Global earth surface coverage mapping method based on domestic satellite image with 2-meter resolution

Abstract

The invention relates to the technical field of electronic maps, in particular to a global earth surface coverage mapping method based on a 2-meter-resolution domestic satellite image, and aims at solving the problem that an existing 30-meter and 10-meter-resolution global earth surface coverage product is low in accuracy. And a mapping strategy combining general and regional feature modeling with layered land surface coverage is adopted. According to the image prepared by the method, the land cover accuracy is high.

Description

Technical Field

[0001] This application relates to the field of electronic map technology, specifically a global land cover mapping method based on 2-meter resolution domestic satellite imagery. Background Technology

[0002] Land cover information is an important source of information characterizing human activities and global change, one of the most important variables in the Global Earth Observation System (GEOS), and a crucial foundation for studying global climate change. It is of great significance to climate change research, biodiversity conservation, global food security, and public health (Foley et al., 2005; Liu et al., 2021; Liang et al., 2010).

[0003] With the rapid advancement of satellite remote sensing technology and computing power, especially the free access to medium-resolution imagery such as Landsat and Sentinel, significant progress has been made in global land cover products. Global land cover products with 30-meter and 10-meter resolutions are continuously being released. 30-meter land cover products include GlobeLand30 (Chen et al., 2015), FROM-GLC30 (Gong et al., 2012), and GLC_FCS30 (Zhang et al., 2021), while 10-meter land cover products include FROM-GLC10 (Gong et al., 2019), ESA WorldCover (Buchhorn et al., 2020), ESRI Land Cover (Karra et al., 2021), and GLC_FCS10 (Zhang et al., 2025). Zhao et al. (2023) pointed out that the accuracy of FROM_GLC10, ESRI LandCover, and ESA WorldCover is relatively low. Summary of the Invention

[0004] The purpose of this invention is to address the low accuracy of existing global land cover products with resolutions of 30 meters and 10 meters by providing a global land cover mapping method based on domestically produced satellite imagery with a resolution of 2 meters.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] A global land cover mapping method based on 2-meter resolution domestic satellite imagery includes the following steps:

[0007] Step 1: Use domestically produced satellites with a resolution of 2 meters to acquire global remote sensing image data, and obtain various types of land cover based on the global remote sensing image data;

[0008] Step 2: Based on prior knowledge, obtain global glacier surface cover and global wetland surface cover;

[0009] Step 3: Overlay various types of land cover, global glacier land cover, and global wetland land cover to obtain a global land cover image.

[0010] Furthermore, the specific steps of step 1 are as follows:

[0011] Step 11: Acquire global remote sensing image data and input the global remote sensing image data into the terrestrial ecoregion model to obtain the world terrestrial ecoregion division results;

[0012] Step 12: For each terrestrial ecological zone, use the land type extraction model to obtain the land type 1 extraction result. The land type 1 extraction result includes one or more of the following: cultivated land, forest land, shrubland, grassland, artificial surface, desert and bare surface, water area, glacier and perennial snow cover, wetland and tundra. For artificial surface, the artificial surface needs to be input into the regional culture model for further subdivision. After subdivision, the final land type 1 extraction result is obtained. The regional culture model is characterized by building style, urban layout and urban greening, and divides the input image into cultural zones.

[0013] Step 13: Input the final land cover type 1 extraction result from Step 12 into the land cover extraction filter to obtain various types of land cover.

[0014] Furthermore, the world's terrestrial ecoregions include tropical and subtropical humid broad-leaved forests, tropical and subtropical dry broad-leaved forests, tropical and subtropical coniferous forests, temperate broad-leaved mixed forests, temperate coniferous forests, northern forests / taiga forests, tropical and subtropical grasslands, savannas and shrublands, temperate grasslands, savannas and shrublands, floodplains and savannas, mountain grasslands and shrublands, tundra, Mediterranean forests, woodlands and shrublands, deserts and xerophytic shrublands, and mangroves.

[0015] Furthermore, the global glacier surface cover is obtained through the World Glacier Inventory.

[0016] Furthermore, the global wetland surface cover was obtained through a global lakes and wetlands database.

[0017] Furthermore, the land use extraction model is a pre-trained land use extraction model, which is obtained by training using a sample dataset, which is obtained through the following steps:

[0018] Step 121: Acquire historical remote sensing image data and corresponding labeled land cover data. If the interpreted data is vector data, convert the vector data into raster data with the same resolution as the image data.

