A method for cloud platform-oriented downscaling reconstruction of 10-meter vegetation coverage

Abstract

The present application relates to the field of remote sensing image processing and machine learning, and is a method for cloud platform downscaling reconstruction of 10-meter vegetation coverage, comprising the following steps: S1: data selection; S2: data preprocessing; S3: three-dimensional layered sample collection; S4: grid sample weight distribution; S5: grid sample space optimization; S6: training sample generation; S7: establishment of FVC cross-scale model; S8: model verification; S9: model cross-scale application; S10: model consistency calibration post-processing. The present application comprehensively combines multi-source remote sensing data fusion and random forest downscaling model, and proposes a layered sample training and consistency constraint downscaling modeling scheme in the aspect of FVC spatial resolution improvement, successfully realizing spatial seamless conversion from 500-meter GLASS FVC to 10-meter high-resolution FVC.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing and machine learning, and is a method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform. Background Technology

[0002] Fractional vegetation cover (FVC) refers to the percentage of the vertically projected area of ​​vegetation (including leaves, stems, and branches) on the ground to the total area of ​​a statistical area. It is an important parameter for characterizing land surface vegetation cover. It plays an important role in studies on vegetation change, ecological environment assessment, soil and water conservation, and urban ecological evolution. It is also one of the essential parameters for many weather forecasting models, hydro-climate models, and other land surface models.

[0003] Most common FVC (Front-Veined Surface Cover) data are of medium or low spatial resolution, such as CYCOLPES FVC data, GEOV1-GEOV3 FVC series data, and GLASS FVC data with spatial resolutions of 500 to 1000 meters. These data are limited in certain urban areas and for small-scale fine-grained comparisons, failing to meet the needs for detailed surveys of local vegetation cover distribution. In recent years, with the development of remote sensing technology, research focus has gradually shifted to the extraction of refined land cover information, especially in areas with high spatial heterogeneity such as urban green spaces, desert vegetation, forest edges, and farmland. Medium and low spatial resolution vegetation cover products are insufficient to meet the practical needs of refined monitoring and management. Therefore, developing vegetation cover products with a spatial resolution ≤10m is of significant scientific and practical value for improving the quantitative accuracy of these special areas.

[0004] There are relatively few publicly available 10-meter spatial resolution vegetation cover data. Most methods used directly generate FVC regional-scale datasets using 10-meter resolution Sentinel-2 multispectral data. Wu Bingfang et al. (2025) used a pixel-based bisection model to generate 10-meter vegetation cover data for the Tibetan Plateau in 2020; Niu Bowen et al. (2025) constructed a bidirectional long short-term memory network model (BiLSTM) to generate monthly vegetation cover datasets for the Kherlen Basin in 2022. In addition, many academic studies have used hybrid pixel decomposition models, radiative transfer models, and data fusion methods to directly invert vegetation cover. However, these methods have high requirements for the quality, spatial continuity, and spatial resolution of the original remote sensing data. The results are easily affected by factors such as clouds, shadows, and atmosphere, and may also be affected by dense canopy and background features, making it difficult to obtain large-scale, continuous, and reliable vegetation cover products. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for downscaling and reconstructing 10-meter vegetation coverage on a cloud platform.

[0006] The specific technical solution is as follows:

[0007] A method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform includes the following steps:

[0008] S1: Data selection;

[0009] This includes high spatial resolution auxiliary data and low spatial resolution data;

[0010] The high spatial resolution auxiliary data is the GSE dataset, and the low spatial resolution data used is the GLASS FVC remote sensing data product.

[0011] S2: Data preprocessing;

[0012] GLASS FVC remote sensing products are integrated and uploaded to the GEE cloud platform for data preprocessing.

[0013] S3: Three-dimensional layered sample acquisition;

[0014] A three-dimensional stratified sampling strategy was adopted to select sample points: 7 levels of elevation, 19 types of climate, and 5 levels of vegetation cover (FVC).

[0015] S4: Grid sample weight allocation;

[0016] The entire China region was divided into grids according to latitude and longitude, and the sample size was allocated using area ratio weighting.

[0017] S5: Grid sample space optimization;

[0018] S6: Training sample generation;

[0019] All the sample points obtained from the grid are merged to generate a training sample set, which is then imported into the GEE cloud platform in shp format. The preprocessed FVC data at the location of the sample points and the corresponding pixel values ​​of the 500m GSE data are extracted and organized into a FeatureCollection, which is used as training samples for model building.

