Global cropland mapping method based on coupling prior knowledge and weakly supervised learning strategy

Abstract

The application discloses a global cultivated land mapping method based on coupling prior knowledge and a weakly supervised learning strategy; a hierarchical sampling strategy is established based on a simulated annealing algorithm, a certain number of primary selection grids are divided in the overlapping area of a global cultivated land data set, the optimal screening grid is selected through the simulated annealing algorithm, and representative point-shaped cultivated land samples and point-shaped non-cultivated land samples are further selected in each optimal screening grid through the simulated annealing algorithm; an integrated learning method is used to expand the sparse point-shaped sample set into a planar sample set; NDVI time-series image data and six global cultivated land data products are used as input data sources, and the cultivated land and non-cultivated land prediction results of the planar sample set are used as training labels to train a U-Net model and complete global cultivated land mapping. The application provides a stable high-precision mapping method for various global cultivated land recognition by using a deep learning model and various prior knowledge.

Description

Technical Field

[0001] This invention relates to the field of remote sensing feature information extraction technology, specifically to a global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategies. Background Technology

[0002] Global food demand is expanding with a growing population, placing immense pressure on global food production. Meanwhile, climate change and market volatility may further exacerbate this pressure. Furthermore, ensuring food security while protecting the environment remains a significant challenge. To effectively alleviate food production pressures and address environmental challenges, a thorough understanding of the quantity and detailed geospatial distribution of global arable land is crucial.

[0003] Remote sensing technology is currently the most efficient method for large-scale monitoring and identification of arable land distribution. Over the past two decades, numerous low-to-medium resolution (100–1000 m) global arable land cover products based on remote sensing satellites have been developed. These products have provided macroscopic mapping of global arable land distribution and significantly advanced research in global arable land monitoring and management. However, due to their low spatial resolution, these products are heavily affected by pixel mixing and are prone to missing small areas of arable land, resulting in lower mapping reliability in highly heterogeneous regions.

[0004] In recent years, due to the increase in freely available medium-to-high resolution satellite data sources (such as the Landsat and Sentinel-2 series satellites) and the gradual enhancement of cloud platform computing power, more and more medium-to-high spatial resolution (10~30m) global arable land products have emerged. The challenge of medium-to-high spatial resolution global arable land products lies in the need for effective management of large amounts of remote sensing satellite data, coordination of computing resources, and acquisition of large sets of real training data samples.

[0005] Furthermore, mapping of globally covered arable land products requires complex geography. In these regions, particularly geographically transitional areas, areas with complex topography, and areas with high landscape heterogeneity, existing global arable land products exhibit significant inconsistencies in classification results. These discrepancies in arable land area and spatial distribution mapping results can be attributed to unscientific training sample distribution, high levels of noise in the input remote sensing data, and outdated classification methods.

[0006] Therefore, there is an urgent need for a sampling and classification method to further improve the accuracy of global arable land mapping, in order to enhance the accuracy of global arable land identification, especially for regions with complex geographical features. Summary of the Invention

[0007] The purpose of this invention is to provide a global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategies, which can effectively improve the accuracy and precision of farmland extraction based on remote sensing.

[0008] The present invention achieves the above objectives mainly through the following technical means: A global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategies includes the following steps: Step S1: Acquire NDVI time-series image data; Step S2: Obtain global climate data, global DEM data, and six types of global arable land datasets; Step S3: Based on NDVI time-series imagery data, global climate data, global DEM data, and six global farmland datasets, use the simulated annealing algorithm to select a set number of optimal datasets. Grid, determining each optimal filter The final point samples of the grid are generated, and each optimal filter is selected. The final point sample set of the grid; Step S4: Based on each optimal filter The final point samples of the grid, along with NDVI time-series imagery data and six types of global farmland data, were used to train and validate the selected machine learning models. Based on the validation results, the optimal model that meets the requirements was selected. The grid serves as a set of areal samples, and the predicted farmland and non-farmland land use results from the areal sample set are used as labels. Step S5: Train the U-Net deep learning model based on the label and areal sample set. After the U-Net deep learning model is trained, input the global NDVI time-series image data and six global farmland datasets to perform global farmland and non-farmland identification calculations, and finally obtain the global farmland mapping results.