[0019] Step 122: Crop the image data and interpretation data into sample data of the same size;

[0020] Step 123: Perform image enhancement processing on the cropped image data, wherein the image enhancement processing is histogram adjustment;

[0021] Step 124: Generate pseudo-color label samples from the label samples using color mapping;

[0022] Step 125: Construct a sample dataset using the results of Steps 123 and 124, and verify the validity of the sample dataset through the deep learning model training process to obtain valid data;

[0023] Step 126: Use the SAM segmentation model to classify the valid data, generate label information, and compare the label information with the valid data. If the overlap does not reach the threshold, the sample will not be included in the final sample dataset.

[0024] Furthermore, the specific steps for obtaining valid data in step 125 are as follows:

[0025] If the model training does not report any errors, and the upward trend of the model training accuracy is consistent with the downward trend of the loss function, indicating that the model is converging, then the sample dataset is valid data.

[0026] Furthermore, the SAM segmentation model is an improved SAM segmentation model, expressed as:

[0027] ,

[0028] in, Image with a resolution of 2 meters. For multispectral images, For the improved segmentation model, For sample data, For fusion algorithms, This is a superpixel segmentation algorithm.

[0029] Furthermore, the land category extraction filter is represented as follows:

[0030]

[0031] in, For the final land cover image, Results extracted from multiple land categories. The extraction results are for land category 1. To improve the accuracy of multi-category classification, For single-category classification accuracy, This represents the average classification accuracy for a single land category. The coefficient of variation for single land category classification accuracy. To merge the extraction results of multiple land types, This is for the fusion of multi-location category extraction results and multiple single-location category extraction results.

[0032] The beneficial effects of this invention are:

[0033] This application is based on 2-meter resolution domestic satellite remote sensing imagery, combined with prior knowledge, and employs a mapping strategy that combines general and regional feature modeling with layered land cover. The images obtained through this application demonstrate high land cover accuracy. Attached Figure Description

[0034] Figure 1 Flowchart for land cover mapping at 2-meter resolution;

[0035] Figure 2 Classification diagram of first-level and second-level categories Figure 1 ;

[0036] Figure 3 Classification diagram of first-level and second-level categories Figure 2 ;

[0037] Figure 4 Create a flowchart for the sample dataset;

[0038] Figure 5 A schematic diagram of partial ground cover in temperate broad-leaved forest and mixed forest areas;

[0039] Figure 6 A schematic diagram of local surface cover in tropical and subtropical arid broadleaf forest areas;

[0040] Figure 7 This is a schematic diagram showing the results of classifying and extracting local artificial land surfaces. Detailed Implementation

[0041] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0042] Specific Implementation Method 1: The global land cover mapping method based on 2-meter resolution domestic satellite imagery described in this implementation method includes the following steps:

[0043] Step 1: Use domestically produced satellites with a resolution of 2 meters to acquire global remote sensing image data, and obtain various types of land cover based on the global remote sensing image data;

[0044] Step 2: Based on prior knowledge, obtain global glacier surface cover and global wetland surface cover;

[0045] Step 3: Overlay various types of land cover, global glacier land cover, and global wetland land cover to obtain a global land cover image.

[0046] Satellite remote sensing images

[0047] The ZY-3 satellite imagery includes panchromatic imagery, multispectral imagery, and image metadata. The panchromatic imagery has a ground resolution of 2 meters, and the multispectral imagery has a ground resolution of 6 meters. It is in IMG format, WGS-84 coordinate system, UTM projection, 6° zoning, and the coordinate unit is meters. The imagery was acquired between 2015 and 2018.

[0048] Existing surface cover products

[0049] The GlobeLand30 dataset primarily includes 10 major land cover types: cultivated land, forest, grassland, shrubland, wetland, water bodies, tundra, artificial land, bare land, glaciers, and permanent snow. The ground resolution is 30 meters, in TIF format, WGS-84 coordinate system, UTM projection, 6° zoning, and the coordinate unit is meters. The dataset is current as of 2020.

[0050] The 10-meter ground cover products mainly include 21 categories such as paddy fields, dry land, arbor forests, other woodlands, shrublands, and grasslands. Vector data, ESRI Geodatabase format, WGS-84 coordinate system, unit is degrees.

[0051] wetland data

[0052] The wetland data includes a global database of lakes and wetlands, a wetland register, and global mangrove data.