[0020] S7: Establish an FVC cross-scale model;

[0021] A random forest regression model was used to fit the nonlinear mapping relationship between the input variable GSE and the output variable FVC, and an FVC cross-scale model was established.

[0022] S8: Model Validation;

[0023] After the model is built, the root mean square error (RMSE) and the coefficient of determination (R²) are calculated using the validation set test_sample. 2 This is used to evaluate the performance of the regression model;

[0024] S9: Cross-scale application of models;

[0025] The trained model is applied across scales to the study region; the model is applied regionally; the study region is divided into 1°×1° grids, and the trained random forest regression model is applied to GSE feature data with a spatial resolution of 10m in each grid to obtain the FVC estimate with a spatial resolution of 10m. fine ;

[0026] S10: Post-processing for model consistency calibration;

[0027] The random forest downscaling model establishes a mapping relationship between GSE feature variables and FVC based on training samples. Then, the established model is applied to high spatial resolution feature variables to generate a high spatial resolution FVC dataset, thereby achieving spatial downscaling. Consistency calibration is then used for post-processing.

[0028] The reconstructed FVC was output in grid form to obtain an FVC with a spatial resolution of 10m for the entire study area, thus constructing a time series FVC downscaling production framework.

[0029] Further:

[0030] In S1, the GSE dataset is an annual dataset with a spatial resolution of 10m; the GLASS FVC remote sensing data product has a temporal resolution of 8 days and a spatial resolution of 500m, with the time span consistent with the GSE data.

[0031] In S2, 8 days of GLASS data are synthesized into annual-scale data, keeping the time interval consistent with that of GSE data; annual normalization is performed on the annual GLASS FVC data, and cross-year correction is implemented to ensure that all data values ​​are within the range of 0-100%; GSE data is called, and the dataset with a spatial resolution of 10m is reorganized into 500m by aggregating the average value, keeping the spatial resolution consistent with that of GLASSFVC data.

[0032] S3 specific three-dimensional layered sample acquisition:

[0033] Using DEM data, China is divided into 7 parts according to elevation: E={E1, E2, ..., E7}. The DEM value less than 0 represents the area below sea level (E1), the DEM value between 0 and 500 represents the low elevation area (E2), the DEM value between 500 and 1000 represents the medium elevation area (E3), the DEM value between 1000 and 2000 represents the mountain area (E4), the DEM value between 2000 and 3000 represents the high mountain area (E5), the DEM value between 3000 and 4000 represents the extremely high mountain area (E6), and the DEM value greater than 4000 represents the area above the snow line (E7).

[0034] The Köppen climate classification system was used to divide the study area into 19 categories, C = {C1, C2, ..., ...}. C19: C1 Tropical Monsoon Climate Am, C2 Tropical Woodland-Savanna Climate Aw, C3 Cold Desert Climate BWk, C4 Tropical Grassland Climate BSH, C5 Cold Grassland Climate BSK, C6 Hot Summer-Dry Summer-Warm Temperate Climate Csa, C7 Hot Summer-Dry Winter-Warm Temperate Climate Cwa, C8 Warm Summer-Dry Winter-Warm Temperate Climate Cwb, C9 Hot Summer-Humid Warm Temperate Climate Cfa, C10 Warm Summer-Humid Warm Temperate Climate Cfb, C11 Hot Summer-Dry Summer-Cold Temperate Climate Dsa, C12 Warm Summer-Dry Summer-Cold Temperate Climate Dsb, C13 Hot Summer-Winter-Dry Cold Temperate Climate Dwa, C14 Warm Summer-Winter-Dry Cold Temperate Climate Dwb, C15 Cold Summer-Winter-Dry Cold Temperate Climate Dwc, C16 Hot Summer-Humid Cold Temperate Climate Dfa, C17 Warm Summer-Humid Cold Temperate Climate Dfb, C18 Cold Summer-Humid Cold Temperate Climate Dfc, C19 Polar Tundra Climate ET.

[0035] The FVC of the area under study is divided into 5 levels F={F1, F2, F3, F4, F5}. FVC between 0 and 0.2 is low coverage F1, FVC between 0.2 and 0.4 is low-to-medium coverage F2, FVC between 0.4 and 0.6 is medium coverage F3, FVC between 0.6 and 0.8 is medium-to-high coverage F4, and FVC between 0.8 and 1 is high coverage F5.

[0036] In summary, each coverage type is represented by a triplet t=(Ei, Cj, Fk), where i=1, 2,...7, j=1, 2,...19, and k=1, 2,...5.