[0009] As described above, step 3 is based on the following steps: Step S31: Divide the overlapping areas of six global arable land datasets into N initial selections. Grid, for each initial selection The grid is rated for "inconsistency", with levels including "very high", "high", "medium" and "low". Step S32, based on the initial selection The mesh is determined using simulated annealing algorithm to identify the final optimal selection mesh set. The selection meshes in the final optimal selection mesh set are the optimal selection meshes. Grid; Step S33: Use the simulated annealing algorithm to determine each optimal filter in the optimal filter grid set. The final point samples in the grid are divided into "arable land" and "non-arable land" to obtain point-based arable land sample sets and point-based non-arable land sample sets. The point-based arable land sample sets and point-based non-arable land sample sets constitute the final point sample set.

[0010] As mentioned above in the preliminary selection The final optimal mesh set is determined using the simulated annealing algorithm based on the following steps: Step S321, in all initial selections Within the grid, select the initial number of filters corresponding to each level of "inconsistency". In the grid, the number of selections corresponding to the "inconsistency" level decreases sequentially from high to low. A total of four "inconsistency" levels filter out the set number of items. The grid forms the selection grid set; the historical best grid objective function value and the historical best selection grid set are initialized; Step S322: Calculate the minimum objective function value of the current filter grid set. If the minimum objective function value of the pre-selected grid set If the objective function value is less than the historical best grid value, update the historical best grid objective function value to the minimum objective function value of the current selected grid set. The minimum objective function value The corresponding set of filter grids is used as the historical best set of filter grids; Step S323: Repeat steps S321 to S322 until the historical optimal filtering grid set stops updating. The final historical optimal filtering grid set is taken as the final optimal filtering grid set, and the filtering grids in the final optimal filtering grid set are the optimal filtering grids. Grid.

[0011] As described above, calculate the minimum objective function value of the current filter grid set. Based on the following formula: ; ; in, Represents continuous variables within the total grid set; Indicates the category number of a continuous variable; Indicates the number of categories for a continuous variable; Indicates the initial selection in the total grid set The first in the grid Six continuous variables, specifically including six types of global arable land datasets, in which arable land area accounts for respectively The grid area ratio, NDVI time-series image data, and global DEM data; Indicates filtering within the grid set The grid number; Filtering in a grid collection Number of grids; and This indicates that the range of each continuous variable within the total grid set is... When divided equally, the first in the total grid set The first type of continuous variable range The quantile and the first quantile quantiles, when hour, The first in the total grid set The minimum value of the range of a continuous variable, when hour, The first in the total grid set The maximum value of the range of a continuous variable; express Located in the interval [ Internal filtering Number of grids; Representing discrete variables, This indicates the total number of climate categories. This indicates the first corresponding global climate data. Climate type; Indicates that it is classified as the first Screening of climate types Number of grids; Indicates having the first Preliminary selection of climate types The percentage of a grid in the total grid set; This indicates determining the first [number] in the filter grid set. Filter Grid, and calculate the filter. Grid and Filter: Other Remaining Filters in the Grid Set The distance between grids; Indicates the first Filter Grid and Filter: Other Remaining Filters in the Grid Set Minimum distance between grids This represents the area of ​​the total grid set; For the preliminary selection Representativeness bias of grid continuous variables For the preliminary selection Representativeness bias of grid discrete variables For the preliminary selection Grid nearest neighbor index.

[0012] As described above, the simulated annealing algorithm is used to determine each optimal filter in the optimal filter grid set. The final point samples in the grid are based on the following steps: Step S331: In any optimal selection Random selection within the grid A set of point samples is used as the point sample set. The objective function value of the historical optimal point sample and the historical optimal point sample set are initialized. Step 332: Calculate the minimum objective function value corresponding to the point sample set. If the minimum objective function value corresponding to the point sample set If the objective function value of the current point sample is less than the historical best objective function value, update the historical best objective function value to the minimum objective function value of the current point sample set. The minimum objective function value The corresponding point sample set serves as the historical optimal point sample set; Step S333: Repeat steps S331 to S332 until the historical optimal state sample set no longer changes. The final historical optimal state sample set is used as the optimal selection. The final optimal state sample set of the grid.