[0053] The Global Lakes and Wetlands Database is in ArcGIS ShapeFile polygon data format, WGS-84 coordinate system, and units are degrees. The data is divided into three feature layers (GLWD-1, GLWD-2, and GLWD-3), containing vector boundaries and related attribute information for large lakes, reservoirs, and rivers worldwide. The Wetlands Register data is in ArcGIS ShapeFile point data format, WGS-84 coordinate system, and units are degrees. The Global Mangrove Data is in ArcGIS ShapeFile polygon data format, WGS-84 coordinate system, and units are degrees.

[0054] Glacier data

[0055] The data for the glacier inventory comes from the World Glacier Inventory, which includes information such as the latitude, longitude, and names of 107,009 glaciers worldwide.

[0056] To ensure the mapping accuracy of 2-meter resolution land cover products, this application proposes a strategy combining general and regional feature modeling with layered land cover mapping, forming a technical route and method suitable for the production of 2-meter resolution land cover products. The main technical process is described in [link to technical details]. Figure 1The first step is sample dataset creation, generating a labeled sample dataset based on historical data and prior products. The second step is model training and land cover extraction. For a global scale, natural and artificial land cover zoning is used, and land cover mapping combining single and multiple land cover types is performed based on a general + regional model. The image production process provides the main data source for land cover extraction. The third step is quality verification and post-processing, assessing and verifying the quality of the extracted land cover results, and then overlaying data such as glaciers and wetlands for further verification, ultimately forming the land cover product. Figure 1 As shown.

[0057] This application further refines the land cover classification system into secondary categories based on the commonly used primary categories. The land cover classification system includes 10 primary categories and 20 secondary categories. Specifically, as follows... Figure 2 and Figure 3 As shown.

[0058] Sample Dataset Creation

[0059] While no globally comprehensive 2-meter resolution land cover products have been reported, there are numerous 2-meter resolution land cover products or land cover classification sample data for local areas. Examples include the Gaofen Image Datasets (GID) (Tong et al., 2020), the bigearthnet dataset (Sumbul et al., 2019), and the Million-scale FinE-grained Geospatial Classification Datasets (MEET) (Li et al., 2025). Furthermore, 10-meter land cover products in global data resource development also utilize 2-meter resolution satellite remote sensing imagery, serving as effective sample data. However, for some developing countries, available sample data resources are relatively scarce.

[0060] To meet the needs of sample datasets in different partitions, two main technical approaches are adopted. The first is sample creation based on historical data, primarily involving automated sample creation and sample cleaning and refinement. The second is sample creation based on a general classification model, primarily involving segmentation models, interactive annotation, transfer learning, and sample cleaning and refinement.

[0061] like Figure 4 As shown. The main steps in creating a sample based on historical data include:

[0062] ① Perform coordinate system checks and preprocessing on historical remote sensing image data and interpreted data (vector or raster data) to ensure consistency in data coordinates and data format. If the interpreted data is vector data, it needs to be converted into raster data with the same resolution as the image data.

[0063] ②Based on the size of the classification samples, the image data and interpretation data are cropped into sample data of the same size;

[0064] ③ Perform image enhancement processing such as histogram adjustment on the cropped image data;

[0065] ④ Generate pseudo-color label samples by using color mapping to facilitate sample viewing;

[0066] ⑤ Label sample validity verification: The generated sample dataset is augmented using a deep learning network through random cropping, rotation, and other data enhancement techniques. Model training and evaluation are then performed. If the model training does not report any errors, the size, format, and classification labels of the sample dataset are correct. If the upward trend of the model's training accuracy and the downward trend of the loss function are essentially consistent, indicating that the model is converging, then the sample dataset is considered valid data.

[0067] ⑥ Sample data cleaning: The SAM (segment anything model) segmentation model is used to classify the image samples to be cleaned, generate label information, and compare it with the labeled samples. If the overlap does not reach the threshold, the sample will not be included in the final sample dataset.

[0068] The main steps in creating samples based on a general classification model include:

[0069] ① Use a general classification model to classify and interpret representative satellite remote sensing images, and process the interpretation results as sample data.

[0070] Steps ② to ⑥ are the same as described above.

[0071] ⑦ The generated sample dataset is used to form a land cover classification model suitable for this region through transfer learning, and steps ① to ⑥ are repeated. In step ⑥, the improved SAM is used for interactive annotation. The boundary and attributes of the labeled samples are adjusted through human-computer collaboration to refine the samples.