[0037] S4 specific grid sample weight allocation:

[0038] (1) Divide the entire study area into a 10°×10° grid according to latitude and longitude, and generate training samples in each grid area;

[0039] (2) The area of ​​the grid located in the region to be studied is counted; if it is not in the region to be studied, no sampling is performed.

[0040] (3) Count the area of ​​each category in the grid and allocate the number of samples according to the area ratio; for each category t with area, calculate the number of samples n to be allocated. t :

[0041]

[0042] Where A t A represents the area occupied by type t. total Let N be the total area of ​​the entire study region, and N be the total number of sampling points planned. t The number of sample points to be assigned to type t, taken as an integer, and ,like ;

[0043] (4) To ensure that each type has at least one representative sample, a mandatory minimum sample constraint is set when At least one sample point should be taken at a time.

[0044] S5 specific grid sample space optimization:

[0045] Within each grid, for each class t with area, randomly select sample points within the spatial range covered by class t; set the spatial distance for generating sample points. If the randomly generated points are too close together, then those points will be removed.

[0046] S7 specifically establishes the FVC cross-scale model:

[0047] Using the random forest function ee.Classifier.smileRandomForest in GEE, a random forest regression model RF for FVC and GSE feature variables is trained using the training set train_sample. The mapping relationship between the multidimensional features of GSE and GLASSFVC is nonlinearly fitted to establish the FVC cross-scale model. In the parameter selection, the number of random forest regression trees is set to 200, the feature subset is the square root of the feature number and is set to 8, and the minimum number of leaf node samples is set to 2.

[0048] In S10, a simple linear model is used to perform consistency constraints on the FVC estimate. Specifically, at a 500m scale, the established random forest regression model is applied to the GSE feature data at the 500m scale to estimate the FVC at a coarse resolution scale, thus obtaining the FVC. coarse Simultaneously, the predicted FVC values ​​with a spatial resolution of 10m are aggregated on a scale of 500m to obtain the FVC. aggregated Calculate FVC per pixel coarse With FVC aggregated Deviation δ 500The bias δ ​​is resampled to a spatial resolution of 10m and compared with FVC. fine Maintain consistency, then utilize δ for high-resolution FVC fine Term-by-term correction is performed to obtain FVC. correct This is the FVC reconstructed at high resolution;

[0049]

[0050] The reconstructed FVC was output in grid form to obtain an FVC with a spatial resolution of 10m for the entire study area, thus constructing a time series FVC downscaling production framework.

[0051] The technical effects of this invention are as follows:

[0052] 1. For the first time, 10-meter resolution FVC data at a nationwide scale was generated;

[0053] This invention proposes a method for downscaling and reconstructing 10-meter vegetation cover on cloud platforms. By integrating multi-source remote sensing data fusion and a random forest downscaling model, it generates 10-meter resolution FVC data at a national scale for the first time.

[0054] In terms of technical process, this invention establishes a representative sample system covering the three-dimensional space of altitude-climate-FVC through (1) a three-dimensional hierarchical sampling strategy across the country; (2) using GLASS FVC 500-meter resolution products as target variables and Google Satellite Embedding (GSE) 10-meter resolution feature datasets as main input variables, and using the GEE cloud computing platform to construct a nationwide random forest downscaling model, and nonlinearly fitting the mapping relationship between GSE multidimensional features and GLASS FVC; (3) generating nationwide 10-meter FVC products by batch inversion in different regions.

[0055] This method technically breaks through the previous limitation that high spatial resolution FVC products could only be generated locally or in a single region, and realizes the unified production of 10-meter FVC products nationwide, providing new data support for national-scale ecological monitoring and refined surface parameter inversion.

[0056] 2. A 10-meter seamless FVC product was generated by downscaling from a 500-meter unit;

[0057] In terms of improving the spatial resolution of FVC, this invention proposes a downscaling modeling scheme with hierarchical sample training and consistency constraints, which successfully achieves a seamless spatial conversion from 500-meter GLASS FVC to 10-meter high-resolution FVC.

[0058] At the model level, by building an FVC cross-scale model on the GEE platform, a nonlinear mapping relationship between GSE features and GLASS FVC is realized; in the production stage, by implementing a random forest downscaling model strategy in 1°×1° blocks, the efficiency of big data processing across the country is improved, and the problem of cloud platform computing overflow is avoided.