[0013] As described above, calculate the minimum objective function value corresponding to the point sample set. Based on the following formula: ; ; in, For the continuity deviation of point samples, The nearest neighbor index for point samples; Indicates optimal selection Continuous variables in the grid; Indicates the category number of a continuous variable; Indicates the number of categories for a continuous variable; Represents the first point of the point sample in the point sample set. The continuous variables include the proportion of cultivated land area to the grid area in six global cultivated land datasets and the NDVI time-series image data values. This indicates the sequence number of each point sample; Indicates the number of point samples in the point sample set; and This indicates the optimal selection. The range of each continuous variable in the grid is... When dividing equally, the optimal selection The first grid The first type of continuous variable range The quantile and the first quantile quantiles; express Located in the interval [ The number of point samples within; Indicates an optimal screening The first determined in the grid A number of point samples are selected, and the relationship between these point samples and the optimal selection criteria is calculated. The distance between other remaining point samples in the grid; Indicates optimal selection The first determined in the grid A point sample and the optimal screening The minimum distance between other remaining point samples in the grid. Indicates the optimal selection of point samples The area of ​​the grid.

[0014] As described above, step 4 is based on the following steps: Step S41: Divide the final point sample set into a point sample training dataset and a point sample validation dataset; Step S42: Select four machine learning models: Random Forest, Gradient Boosting Decision Tree, Multilayer Perceptron, and Support Vector Machine, based on optimal selection. The point-like training sample set within the grid is used to optimally select each... Four corresponding machine learning models are trained within the grid; each is optimally selected. NDVI time-series imagery data and global farmland data products within the grid serve as input features for various machine learning models, which are then used to train these models. Each model is then optimized. Four machine learning models trained within the grid are used to predict arable land and non-arable land. The logistic regression algorithm is used to integrate the arable land and non-arable land prediction results of the four machine learning models. Step S43, Optimal Selection The point-based validation dataset within the grid, containing corresponding NDVI time-series imagery and global farmland data, is input into various pre-trained machine learning models. A logistic regression algorithm is then used to integrate the farmland and non-farmland prediction results from the four models. The integrated farmland and non-farmland prediction results are then validated, and the optimal selection is made, retaining those with an overall accuracy greater than 95% and an F1 score greater than 0.95. The grid serves as a isometric sample set, and the predicted farmland and non-farmland land values ​​from the isometric sample set are used as labels.

[0015] As described above, step 5 is based on the following steps: Step S51: Based on the label, the corresponding area sample set U-Net deep learning model was trained using NDVI time-series imagery data within the grid and six global farmland datasets. Step S52: Input global NDVI time-series image data and global six types of arable land datasets into the trained U-Net deep learning model to perform global arable land and non-arable land identification calculations, and finally obtain global arable land mapping results.

[0016] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the above-described method.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0018] By adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention considers factors such as farmland fragmentation, spectral characteristics, climate, and topography globally, and selects a predetermined number of farmland inconsistencies as an initial screening factor. The mesh, through simulated annealing algorithm in the initial selection Optimal filtering in the grid Grid, in each optimal filter In the grid, the simulated annealing algorithm is used to further select representative point-like cultivated land samples and point-like non-cultivated land samples, making the location of the point-like samples more scientific and reasonable, and more conducive to cultivated land mapping and identification in areas with complex geographical features.

[0019] 2. By integrating multiple machine learning algorithms with the U-Net model, deep semantic information is extracted; the optimal joint model is obtained through parameter optimization, which can achieve farmland extraction and identification with a small amount of training data, and effectively improves the accuracy and precision of global farmland extraction. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention.

[0021] Figure 2 This is a comparison of the accuracy of the global arable land distribution results of this invention with other global arable land data products. The left vertical axis represents the overall accuracy (OA), and the right vertical axis represents the F1-score. Detailed Implementation

[0022] This embodiment provides a global farmland mapping method based on coupled prior knowledge and a weakly supervised learning strategy. This addresses the problems of unreliable mapping results in heterogeneous regions due to unreasonable sample location selection in existing technologies, as well as the low accuracy of the final farmland mapping. To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] like Figure 1 As shown, the global farmland mapping method based on coupled prior knowledge and a weakly supervised learning strategy includes the following steps: Step S1: Download Landsat-7 and Landsat-8 time-series remote sensing images, and perform image de-clouding and de-snowing processing, image co-processing, median quarterly composite processing, and Normalized Difference Vegetation Index (NDVI) calculation to obtain de-clouding and de-snowing NDVI time-series image data. Then, perform interpolation processing on the de-clouding and de-snowing NDVI time-series image data to obtain NDVI time-series image data. The specific process of step S1 is as follows: Step S11: Download Landsat-7 and Landsat-8 time-series remote sensing images respectively using the Google Earth Engine (GEE) platform. Data was downloaded via the GEE platform, including time-series remote sensing images of all available Landsat-7 ETM+ and Landsat-8 OLI sensors worldwide from January 1, 2019 to December 31, 2021. The downloaded bands include near-infrared and red bands, with a spatial resolution of 30 m.