[0072] This application primarily employs an improved SAM visual segmentation model for sample refinement and interactive annotation. SAM utilizes the largest segmentation dataset to date, enabling zero-shot transfer to new image segmentation tasks to obtain the image mask. Its performance surpasses previous fully supervised results. However, at a global scale, directly utilizing SAM often results in poor segmentation within geographic regions and inaccurate boundaries. This application proposes: ① For natural geographic units with distinct vegetation features, fusing 2-meter resolution imagery with multispectral satellite imagery (more than 5 bands) better supports inter-type classification for forests, grasslands, and cultivated land (Formula 1). ② For natural geographic units dominated by man-made surfaces, superpixel segmentation of 2-meter resolution imagery followed by fusing with multispectral satellite imagery (more than 5 bands) yields better classification results and boundary features (Formula 2).

[0073] (1)

[0074] (2)

[0075] in, Image with a resolution of 2 meters. For multispectral images, For the improved segmentation model, For sample data, For fusion algorithms, This is a superpixel segmentation algorithm.

[0076] For regions with abundant historical data, a sample creation method based on historical data is adopted; for regions without historical data, a sample creation method based on a general classification model is mainly adopted; for regions with historical data but limited data and poor representativeness, a combination of the two methods is used to expand and refine the sample data and interpret the iterative update of the model.

[0077] Model training and land type extraction

[0078] (1) Land cover zoning

[0079] Land cover is influenced by a combination of factors, including topography, climate, and human activities, resulting in significant differences in land cover across different regions. This application divides natural land surfaces into 14 sub-regions based on the Terrestrial Ecoregions of the World (TEOW), and artificial land surfaces into 12 sub-regions based on the World Cultural Regions.

[0080] The Terrestrial Ecoregions of the World (TEOW) are biogeographical regions of Earth's terrestrial biodiversity. They are a comprehensive classification framework based on existing biogeographical knowledge. They can be divided into 14 distinct biomes, such as tropical and subtropical humid broadleaf forests, tropical and subtropical dry broadleaf forests, tropical and subtropical coniferous forests, temperate broadleaf mixed forests, temperate coniferous forests, northern forests / taiga forests, tropical and subtropical grasslands, savannas and shrublands, temperate grasslands, savannas and shrublands, floodplains and savannas, mountain grasslands and shrublands, tundra, Mediterranean forests, woodlands and shrublands, deserts and xerophytic shrublands, and mangroves.

[0081] A 2-meter resolution land cover model, excluding large land cover types such as forests, grasslands, and water bodies, can better support the subdivision of human-disturbed land cover types, such as man-made surfaces. However, the characteristics of man-made surfaces are influenced by regional culture, exhibiting significant differences in building styles, urban layouts, and urban greening. When extracting secondary classes of man-made surfaces, a world cultural partitioning model is used for regional modeling, which effectively avoids the problem of poor model generalization caused by differences in man-made surfaces within the same geographic ecological zone.

[0082] (2) Model training

[0083] Based on land cover zoning and using a general classification model, the sample dataset within the region is typically divided into training and validation sets in an 8:2 ratio for model training and accuracy verification. There are two main approaches to building models for land cover classification scenarios: one is single-land-class model construction, such as water bodies, forest land, cultivated land, and artificial land surfaces; the other is multi-land-class model construction, such as models for forest land, cultivated land, and grassland.

[0084] The single-land-class model is constructed, and the classification model is constructed by performing a binary classification task (i.e., classifying each pixel into the target land-class and other land-classes) each time. The regional sample dataset is reclassified according to the classification code, that is, multiple land-class classifications are merged into the target land-class and other land-classes. A deep learning framework is used for model training and accuracy verification.

[0085] The multi-land class model is constructed by using a single model to obtain all land cover types at once. The regional sample dataset is classified according to the needs of the multi-land class classification scenario, such as 10 primary classes and 20 secondary classes.

[0086] (3) Land category extraction

[0087] Based on the significant differences in vegetation, water bodies, and man-made structures under different land cover zones, land cover extraction results for each land cover type are generated through a land cover extraction filter.

[0088] The algorithm for land category extraction and filtering is as follows:

[0089]

[0090] in, For the final land cover image, Results extracted from multiple land categories. The extraction results are for land category 1. To improve the accuracy of multi-category classification, For single-category classification accuracy, This represents the average classification accuracy for a single land category. The coefficient of variation for single land category classification accuracy. To merge the extraction results of multiple land types, This is for the fusion of multi-location category extraction results and multiple single-location category extraction results.

