GF6RTSChange2021–2025: A GF-6 Long-Term Dataset for Retrogressive Thaw Slump Deep Learning Change Detection

The full interval from 03 February 2021 to 22 October 2025 is 1,722 d. The successive gaps are 660, 365, 332, and 365 d. Seasonal consistency reduces interannual differences caused by illumination, phenology, surface wetness, and atmosphere.

The date embedded in each product name follows the YYYY_MMDD convention. The five source images were processed in ENVI through radiometric correction, atmospheric correction, orthorectification, and pansharpening. Radiometric correction placed sensor measurements on a comparable radiometric scale; atmospheric correction reduced residual path effects; orthorectification addressed terrain and acquisition geometry; and pansharpening combined multispectral information with fine spatial detail to provide the 2 m, four-band analysis imagery.

Following geometric alignment, all five images underwent a common relative radiometric correction and histogram matching. The purpose was to map the dates into a shared radiometric domain and suppress non-surface variation arising from acquisition conditions, residual atmosphere, and sensor-response differences. The operation leaves the common grid unchanged and modifies only the radiometric comparability of pixel values.

Semantic-segmentation outputs served as initial candidates, after which individual RTS objects were visually checked, revised, and manually delineated to create inventory-style binary masks. Both masks were standardised to uint8 values 0 and 1: class 1 denotes RTS and class 0 denotes background, former NoData, or other invalid values. No mixed NoData category remains in the binary training masks.