[0024] Step S12: Perform cloud removal, snow removal, and cirrus removal processing on the downloaded Landsat-7 and Landsat-8 time-series remote sensing images. Identify missing observations in Landsat-7 and remove invalid pixels using image metadata information. In this embodiment, the CFMASK algorithm was used to remove pixels in Landsat-7 and Landsat-8 time-series remote sensing images affected by clouds, cloud shadows, and snow. Furthermore, due to a malfunction in the Landsat-7 ETM+ sensor, some Landsat-7 time-series remote sensing images contained missing pixels. The locations of these missing pixels were identified using image metadata, and abnormal pixel values ​​were removed.

[0025] Step S13: Use the least squares regression method to perform image co-processing on Landsat-7 and Landsat-8 time-series remote sensing images, then perform median quarterly composite and NDVI vegetation index calculation to obtain cloud- and snow-free NDVI time-series image data. Then, perform interpolation processing on the cloud- and snow-free NDVI time-series image data to obtain NDVI time-series image data.

[0026] As a preferred option, missing pixel values ​​in de-clouded and de-snowed NDVI time-series images are filled by interpolating pixel values ​​from de-clouded and de-snowed NDVI time-series images of adjacent years.

[0027] Given the spectral differences between the Landsat-7 ETM+ and Landsat-8 OLI sensors, ordinary least squares regression was used to adjust the reflectance values, aligning the reflectance bands of the Landsat-7 and Landsat-8 time-series remote sensing images. Subsequently, this invention used the 2020 Landsat-7 and Landsat-8 time-series remote sensing images to perform median composite analysis on the reflectance data of the Landsat-7 and Landsat-8 time-series remote sensing images. The composite analysis period was three months, resulting in composite images. For each composite image, this invention used the near-infrared and red bands to calculate the NDVI, obtaining the normalized difference vegetation index (NDVI) for each composite image, thus constructing NDVI time-series image data, as shown in formula (1): (1) in, and It refers to the spectral reflectance in the near-infrared and red bands.

[0028] Finally, to fill in the missing or invalid pixels in the 2020 composite imagery due to cloud, cloud shadows, and snow, this invention uses pixel values ​​from corresponding cloud- and snow-free NDVI time-series images from 2019 and 2021 to fill in the missing pixel values ​​from the 2020 cloud- and snow-free NDVI time-series images. This yields four global median quarterly composite NDVI time-series image data from 2020 with no missing values.

[0029] Step S2: Collection and preprocessing of global climate data, global DEM data (DEM, Digital Elevation Model), and six types of global arable land datasets; The specific process of step S2 is as follows: Step S21: Collect global climate data, global DEM data, and six global arable land datasets.

[0030] This embodiment downloaded global climate data (https: / / hess.copernicus.org / articles / 11 / 1633 / 2007 / ) and global DEM data, specifically the GMTED2010 global DEM data. The spatial resolution of the global climate data is [missing information]. (Longitude interval × Latitude interval), the spatial resolution of the GMTED2010 global DEM data is 2km.

[0031] In addition, this embodiment also downloaded six global land cover classification data products from around 2020, extracted the six global land cover classification data products according to the "arable land" land category code, and obtained six global arable land data products. The specific information is shown in Table 1. The datasets corresponding to the six global arable land data products are the six global arable land datasets.

[0032] Table 1 Six Global Farmland Products

[0033] Step S22: Crop the global climate data, global DEM data, and six global farmland datasets; project the global climate data to ensure that the global climate data, global DEM data, and six global farmland datasets have the same projection coordinate system; resample the global climate data, global DEM data, and six global farmland datasets using the nearest neighbor method, with a unified spatial resolution of 30 m.