[0091] 1) Multi-land cover classification and extraction results. If the coefficient of variation of the classification accuracy of each individual land cover type is less than 10%, and the difference between the overall classification accuracy of the multi-land cover extraction and the classification accuracy of each individual land cover type is less than 5%, then the multi-land cover classification strategy is adopted to interpret the data to be classified according to multiple land cover types in one go. This strategy is usually applied to scenarios where the differences between land cover types in the region are obvious and the distinction is high.

[0092] 2) Extraction results of single and multiple land cover types. By comprehensively analyzing the spatial, spectral, and boundary differences of land cover types such as forest, grassland, cultivated land, and artificial land within land cover zones, and considering that the coefficient of variation of the classification accuracy of each single land cover type is greater than 10%, the accuracy of single-type and multi-type extraction is compared to formulate a combined extraction strategy. For example, water area → artificial land + forest → cultivated land, or water area → artificial land secondary category (multiple categories) → paddy field + dry land, etc. This strategy is commonly used in scenarios where some land cover types within a region have obvious classification characteristics, while the differences in classification characteristics between other land cover types are small.

[0093] 3) Layered single-land-category land cover combination extraction results. If the coefficient of variation of the classification accuracy of each single land-category is less than 10%, and the difference between the overall classification accuracy of multi-land-category extraction and the classification accuracy of each single land-category is greater than 5%, then each land-category is extracted separately using a single-land-category model. The order of extraction is then determined based on the accuracy of each land-category extraction, typically from highest to lowest classification accuracy, to determine the masking order of the extracted land-category results. This method is usually applied to scenarios where the distribution of land-category areas within a region is relatively balanced, and the differences between land-category areas are small.

[0094] Post-treatment of land cover classification

[0095] Land cover classification primarily employs a strategy of first defining primary categories, followed by secondary categories. First, the distribution range of the primary categories is determined, then the distribution of secondary categories is distinguished within the spatial mask area of ​​the primary categories. Based on the primary category classification results, for wetland land cover, data from the Global Lakes and Wetlands Database, the Wetlands Register, and the Global Mangrove Register are overlaid to further optimize land cover patches; for glaciers, glacier register data is overlaid to further refine the glacier extent. Human-machine collaborative processing is used to handle unreasonable fragmentation and inappropriate patches, resulting in a primary category land cover map.

[0096] Within the scope of the primary land cover mask, secondary classifications are refined, such as artificial land cover classification, using a multi-land cover one-time classification strategy for land use classification. These hierarchical levels form the final land cover product.

[0097] Experimental Analysis / Results and Analysis

[0098] Natural land cover classification experiment

[0099] The natural land cover classification experiment selected temperate broad-leaved forest and mixed forest areas, and tropical and subtropical arid broad-leaved forest areas. Four sample areas were selected within each area, each measuring 25 km × 25 km, which is half the size of a YZ-3 satellite image. Detailed sample data for the two experimental areas are shown in Table 1.

[0100] Table 1. Proportion of different land types in the sample dataset of the experimental area

[0101]

[0102] Based on the single-location-class sample data, labeled samples were created at 512×512 pixels and divided into training and validation sets at an 8:2 ratio to construct a single-location-class classification model. The network model used was ResUNet, the pre-trained model used was ImageNet, the optimizer was SGD, the initial learning rate was set to 0.001, a cosine descent learning rate strategy was adopted, the batch size was 4, the loss function was Cross-Entropy, and training was performed 200 times. An iterative optimization strategy was used to retain the model with the highest accuracy. Two images outside the sample area were selected for interpretation, and 30,000 random points were randomly generated to create a confusion matrix for accuracy evaluation.

[0103] Accuracy is evaluated using overall accuracy (OA), precision, and recall. The calculation formulas are as follows:

[0104] ,

[0105] ,

[0106] ,

[0107] In the formula, TP is the number of pixels correctly predicted as the correct class; FP is the number of pixels incorrectly predicted as the correct class; TN is the number of pixels correctly predicted as the incorrect class; and FN is the number of pixels incorrectly predicted as the incorrect class.

