A 2 m Resolution Impervious Surface Distribution Product of the Qinghai–Xizang Plateau Based on SegformerAT Model

High-resolution impervious surface data is essential for studies related to urban development, human activities, carbon emissions, and ecological environments. Owing to the unique geographic and climatic conditions, impervious surfaces in the Qinghai-Xizang (Tibet) Plateau (QXP) are often surrounded by sparsely vegetated areas with low canopy cover and high land exposure. Additionally, the QXP is characterized by numerous, scattered small-scale impervious surfaces (SSIS). Existing impervious surface products, primarily at global or national scales, typically provide medium spatial resolutions (30-m or 10-m) and exhibit significant omission and commission errors for impervious surfaces in QXP, particularly in mapping SSIS. Consequently, a 2 m impervious surface distribution product for QXP in 2020 (QXP-ISD2m product) was created using Gaofen imagery. This product categorizes roads as a separate class to enhance the data's applicability and defines non-road impervious surfaces as Primary Impervious Surfaces (PIS). Specifically, a framework for large-scale impervious surface mapping was proposed. Initially, the QXP 2 m Impervious Surface Unique Dataset, consisting of 2235 samples, was constructed using a target-scale-based stratified sampling strategy and augmented to 4807 samples by incorporating spatially overlapping multi-temporal and multi-sensor images (2 m resolution, 512× 512 pixels). This dataset accommodates various scales of impervious surface types in the QXP and ensures accurate impervious surface extraction. To address significant scale discrepancies between PIS and road targets, an improved Segformer model, named SegformerAT, was introduced. This model employs patch-wise normalization and utilizes the augmented dataset for training, thus ensuring high accuracy in PIS extraction and significantly improving road detection accuracy. Finally, the Posterior-Knowledge-Based Merging Rule (PKBMR) was employed to achieve effective large-scale mapping. Validation of the QXP-ISD2m product with a set of 7000 independent samples yielded F1 scores of 0.894 for PIS and 0.895 for roads. Comparative experiments with other algorithms further confirmed the superiority of the SegformerAT model.