[0034] Step S3: Based on NDVI time-series imagery data, global climate data, global DEM data, and six global farmland datasets, use the simulated annealing algorithm to select a set number of optimal datasets. Grid, further determining each optimal filter The final point sample locations of the grid are used to further obtain point-based cultivated land sample sets and point-based non-cultivated land sample sets, which together constitute the final point sample set. The specific process of step S3 is as follows: Step S31: Divide the overlapping areas of six global arable land datasets into N initial selections. (Longitude interval × Latitude interval) grid, by evaluating each initial selection The representativeness of the grid in terms of NDVI values, climatic conditions, and topographic features for each initial selection. The grid is rated for "inconsistency", with levels including "very high", "high", "medium" and "low".

[0035] In this embodiment, 16,284 initial selections are made from overlapping regions in the global arable land dataset. Grid, 16284 initial selections The grid, as a total grid set, utilizes each of the six initial selections from global arable land datasets. Farmland inconsistency within the grid was categorized into four levels: "Very High," "High," "Medium," and "Low." This grading was based on an initial selection. The number of consistent cells among the six global farmland datasets within the grid. For example, if the six global farmland datasets classify a cell consistently, the cell is marked as "consistent"; otherwise, it is marked as "inconsistent." For each initial selection... If 50% or more of the cells in a grid are marked as "inconsistent," then the initial selection... The level of grid inconsistency is categorized as "very high". If the proportion of "inconsistent" cells is between 30% and 50%, then the initial selection... The degree of grid inconsistency is categorized as "high"; if the proportion of "inconsistent" pixels is between 10% and 30%, then the initial selection is... The degree of grid inconsistency is rated as "moderate"; if the proportion of "inconsistent" pixels is less than 10%, then the initial selection is considered "inconsistent". The grid inconsistency level is "low".

[0036] Step S32, based on the initial selection The mesh is determined using simulated annealing algorithm to identify the final optimal selection mesh set. The selection meshes in the final optimal selection mesh set are the optimal selection meshes. Grid; specifically: Step S321, in all initial selections Within the grid, select the initial number of filters corresponding to each level of "inconsistency". The grid, with four levels of "inconsistency," filters out a set number of filters (1000 in this example). The grids form the selection grid set, where N is the initial selection. The total number of grids, N=16284; initialize the historical best grid objective function value and the historical best selected grid set; Considering that areas with inconsistent "very high" and "high" levels pose a greater challenge to accurately identifying arable land, this invention selects a first predetermined number (400 in this embodiment) of areas with a "very high" level of inconsistency. Grid as a filter The grid, selecting a second set number (300 in this example) of initial selections with a "high" degree of inconsistency, is used. Grid as a filter The grid is initially selected based on a third set number of grids (200 in this example) with a "moderate" degree of inconsistency. Grid as a filter The grid, and the initial selection of a fourth set number (100 in this embodiment) with a "low" degree of inconsistency. Grid as a filter Grid. The number of "inconsistencies" decreases sequentially from high to low for the first, second, third, and fourth settings. Each filter... The grid forms a set of filter grids.

[0037] Step S322: Calculate the minimum objective function value of the current filter grid set using formulas (2) and (3). If the minimum objective function value of the pre-selected grid set If the objective function value is less than the historical best grid value, update the historical best grid objective function value to the minimum objective function value of the current selected grid set. The minimum objective function value The corresponding set of filter grids is used as the historical best set of filter grids; Step S323: Repeat steps S321 to S322 until the historical optimal filtering grid set no longer updates. Based on this method, the final historical optimal filtering grid set is taken as the final optimal filtering grid set, and the filtering grids in the final optimal filtering grid set are the optimal filtering grids. A grid was used to ensure that it could broadly represent the diverse agricultural landscapes around the world.