[0059] To address the systematic bias in local regions after downscaling, this invention further proposes a consistency constraint post-processing procedure: the model-predicted 10-meter FVC is upsampled to 500 meters and compared with the FVC predicted by the model at a 500-meter scale, and the systematic bias δ ​​is calculated. 500 Then, the deviation is resampled to a 10-meter resolution and the 10-meter FVC is corrected pixel by pixel to ensure that the data remains consistent and spatially continuous in a statistical sense before and after downscaling.

[0060] This method avoids the regional breakage and segmentation splicing marks commonly found in traditional FVC product zoning production, and achieves seamless spatial transition and continuous expression of FVC data across the country.

[0061] 3. Capable of producing annual FVC products, ensuring the annual variation characteristics of the data;

[0062] This invention constructs an annual FVC production framework, using GLASS FVC annual data and GSE data from 2017 to 2024 to generate 10-meter resolution national FVC products year by year.

[0063] In terms of time, by synthesizing GLASS 8-day data into an annual average, consistency with the time interval of the GSE annual-scale product is ensured. In terms of model construction, a random forest model is trained independently for each year to avoid systematic errors caused by cross-year data differences. Simultaneously, annual normalization and cross-validation are used to ensure that the model results for each year are comparable in time series. This guarantees that the spatial distribution trend between years remains consistent with the data before downscaling, ensuring that the downscaled FVC product can accurately reflect the annual changes in vegetation cover.

[0064] Compared with existing methods, this invention can not only reflect the annual dynamic changes of vegetation at high resolution, but also maintain the consistency and stability between data, providing high spatiotemporal accuracy data support for ecological evolution, carbon cycle and environmental response analysis.

[0065] 4. The latest data source (Google Satellite Embedding, GSE) was used, which is rich in information and can provide the spatial details necessary for downscaling.

[0066] This invention is the first to introduce Google Satellite Embedding (GSE) deep learning model feature data into the FVC downscaling process. The GSE data, officially released in July 2025, extracts multidimensional spectral, textural, and spatial semantic features from Google's high-resolution satellite imagery using a deep learning network, offering rich information and nationwide coverage.

[0067] This invention uses the 64-dimensional depth feature vector of GSE as input variables to establish a nonlinear fitting of the mapping relationship between multidimensional features and GLASS FVC. Compared with traditional downscaling methods that rely on band features or single optical indices (such as NDVI or EVI), this invention significantly enhances the model's detail recovery ability and spatial heterogeneity expression ability by utilizing the high-dimensional embedding features of GSE, achieving high-fidelity modeling of FVC with fine detail nationwide. Attached Figure Description

[0068] Figure 1 High-resolution Google imagery of the Weinan area, as shown in the example.

[0069] Figure 2 For the example, the GLASS 500m FVC in Weinan area before downscaling in 2023 is shown. The greener the color, the higher the FVC value.

[0070] Figure 3 The downscaled 10m FVC for the Weinan area in 2023 is used as an example. The greener the color, the higher the FVC value.

[0071] Figure 4 This is a flowchart of the present invention. Detailed Implementation

[0072] The specific technical solution of the present invention will be described with reference to the embodiments. The specific process of the present invention is shown below. Figure 4 As shown.

[0073] This embodiment takes the China region as an example and includes the following steps:

[0074] S1: Data Selection

[0075] The high spatial resolution auxiliary data used was the Google Satellite Embedding (GSE) dataset, an annual dataset with a spatial resolution of 10m and a time range of 2017-2024. The low spatial resolution data used was the GLASS (Global Land Surface Satellite) FVC remote sensing data product. This data has a temporal resolution of 8 days and a spatial resolution of 500m. To maintain consistency with the time span of the GSE data, the GLASS FVC data time range of 2017-2024 was selected.

[0076] S2: Data Preprocessing

[0077] The GLASS FVC remote sensing product was integrated and uploaded to the GEE cloud platform for data preprocessing. Eight days of GLASS data were synthesized into annual-scale data, maintaining consistency with the GSE data time interval. Annual normalization was performed on the yearly GLASS FVC data, with cross-year corrections to ensure all data values ​​fall within the 0-100% range. GSE data was then used, and the 10m spatial resolution dataset was consolidated into a 500m dataset through average aggregation, maintaining consistency with the GLASS FVC data spatial resolution.

