Wang et al. (2025) Application of High-Precision Classification Method Based on Spatiotemporal Stable Samples and Land Use Policy in Oasis–Desert Mosaic Landscape Areas

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-11-28
• Authors: Jinghan Wang, Yuefei Zhou, Miaohang Zhou, Zengjing Song, Xiangyu Ji, Xujun Han
• DOI: 10.3390/rs17233859

Research Groups

• Chongqing Engineering Research Center for Remote Sensing Big Data Application, Chongqing Jinfo Mountain National Field Scientific Observation and Research Station for Karst Ecosystem, School of Geographical Sciences, Southwest University, Chongqing 400715, China
• Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, 7500 AE Enschede, The Netherlands

Short Summary

This study developed a high-precision land cover classification method for oasis–desert mosaic landscapes by integrating spatiotemporally stable samples, a novel Canopy Growth Index, and land-use policy constraints. The method achieved an overall accuracy of 91.9% and significantly reduced spatiotemporal inconsistencies in land cover products.

Objective

• To develop a high-precision land cover classification framework that addresses spatiotemporal inconsistencies and frequent, repetitive classification fluctuations in existing products, particularly in complex oasis–desert mosaic landscapes, by leveraging spatiotemporally stable samples and land-use policy constraints.

Study Configuration

• Spatial Scale: Ganzhou District, Zhangye City, Gansu Province, China, covering approximately 3661 km².
• Temporal Scale: Landsat time-series data from 2010 to 2020 (aggregated to an 11-year composite for analysis); CLCD data from 2010 to 2022; ESA WorldCover 2020–2021; Global PALSAR-2/PALSAR Mosaics and FNF maps 2010, 2015–2022.

Methodology and Data

• Models used: Improved Hampel filter (for outlier detection and correction), Fourier series fitting (for temporal trend modeling), Extreme Gradient Boosting (XGBoost) classifier (primary), Random Forest (RF), Logistic Regression (LR), k-Nearest Neighbors (KNN).
• Data sources: Landsat 5/7/8/9 satellite imagery (30 m spatial resolution) processed on Google Earth Engine (GEE); China Land Cover Dataset (CLCD); ESA WorldCover; Global PALSAR-2/PALSAR Mosaics and Forest/Non-Forest (FNF) maps; Remote sensing indices (Normalized Difference Vegetation Index (NDVI), Kernel NDVI (kNDVI), Soil-Adjusted Vegetation Index (SAVI), Modified Normalized Difference Water Index (MNDWI), Normalized Difference Built-up Index (NDBI), Bare Soil Index (BSI), Normalized Difference Fraction Index (NDFI)); Auxiliary data (Elevation, Slope, Aspect).

Main Results

• The proposed method achieved an overall classification accuracy of 91.9% (±0.037) and a Kappa coefficient of 0.8699, significantly outperforming other machine learning algorithms (RF, LR, KNN) and existing land cover products (CLCD, GlobeLand, GLC_FCS, ESA).
• The integration of land-use policy constraints markedly improved spatiotemporal consistency, reducing pixels with unreasonable repetitive land-cover changes from 56.67% to 34.03% (a reduction of 22.64 percentage points).
• A novel Canopy Growth Index (CGI) was introduced, enhancing the separability among various vegetation types, particularly deciduous forests, grasslands, and shrublands.
• The two-step filtering framework (Improved Hampel filter and Fourier series fitting) effectively identified and retained 881 spatiotemporally stable samples from an initial 1408, ensuring high sample purity and temporal robustness.
• The method demonstrated robust accuracies for grasslands, shrublands, and croplands, while classification of barren land and impervious surfaces remained challenging due to spectral similarities in the oasis–desert environment.

Contributions

• Introduced a novel Canopy Growth Index (CGI) to enhance the separability of vegetation types in complex oasis–desert mosaic landscapes.
• Developed a two-step filtering framework (Improved Hampel filter and Fourier series fitting) for automated extraction of spatiotemporally stable samples, which significantly reduced manual effort and enhanced sample purity and temporal robustness.
• Integrated national and regional land-use policies as post-classification logical refinements to detect and correct temporal–spatial inconsistencies, effectively mitigating spurious land-cover changes and improving overall spatiotemporal consistency.
• Provided a simple, efficient, and transferable workflow for high-precision land cover mapping in ecologically fragile arid and semi-arid regions, addressing the limitations of large-scale products at local scales.
• Demonstrated superior accuracy and temporal stability compared to existing land cover datasets and conventional machine learning methods, offering a reliable basis for ecological restoration and sustainable land-use planning.

Funding

• Special fund for youth team of Southwest University project (grant numbers: SWU-XJLJ202305).

Citation

@article{Wang2025Application,
  author = {Wang, Jinghan and Zhou, Yuefei and Zhou, Miaohang and Song, Zengjing and Ji, Xiangyu and Han, Xujun},
  title = {Application of High-Precision Classification Method Based on Spatiotemporal Stable Samples and Land Use Policy in Oasis–Desert Mosaic Landscape Areas},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17233859},
  url = {https://doi.org/10.3390/rs17233859}
}