[0038] To obtain the optimal set of filtering grids, this embodiment uses three metrics: initial selection... Representativeness bias of grid continuous variables Preliminary selection Representativeness bias of grid discrete variables and primary election Grid nearest neighbor index ( Preliminary selection Representativeness bias of grid continuous variables and primary election Representativeness bias of grid discrete variables Used to measure 16,284 primary selections In the grid, the representativeness deviation between grids; initial selection Grid nearest neighbor index Used to evaluate 16,284 preliminary selections Initial selection of a set number of items randomly selected within the grid. The spatial dispersion of the grid is shown in Equation (2): (2)

[0039] Represents continuous variables within the total grid set; Indicates the category number of a continuous variable; This indicates the number of categories for a continuous variable. This represents the total grid set (16284 initial selections). Preliminary selection in the grid) The first in the grid Six continuous variables, specifically including six types of global arable land datasets, in which arable land area accounts for respectively The initial selection was based on the grid area ratio, NDVI time-series image data, and global DEM data. The NDVI time-series image data values ​​of the grid are initially selected. The average value of the NDVI time-series image data for all points in the grid, initially selected. The global DEM data values ​​of the grid are initially selected. The average value of global DEM data for all points in the grid; Indicates filtering within the grid set The grid number; Filtering in a grid collection The number of grid cells in this step, ; and This indicates that the range of each continuous variable within the total grid set is... When divided equally, the first in the total grid set The first type of continuous variable range (the range formed by the minimum and maximum values ​​of the variable) The quantile and the first quantile quantiles, among which hour, The first in the total grid set The minimum value of the range of a continuous variable. hour, The first in the total grid set The maximum value of the range of a continuous variable; express Located in the interval [ Internal filtering The number of grids.

[0040] Represents discrete variables; This indicates the total number of climate categories; in this embodiment, there are 13 categories. This indicates the first of 13 climate types corresponding to global climate data. Climate type; Indicates that it is classified as the first Screening of climate types Number of grids; Indicates having the first Preliminary selection of climate types The percentage of a grid in the total grid set.

[0041] This indicates determining the first [number] in the filter grid set. Filter Grid, and calculate the filter. Grid and Filter: Other Remaining Filters in the Grid Set The distance between grids; Indicates the first Filter Grid and Filter: Other Remaining Filters in the Grid Set Minimum distance between grids. This represents the area of ​​the total grid set.

[0042] In the calculation , and Then, construct the minimum objective function value corresponding to the set of filtering grids. ), Filter the minimum objective function value corresponding to the grid set Consider each filter The representativeness of the mesh in terms of spatial properties and attribute characteristics is shown in formula (3): (3)

[0043] Step S33: Use the simulated annealing algorithm to determine each optimal filter in the optimal filter grid set. The final point samples in the grid, combined with the seasonal NDVI trend of the final point samples over the past three years (2019-2021 in this example), are divided into "cultivated land" and "non-cultivated land," resulting in a point-based cultivated land sample set and a point-based non-cultivated land sample set. The point-based cultivated land sample set and the point-based non-cultivated land sample set constitute the final point sample set; specifically, the following steps are included: Step S331: In any optimal selection Random selection within the grid We initialize the objective function value of the historical optimum sample and the historical optimum sample set with a set of point samples. Step 332: Since the terrain and climate conditions within the selected grid are relatively similar, input the proportion of cultivated land area to the selected grid area and the NDVI time series data value of each of the six global cultivated land data products of each point sample in the point sample set into formulas (4) and (5) to calculate the minimum objective function value corresponding to the point sample set. If the minimum objective function value corresponding to the point sample set If the objective function value of the current point sample is less than the historical best objective function value, update the historical best objective function value to the minimum objective function value of the current point sample set. The minimum objective function value The corresponding point sample set serves as the historical optimal point sample set; (4) (5)

[0044] For the continuity deviation of point samples, The nearest neighbor index for point samples; Indicates optimal selection Continuous variables in the grid; Indicates the category number of a continuous variable; Indicates the number of categories for a continuous variable; Represents the first point of the point sample in the point sample set. The continuous variables include the proportion of cultivated land area to the grid area in six global cultivated land datasets and the NDVI time-series image data values. This indicates the sequence number of each point sample; This indicates the number of point samples in the point sample set; in this step, ; and This indicates the optimal selection. The first point sample of each grid The range of continuous variables is When dividing equally, the optimal selection The first grid The first type of continuous variable range (the range formed by the minimum and maximum values ​​of the variable) The quantile and the first quantile quantiles, among which hour, For optimal screening The first grid The minimum value of the range of a continuous variable, taking the value 0. hour, For optimal screening The first grid The maximum value within the range of continuous variables is the optimal selection factor. In the grid The maximum value of a continuous variable; express Located in the interval [ The number of point samples within.