[0108] Table 2 shows a comparison of the extraction accuracy of single land cover types in the two experimental areas. In the temperate broadleaf forest and mixed forest areas, the overall classification accuracy for cultivated land, forest land, artificial land surfaces, and water bodies was above 97%, with good precision and recall rates for individual land cover types. In the tropical and subtropical arid broadleaf forest areas, the overall classification accuracy for cultivated land, forest land, artificial land surfaces, and water bodies was above 92%, while the precision and recall rates for artificial land surfaces and water bodies were poor.

[0109] Table 2 Comparison of Extraction Accuracy of Single Land Category in Experimental Areas

[0110]

[0111] Based on the sample dataset, multi-location category label samples were created at 512×512 pixels and divided into training and validation sets at an 8:2 ratio for model training. The model training parameters remained consistent with the single-location category classification model. A confusion matrix was generated using 30,000 random points for single-location category validation to evaluate accuracy. Table 3 compares the multi-location category extraction accuracy of the two experimental areas. The overall classification accuracy for temperate broadleaf forests and mixed forests was 95.38%, with a Kappa coefficient of 0.84. The overall classification accuracy for tropical and subtropical arid broadleaf forests was 69.20%, with a Kappa coefficient of 0.503.

[0112] Table 3 Comparison of Extraction Accuracy of Multiple Land Categories in the Experimental Area

[0113]

[0114] Based on the accuracy comparison of single-land cover extraction and multi-land cover extraction in the two experimental areas, the differences in land cover characteristics between temperate broad-leaved forests and mixed forests are significant, with clear boundaries. The accuracy difference between single-land cover extraction and multi-land cover extraction is relatively small. Due to the layered single-land cover combination extraction, the results of single-land cover extraction at each level are masked and superimposed, which easily leads to the generation of fragmented and unreasonable patches. For this feature-based zoning, it is recommended to adopt a one-time classification strategy for multi-land cover. In tropical and subtropical arid broad-leaved forest areas, due to the small water area and indistinct water surface characteristics, the area is more often semi-arid. Natural vegetation such as forest land is affected by climatic conditions, resulting in low canopy closure and a certain area of ​​dry bare land surface, which easily causes confusion between forest land, artificial surfaces, and water bodies. For this feature-based zoning, it is recommended to adopt a layered single-land cover combination extraction strategy, with the priority order of land cover extraction as follows: cultivated land → forest land → artificial surfaces → water bodies. Figure 5 and Figure 6 As shown.

[0115] Artificial land cover classification experiment

[0116] For the classification of artificial land cover, the distribution area of ​​a single subclass of artificial land cover is relatively small. Therefore, a multi-subclass extraction method was adopted, and comparative experiments were conducted on the extraction of all subclasses and the extraction of multiple subclasses within the artificial land cover area. The comparison shows that determining the distribution range of artificial land cover first, and then performing multi-subclass extraction of subclasses, yields better data accuracy. (See Table 4 and...) Figure 7 As shown.

[0117] Table 4 Comparison of Accuracy in Artificial Land Surface Classification Extraction

[0118]

[0119] Conclusion / Discussion

[0120] With the enhanced acquisition capabilities of domestically produced high-resolution satellite remote sensing data, the temporal and coverage aspects of high-resolution satellite imagery in overseas regions are gradually improving. However, high-resolution global land cover products (better than 5 meters) have not yet been released. This application focuses on the production of global 2-meter resolution land cover products, analyzing the differences in natural land cover types worldwide due to variations in climate conditions and topography. Influenced by different human religious beliefs and lifestyles, artificial land cover also exhibits certain differences across regions. This application proposes a land cover classification sample design and optimization method based on historical data and prior knowledge, a mapping strategy combining general and regional feature modeling with layered land cover, and a two-level mapping method for artificial land cover. The feasibility of the technical route and method is verified through experiments, aiming to provide a reference for the production of high-resolution global land cover products.

[0121] This application primarily relies on 2-meter-resolution domestic public satellite remote sensing data resources. Based on historical data and prior knowledge, it constructs a sample dataset and proposes a strategy combining general and regional feature modeling with hierarchical land cover mapping. This strategy constructs a land cover mapping model applicable to different global landscape characteristics, improving the mapping accuracy and reliability of 2-meter-resolution land cover. Currently, common land cover mapping methods have low accuracy in classifying forest land subclasses and shrubland / grassland, making it difficult to obtain high-precision sample data. This application proposes a sample refinement method based on an improved SAM (Self-Assessing and Mapping Aspect) that can quickly expand and apply sample data. For the refined classification of artificial land surfaces, it proposes a technical route and method suitable for the production of 2-meter-resolution land cover products and conducts experimental verification, which is of great significance for the routine monitoring of artificial land surfaces.

