Mapping land-cover dynamics in arid regions using spectral mixture analysis and representative training samples

Timely and accurate land-cover surveys are essential for understanding the drivers of global climate change and maintaining the carbon cycle. However, monitoring land-cover dynamics in arid regions presents challenges due to landscapes complexity and limited training samples. To address these challenges, this study proposes a novel approach that combines Spectral Mixture Analysis (SMA) and adaptive strategies to automatically generate high-quality training samples for mapping large-scale land-cover dynamics in Guanzhong Plain (GZ), China. SMA was employed to develop new indices for detecting spatiotemporal changes and boundaries in land cover: the Endmember Abundance Vegetation Index (EAVI) and the Heterogeneity Detection Index (
HDI_RMSE
). Subsequently, the spatial consistency region of multi-source products were automatically generated using spatial superposition analysis, the morphological corrosion method, and
HDI_RMSE
. Once the initial training sample pool was created, it was optimized and updated through an adaptive sampling strategy to ensure a representative dataset. Additionally, by integrating the EAVI with CCD algorithm (CCD_EAVI), spatiotemporally stable regions of land cover change were identified, enabling the automatic updating of annual training samples. Finally, a spatiotemporal consistency optimization approach was employed to improve the temporal stability of the time-series product. Based on independent validation samples, the overall accuracy of the land-cover mapping of the GZ in 2020 reached 97.92%, with an average mapping accuracy exceeding 96% over the past 32 years. Compared to benchmark products, these mapping results significantly improved spatial clarity, noise reduction, and classification accuracy. These findings highlight the proposed method’s effectiveness for automated sampling and time-series land-cover mapping.