87.17 %). The weighted voting strategy prioritized the predictions of each type, resulting in a 1.41 %∼7.91 % improvement in OA compared to the single-scale classification and average voting methods. This innovative approach facilitated the generation of a high-quality, fine-grained, and global-scale SDP map at 30 m resolution (GSDP30), which not only directly provides the spatial distribution of global SDPs but also serves as valuable support for understanding aeolian processes. This study represents the first instance of producing such a comprehensive and globally applicable SDP map at this fine resolution.