[0078] S3: Three-dimensional layered sample acquisition

[0079] Building a downscaling model requires selecting representative training samples nationwide. Considering the significant east-west elevation variations and north-south climate differences in China, and the need for sample points to cover the entire range of FVC values, a three-dimensional stratified sampling strategy was adopted to select sample points: 7 elevation levels, 19 climate types, and 5 vegetation cover levels (FVC). The goal is to stratify China according to geographical distribution and environmental heterogeneity, allocating the number of samples within each stratum based on area proportion, while ensuring that each type with area has at least one sample point, thus balancing overall representativeness with the learnability of rare / small-area units.

[0080] Considering the significant elevation changes across China, DEM data is used to divide China into seven regions based on elevation: E={E1, E2, ..., E7}. Regions with DEM values ​​less than 0 are below sea level (E1); regions with DEM values ​​between 0 and 500 are low-altitude regions (E2); regions with DEM values ​​between 500 and 1000 are mid-altitude regions (E3); regions with DEM values ​​between 1000 and 2000 are mountainous regions (E4); regions with DEM values ​​between 2000 and 3000 are high-altitude regions (E5); regions with DEM values ​​between 3000 and 4000 are extremely high-altitude regions (E6); and regions with DEM values ​​greater than 4000 are above the snow line (E6).

[0081] Considering China's vast geographical span and significant climate variations, the Köppen climate classification method is used to divide China into 19 categories C={C1, C2, ..., C19}: C1 Tropical Monsoon Climate (Am), C2 Tropical Woodland-Savanna Climate (Aw), C3 Cold Desert Climate (BWk), C4 Tropical Steppe Climate (BSh), C5 Cold Steppe Climate (BSk), C6 Hot Summer-Dry Summer-Warm Temperate Climate (Csa), C7 Hot Summer-Dry Winter-Warm Temperate Climate (Cwa), C8 Temperate Summer-Dry Winter-Warm Temperate Climate (Cwb), C9 Hot Summer-Normal Humid Warm Temperate Climate (Cfa), C10 Temperate Summer-Normal Humid Warm Temperate Climate (Cf) b), C11 Hot summer, dry summer, cold temperate climate (Dsa), C12 Warm summer, dry summer, cold temperate climate (Dsb), C13 Hot summer, dry winter, cold temperate climate (Dwa), C14 Warm summer, dry winter, cold temperate climate (Dwb), C15 Cold summer, dry winter, cold temperate climate (Dwc), C16 Hot summer, humid, cold temperate climate (Dfa), C17 Warm summer, humid, cold temperate climate (Dfb), C18 Cold summer, humid, cold temperate climate (Dfc), C19 Polar tundra climate (ET).

[0082] To ensure that the sample points cover the entire FVC range of 0-1, the FVC in China is divided into 5 levels F={F1, F2, F3, F4, F5}. FVC in the range of 0-0.2 is low coverage F1, FVC in the range of 0.2-0.4 is low-medium coverage F2, FVC in the range of 0.4-0.6 is medium coverage F3, FVC in the range of 0.6-0.8 is medium-high coverage F4, and FVC in the range of 0.8-1 is high coverage F5.

[0083] In summary, each coverage type is represented by a triplet t=(Ei, Cj, Fk), where i=1, 2,...7, j=1, 2,...19, and k=1, 2,...5. In reality, many combinations have zero area within China; therefore, only combinations with an area greater than 0 should be sampled.

[0084] S4: Grid Sample Weight Allocation

[0085] (1) Considering that the training samples need to be evenly distributed within the Chinese region, the entire Chinese region is divided into a 10°×10° grid according to latitude and longitude, and training samples are generated in each grid region.

[0086] (2) The area of ​​the grid located in China is counted. If the area is not located in China, no sampling is performed.

[0087] (3) Calculate the area of ​​each category in the grid and allocate the number of samples according to the area proportion. For each category t with area, calculate the number of samples n to be allocated. t :

[0088]

[0089] Where A t A represents the area occupied by type t. total Let N be the total area of ​​the entire China region, and N be the total number of sampling points in the overall plan. t The number of sample points to be assigned to type t (take an integer, and ,like )

[0090] (4) To ensure that each type has at least one representative sample, a mandatory minimum sample constraint is set when At least one sample point should be taken at a time.

[0091] S5: Grid Sample Space Optimization

[0092] Within each grid cell, for each class t with an area, sample points are randomly selected within the spatial area covered by class t. To avoid excessive clustering of sample points, a spatial distance is set for generating sample points. If the randomly generated points are too close together, the points that are too close together will be removed.