[0122] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solution of the present invention and should not be used to limit the scope of protection. Any modifications made in accordance with the claims and specification of the present invention that are only partial should still fall within the protection scope of the present invention.

Claims

1. A global land cover mapping method based on 2-meter resolution domestic satellite imagery, characterized in that... Includes the following steps: Step 1: Use domestically produced satellites with a resolution of 2 meters to acquire global remote sensing image data, and obtain various types of land cover based on the global remote sensing image data; Step 2: Based on prior knowledge, obtain global glacier surface cover and global wetland surface cover; Step 3: Overlay various types of land cover, global glacier land cover, and global wetland land cover to obtain a global land cover image.

2. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 1, characterized in that... The specific steps of step 1 are as follows: Step 11: Acquire global remote sensing image data and input the global remote sensing image data into the terrestrial ecoregion model to obtain the world terrestrial ecoregion division results; Step 12: For each terrestrial ecological zone, use the land type extraction model to obtain the land type 1 extraction result. The land type 1 extraction result includes one or more of the following: cultivated land, forest land, shrubland, grassland, artificial surface, desert and bare surface, water area, glacier and perennial snow cover, wetland and tundra. For artificial surface, the artificial surface needs to be input into the regional culture model for further subdivision. After subdivision, the final land type 1 extraction result is obtained. The regional culture model is characterized by building style, urban layout and urban greening, and divides the input image into cultural zones. Step 13: Input the final land cover type 1 extraction result from Step 12 into the land cover extraction filter to obtain various types of land cover.

3. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 2, characterized in that... The world's terrestrial ecoregions include tropical and subtropical humid broad-leaved forests, tropical and subtropical dry broad-leaved forests, tropical and subtropical coniferous forests, temperate broad-leaved mixed forests, temperate coniferous forests, northern forests / taiga forests, tropical and subtropical grasslands, savannas and shrublands, temperate grasslands, savannas and shrublands, floodplains and savannas, mountain grasslands and shrublands, tundra, Mediterranean forests, woodlands and shrublands, deserts and xerophytic shrublands, and mangroves.

4. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 3, characterized in that... The global glacier surface cover was obtained through the World Glacier Inventory.

5. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 4, characterized in that... The global wetland surface cover was obtained from the Global Lakes and Wetlands Database.

6. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 5, characterized in that... The land use extraction model is a pre-trained land use extraction model, which is obtained by training using a sample dataset. The sample dataset is obtained through the following steps: Step 121: Acquire historical remote sensing image data and corresponding labeled land cover data. If the interpreted data is vector data, convert the vector data into raster data with the same resolution as the image data. Step 122: Crop the image data and interpretation data into sample data of the same size; Step 123: Perform image enhancement processing on the cropped image data, wherein the image enhancement processing is histogram adjustment; Step 124: Generate pseudo-color label samples from the label samples using color mapping; Step 125: Construct a sample dataset using the results of Steps 123 and 124, and verify the validity of the sample dataset through the deep learning model training process to obtain valid data; Step 126: Use the SAM segmentation model to classify the valid data, generate label information, and compare the label information with the valid data. If the overlap does not reach the threshold, the sample will not be included in the final sample dataset.

7. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 6, characterized in that... The specific steps for obtaining valid data in step 125 are as follows: If the model training does not report any errors, and the upward trend of the model training accuracy is consistent with the downward trend of the loss function, indicating that the model is converging, then the sample dataset is valid data.

8. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 7, characterized in that... The SAM segmentation model is an improved SAM segmentation model, expressed as: ， in, Image with a resolution of 2 meters. For multispectral images, For the improved segmentation model, For sample data, For fusion algorithms, This is a superpixel segmentation algorithm.

9. The global land cover mapping method based on 2-meter resolution domestic satellite imagery according to claim 8, characterized in that... The land category extraction filter is represented as follows: in, For the final land cover image, Results extracted from multiple land categories. The extraction results are for land category 1. To improve the accuracy of multi-category classification, For single-category classification accuracy, This represents the average classification accuracy for a single land category. The coefficient of variation for classification accuracy of a single land category. To merge the extraction results of multiple land types, This is for the fusion of multi-location category extraction results and multiple single-location category extraction results.

Images