The Maowusu Desert is located in the core area of the "Yellow River Key Ecological Zone" in the national "Three Zones and Four Belts" ecological security strategic pattern, and belongs to the key ecological function zone of windbreak and sand fixation in the western section of the northern agricultural pastoral ecotone. Due to the dual impact of long-term agricultural and animal husbandry development and ecological restoration projects, the surface landscape in this area is highly fragmented, leading to frequent misjudgment or poor applicability of existing large-scale open source land classification products. For this purpose, this dataset is specifically tailored to the complex surface features of sandy areas, aiming to compensate for the limited availability of national scale products in local vulnerable areas and provide a more accurate base map for ecological assessment.
This data product is based on a total of 10 Landsat series remote sensing images (5-year time step) from 1980 to 2025, with a spatial resolution of 30 meters. The dataset adopts a first level land use/cover classification system, covering 7 major categories: tree forest, shrub forest, farmland, water body, grassland, bare sand land, and construction land. In terms of production process, a highly reliable ground sample set was first constructed through field investigations, low altitude drone orthophoto, and multi-source open source data fusion; Then, relying on the ENVI platform, multiple supervised classification algorithms are compared and screened, and a subjective and objective collaborative mechanism of confusion matrix quantitative evaluation and expert visual interpretation is introduced; Finally, through manual interactive inspection and land boundary correction, a full sequence spatial distribution product is generated. After verification, the average Kappa coefficient of the entire sequence dataset is 0.71, with an average overall accuracy of 80.36%, which can fully meet the research needs of long-term monitoring of regional ecological succession and macro land change science.
This dataset is in raster format (GeoTIFF), with each pixel value stored in integer encoding, uniquely corresponding to a specific land use/cover type. The classification system consists of 7 major categories, and the specific grid code mapping relationships are as follows: 1: Grassland 2: Shrubs forest 3: Farmland 4: Tree forest 5: Naked Sand 6: Water body 7: Construction land
The main remote sensing data of this dataset comes from Landsat series satellite images (covering MSS, TM, ETM+, and OLI sensors). The time series strictly covers the six-month generation nodes from 1980 to 2025, with a unified spatial resolution of 30 meters (early MSS images were resampled to 30 meters). To eliminate phenological interference, high-quality scenes with vegetation growth season (June October) and panoramic cloud cover below 10% were selected for the images; For a very small number of years with poor quality due to cloudy weather, adjacent seasonal images are used as substitutes. The open-source data for auxiliary cross validation includes the China Annual Land Cover Dataset (CLCD), https://doi.org/10.5194/essd-13-3907-2021 ）Esri Global 10 meter Land Cover Data (Esri 10m, https://doi.org/10.1109/IGARSS47720.2021.9553499 ）The spatial distribution map of 250 meter irrigated farmland in China (CIrrMap250, https://doi.org/10.5194/essd-16-5207-2024 ）And the Global Surface Water Dataset (GSWE), https://doi.org/10.1038/nature20584 ）.
1. Construction of classification sample set
The interpretation samples were obtained from field investigations, low altitude unmanned aerial vehicle orthophoto images, and the intersection confidence zone of multiple high-precision open source products (CLCD, Esri 10m, CIrrMap250, GSWE, etc.). A total of 1808 high-quality training and validation samples were screened and extracted. The Jeffreys Matuxita distance (J-M Distance) of each region of interest (ROI) is greater than 1.7, ensuring high separability of the spectral features of the samples.
2. Multi algorithm classification and result optimization
Based on ENVI 5.6 software, comprehensive remote sensing images and sample data are used to perform supervised classification using six algorithms: support vector machine, minimum distance, maximum likelihood, Mahalanobis distance, parallelepiped, and neural network. In the preliminary evaluation, the parallelepiped and neural network algorithms that have poor classification performance for complex sandy terrain were excluded. Subsequently, a subjective objective collaborative evaluation method combining confusion matrix and visual interpretation was used to cross compare the results of the remaining four algorithms, in order to dynamically select the best classification algorithm for different years' image features. Finally, with the assistance of fine manual interactive visual correction to eliminate plaque noise, this dataset was generated.
This dataset is based on confusion matrix for Kappa coefficient and overall accuracy testing, and the results are as follows:
In 1980 (Kappa 0.74, overall accuracy 84.28%);
In 1985 (Kappa 0.62, overall accuracy 76.26%);
In 1990 (Kappa 0.74, overall accuracy 85.12%);
In 1995 (Kappa 0.74, overall accuracy 85.50%);
In 2000 (Kappa 0.73, overall accuracy 81.39%);
In 2005 (Kappa 0.77, overall accuracy 83.39%);
In 2010 (Kappa 0.69, overall accuracy 76.43%);
In 2015 (Kappa 0.71, overall accuracy 77.63%);
In 2020 (Kappa 0.73, overall accuracy 81.32%);
In 2025 (Kappa 0.64, overall accuracy 72.30%);
Average (Kappa 0.71, overall accuracy 80.36%)
| # | number | name | type |
| 1 | 2024JBGS0020 | Systematic Ecological Governance Integrated Technology Demonstration of Mu Us Sandy Land | Major Demonstration Project of Technological Innovation in Inner Mongolia Autonomous Region with the 'Reveal the List and Take the Lead' Approach |
This work is licensed under
CC BY 4.0 (Creative Commons Attribution 4.0 International License).
CSTR:11738.11.ncdc.nieer.db7325.2026
10.12072/ncdc.nieer.db7325.2026