[0093] S6: Training Sample Generation

[0094] All grid-obtained sample points were merged to generate a training sample set (sample), which was then imported into the GEE cloud platform in shapefile format. The preprocessed FVC data and the corresponding pixel values ​​from the 500m GSE data at each sample point location were extracted and organized into a FeatureCollection, which served as the training sample for model building. This sample contained FVC and 64 GSE feature variables, totaling 65 fields. The training sample was divided into a training set (train_sample) and a validation set (test_sample), with 70% used as the training set and 30% as the validation set.

[0095] S7: Establishing an FVC cross-scale model

[0096] Using the random forest function `ee.Classifier.smileRandomForest` in GEE, a random forest regression model `RF` was trained on the training set `train_sample` to obtain the FVC and GSE feature variables. The mapping relationship between the GSE multidimensional features and GLASSFVC was then nonlinearly fitted to establish a cross-scale FVC model. In the parameter selection, the number of random forest regression trees was set to 200, the feature subset (the square root of the feature number) was set to 8, and the minimum number of leaf node samples was set to 2.

[0097] S8: Model Validation

[0098] After the model is built, the root mean square error (RMSE) and the coefficient of determination (R²) are calculated using the validation set test_sample. 2 R0 is used to evaluate the performance of a regression model. RMSE characterizes the average deviation between predicted and actual values; a smaller value indicates lower prediction error and higher model accuracy. 2 It can measure the model's ability to explain the variation in actual data; the closer the value is to 1, the better the model's interpretability.

[0099] S9: Model Cross-Scale Application

[0100] The trained model was applied across scales to the Chinese region. Due to the large coverage area of ​​China, and considering the computational efficiency of the GEE cloud platform to avoid computational overflow, the model was applied regionally. The Chinese region was divided into 1°×1° grids. Within each grid, the trained random forest regression model was applied to GSE feature data with a 10m spatial resolution, obtaining the 10m spatial resolution FVC estimate. fine .

[0101] S10: Post-processing of model conformance calibration

[0102] The Random Forest downscaling model establishes a mapping relationship between GSE feature variables and FVC based on training samples. Then, the established model is applied to high spatial resolution feature variables to generate a high spatial resolution FVC dataset, thereby achieving spatial downscaling.

[0103] Considering the systematic bias in the FVC data before and after downscaling, a simple linear model is used to constrain the FVC estimates to ensure data consistency. At a 500m scale, the established random forest regression model is applied to GSE feature data at the 500m scale to estimate the FVC at a coarse resolution scale. coarse Simultaneously, the predicted FVC values ​​at a spatial resolution of 10m are aggregated at a scale of 500m to obtain the FVC. aggregated Calculate FVC per pixel coarse With FVC aggregated Deviation δ 500 The bias δ ​​is resampled to a spatial resolution of 10m and compared with FVC. fine Maintain consistency, then utilize δ for high-resolution FVC fine Term-by-term correction is performed to obtain FVC. correct This is the FVC after high-resolution reconstruction.

[0104]

[0105] The reconstructed FVC was output in grid form to obtain FVCs with a spatial resolution of 10m for the entire China region, and a time series FVC downscaling production framework for 2017-2024 was constructed.

[0106] Result evaluation:

[0107] (1) Model validation accuracy

[0108] A random forest regression model was established using the above method, and the model validation results from 2017 to 2024 are summarized in the table below. The model exhibited relatively stable accuracy from 2017 to 2024. The R² value remained generally between 0.89 and 0.93, indicating a good model fit. The RMSE value ranged from 0.078 to 0.094, with small fluctuations, indicating a generally low error level. In summary, the model showed good stability across different years, with high overall fit, small error, and strong temporal consistency and reliability.

[0109] Table 1. Model Validation Results (2017-2024)

[0110]

[0111] (2) Comparison of results before and after downscaling

[0112] The above method was used to downscale the GLASS 500m FVC data, and the spatial distribution before and after downscaling was shown. Taking the Weinan region of Shaanxi Province as an example, as... Figure 1 , Figure 2 and Figure 3 As shown. Figure 1 This is a high-resolution Google image of the Weinan area, which shows a mosaic distribution of towns, farmland, and roads, with mountains concentrated in the southeast region, indicating relatively strong surface heterogeneity. Figure 2 The image shows the GLASS FVC at a spatial resolution of 500m before downscaling in this region in 2023. A greener color indicates a higher FVC value. The image roughly shows the boundaries of urban areas, mountains, and plains. The "mosaic effect" becomes more pronounced as the scale increases. Figure 3 The FVC spatial distribution of the region after downscaling to 10m spatial resolution can roughly show the overall situation of vegetation cover in the region. It not only shows the boundaries of urban areas, mountain and plain boundaries, but also shows more detailed information such as villages, fields and roads more clearly. As the scale increases, the rivers and roads in the figure are clearer, which can better represent the small-scale spatial heterogeneity.

[0113] Compared to existing technologies, this method, based on the Google Earth Engine (GEE) cloud platform, utilizes the embedded learning results of the GSE feature set to establish a cross-scale FVC model and employs consistency calibration for post-processing to achieve FVC reconstruction with a spatial resolution of 10m. High spatial resolution FVC can more accurately characterize small-scale spatial heterogeneity, especially in areas where farmland, woodland, grassland, and bare land are interspersed, enabling clearer identification of these land boundary plots. Furthermore, it has significant implications for supporting refined ecological and agricultural management, improving climate and ecological models, and urban and regional environmental monitoring.

Claims

1. A method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform, characterized in that, Includes the following steps: S1: Data selection; This includes high spatial resolution auxiliary data and low spatial resolution data; The high spatial resolution auxiliary data is the GSE dataset, and the low spatial resolution data used is the GLASS FVC remote sensing data product. S2: Data preprocessing; GLASS FVC remote sensing products are integrated and uploaded to the GEE cloud platform for data preprocessing. S3: Three-dimensional layered sample acquisition; A three-dimensional stratified sampling strategy was adopted to select sample points: 7 levels of elevation, 19 types of climate, and 5 levels of vegetation cover (FVC). S4: Grid sample weight allocation; The entire China region was divided into grids according to latitude and longitude, and the sample size was allocated using area ratio weighting. S5: Grid sample space optimization; S6: Training sample generation; All the sample points obtained from the grid are merged to generate a training sample set, which is then imported into the GEE cloud platform in shp format. The preprocessed FVC data at the location of the sample points and the corresponding pixel values ​​of the 500m GSE data are extracted and organized into a FeatureCollection, which is used as training samples for model building. S7: Establish an FVC cross-scale model; A random forest regression model was used to fit the nonlinear mapping relationship between the input variable GSE and the output variable FVC, and an FVC cross-scale model was established. S8: Model Validation; After the model is built, the root mean square error (RMSE) and the coefficient of determination (R²) are calculated using the validation set test_sample. 2 This is used to evaluate the performance of the regression model; S9: Model cross-scale application; The trained model is applied across scales to the study region; the model is applied regionally; the study region is divided into 1°×1° grids, and the trained random forest regression model is applied to GSE feature data with a spatial resolution of 10m in each grid to obtain the FVC estimate at a spatial resolution of 10m. fine ; S10: Post-processing for model consistency calibration; The random forest downscaling model establishes a mapping relationship between GSE feature variables and FVC based on training samples. Then, the established model is applied to high spatial resolution feature variables to generate a high spatial resolution FVC dataset, thereby achieving spatial downscaling. Consistency calibration is then used for post-processing. The reconstructed FVC was output in grid form to obtain an FVC with a spatial resolution of 10m for the entire study area, thus constructing a time series FVC downscaling production framework.

2. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, In S1, the GSE dataset is an annual dataset with a spatial resolution of 10m; the GLASS FVC remote sensing data product has a temporal resolution of 8 days and a spatial resolution of 500m, with the time span consistent with the GSE data.

3. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, In S2, 8 days of GLASS data are synthesized into annual-scale data, keeping the time interval consistent with that of GSE data; annual normalization is performed on the GLASSFVC data year by year, and cross-year correction is implemented to ensure that all data values ​​are within the range of 0-100%; GSE data is called, and the dataset with a spatial resolution of 10m is reorganized into 500m by aggregating the average value, keeping the spatial resolution consistent with that of GLASS FVC data.

4. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, S3 specific three-dimensional hierarchical sample acquisition: Using DEM data, China is divided into 7 parts according to elevation: E={E1, E2, ..., E7}. The DEM value less than 0 represents the area below sea level (E1), the DEM value between 0 and 500 represents the low elevation area (E2), the DEM value between 500 and 1000 represents the medium elevation area (E3), the DEM value between 1000 and 2000 represents the mountain area (E4), the DEM value between 2000 and 3000 represents the high mountain area (E5), the DEM value between 3000 and 4000 represents the extremely high mountain area (E6), and the DEM value greater than 4000 represents the area above the snow line (E7). The Köppen climate classification system was used to divide the study area into 19 categories, C = {C1, C2, ..., ...}. C19: C1 Tropical Monsoon Climate Am, C2 Tropical Woodland-Savanna Climate Aw, C3 Cold Desert Climate BWk, C4 Tropical Grassland Climate BSH, C5 Cold Grassland Climate BSK, C6 Hot Summer-Dry Summer-Warm Temperate Climate Csa, C7 Hot Summer-Dry Winter-Warm Temperate Climate Cwa, C8 Warm Summer-Dry Winter-Warm Temperate Climate Cwb, C9 Hot Summer-Humid Warm Temperate Climate Cfa, C10 Warm Summer-Humid Warm Temperate Climate Cfb, C11 Hot Summer-Dry Summer-Cold Temperate Climate Dsa, C12 Warm Summer-Dry Summer-Cold Temperate Climate Dsb, C13 Hot Summer-Winter-Dry Cold Temperate Climate Dwa, C14 Warm Summer-Winter-Dry Cold Temperate Climate Dwb, C15 Cold Summer-Winter-Dry Cold Temperate Climate Dwc, C16 Hot Summer-Humid Cold Temperate Climate Dfa, C17 Warm Summer-Humid Cold Temperate Climate Dfb, C18 Cold Summer-Humid Cold Temperate Climate Dfc, C19 Polar Tundra Climate ET. The FVC of the area under study is divided into 5 levels F={F1, F2, F3, F4, F5}. FVC between 0 and 0.2 is low coverage F1, FVC between 0.2 and 0.4 is low-to-medium coverage F2, FVC between 0.4 and 0.6 is medium coverage F3, FVC between 0.6 and 0.8 is medium-to-high coverage F4, and FVC between 0.8 and 1 is high coverage F5. In summary, each coverage type is represented by a triplet t=(Ei, Cj, Fk), where i=1, 2,...7, j=1,2,...19, and k=1, 2,...

5.

5. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, S4's specific training sample size allocation rules: (1) Divide the entire study area into a 10°×10° grid according to latitude and longitude, and generate training samples in each grid area; (2) The area of ​​the grid located in the region to be studied is counted; if it is not in the region to be studied, no sampling is performed. (3) Count the area of ​​each category in the grid and allocate the number of samples according to the area ratio; For each category t with area, calculate the number of assigned samples n. t : ； Where A t A represents the area occupied by type t. total Let N be the total area of ​​the entire study region, and N be the total number of sampling points planned. t The number of sample points to be assigned to type t, taken as an integer, and ,like ; (4) To ensure that each type has at least one representative sample, a mandatory minimum sample constraint is set when At least one sample point should be taken at a time.

6. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, S5 specific grid sample space optimization: Within each grid, for each class t with area, randomly select sample points within the spatial range covered by class t; set the spatial distance for generating sample points. If the randomly generated points are too close together, the points that are too close together will be removed.

7. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, S7 specifically establishes the FVC cross-scale model: Using the random forest function ee.Classifier.smileRandomForest in GEE, a random forest regression model RF for FVC and GSE feature variables is trained using the training set train_sample. The mapping relationship between GSE multidimensional features and GLASS FVC is nonlinearly fitted to establish an FVC cross-scale model. In the parameter selection, the number of random forest regression trees is set to 200, the feature subset is the square root of the feature number and is set to 8, and the minimum number of leaf node samples is set to 2.

8. The method for downscaling and reconstructing 10-meter vegetation cover on a cloud platform according to claim 1, characterized in that, In S10, a simple linear model is used to perform consistency constraints on the FVC estimate. Specifically, at a 500m scale, the established random forest regression model is applied to the GSE feature data at the 500m scale to estimate the FVC at a coarse resolution scale, thus obtaining the FVC. coarse Simultaneously, the predicted FVC values ​​with a spatial resolution of 10m are aggregated on a scale of 500m to obtain the FVC. aggregated Calculate FVC per pixel coarse With FVC aggregated Deviation δ 500 The bias δ ​​is resampled to a spatial resolution of 10m and compared with FVC. fine Maintain consistency, then utilize δ for high-resolution FVC fine Term-by-term correction is performed to obtain FVC. correct This is the FVC reconstructed at high resolution; ； ； The reconstructed FVC was output in grid form to obtain an FVC with a spatial resolution of 10m for the entire study area, thus constructing a time series FVC downscaling production framework.

Images