[0045] Indicates an optimal screening The first determined in the grid A number of point samples are selected, and the relationship between these point samples and the optimal selection criteria is calculated. The distance between other remaining point samples in the grid; Indicates optimal selection The first determined in the grid A point sample and the optimal screening The minimum distance between other remaining point samples in the grid; Indicates the optimal selection of point samples The area of ​​the grid.

[0046] Step S333: Repeat steps S331 to S332 until the historical optimal state sample set no longer changes. Based on this method, the final historical optimal state sample set is used as the optimal selection. The final optimal state sample set of the grid.

[0047] Step S334: Classify the optimal point samples of all optimal point sample sets into point-based cultivated land sample sets and point-based non-cultivated land sample sets. The point-based cultivated land sample sets and point-based non-cultivated land sample sets constitute the final point-based sample sets; the method is as follows: Further analysis of the seasonal variation trend of the Normalized Difference Vegetation Index (NDVI) of the optimal morphological samples over the past three years (2019-2021) revealed that samples with an NDVI curve peak value greater than 0.2 over the three years were labeled as point-shaped cultivated land samples, while the remaining optimal morphological samples were labeled as point-shaped non-cultivated land samples. All point-shaped cultivated land samples constituted the point-shaped cultivated land sample set, and all point-shaped non-cultivated land samples constituted the point-shaped non-cultivated land sample set. The point-shaped cultivated land sample set and the point-shaped non-cultivated land sample set together constituted the final point-shaped sample set.

[0048] Step S4: Based on each optimal filter The final point samples of the grid, along with NDVI time-series imagery data and six types of global farmland data, were used to train and validate the selected machine learning models. Based on the validation results, the optimal model that meets the requirements was selected. The grid serves as a set of areal samples, and the farmland prediction results and non-farmland prediction results of the areal sample set are used as labels. The specific process of step S4 is as follows: Step S41: Divide the final point sample set into a point sample training dataset and a point sample validation dataset, wherein in each optimal selection... 60% of the final point samples in the grid are randomly selected as point training samples, and each point training sample constitutes the point training dataset. The remaining 40% of the point samples are used as point validation samples, and each point validation sample constitutes the point validation dataset.

[0049] Step S42: Select four widely used machine learning methods, including Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Multilayer Perceptron (MLP), and Support Vector Machine (SAM), as base classifiers, and then select the optimal method. Point-like training samples within the grid, for each optimal selection Four machine learning models corresponding to the grid were trained. Each model was then used for optimal selection. The NDVI time-series imagery data and global farmland data corresponding to the grid-like training samples are used as input features for various machine learning models to train them, and each model is then optimized. Four machine learning models trained within the grid are used to predict arable and non-arable land. Then, a logistic regression algorithm is used to integrate the arable and non-arable land prediction results of the four machine learning models.

[0050] Step S43, Optimal Selection The point-based validation dataset within the grid, containing corresponding NDVI time-series imagery and global farmland data, is input into each pre-trained machine learning model. A logistic regression algorithm is then used to integrate the farmland and non-farmland prediction results from the four models. The integrated farmland and non-farmland prediction results are validated, and the optimal selection is made, retaining those with an overall accuracy (OA) greater than 95% and an F1 score greater than 0.95. The grid serves as a planar sample set. Ultimately, 1000 optimal samples are selected. 97 optimal filters were removed from the grid. Grid, 903 optimal filters remaining The grid serves as a isometric sample set, and the farmland prediction results and non-farmland prediction results of the isometric sample set are used as labels.

[0051] Step S5: Based on the labels and the corresponding areal sample sets The U-Net deep learning model is trained using NDVI time-series imagery data within the grid and six global farmland datasets. After the U-Net deep learning model is trained, the NDVI time-series imagery data (global scope) and the six global farmland datasets (global scope) are input to obtain global farmland mapping results.

[0052] The specific process of step S5 is as follows: Step S51: Divide the areal sample set obtained in step S43 into 80% areal samples (the optimal selection corresponding to the areal sample set). The grid consists of planar samples, rounded to 722. NDVI time-series imagery within the grid and six global farmland datasets were used as input features for training. Labels were used to guide model training and were used for U-Net deep learning model training. 20% of the samples were areal samples (the optimal selection corresponding to the areal sample set). The grid consists of planar samples, rounded to 181. The NDVI time-series imagery data within the grid and six global farmland datasets were used to verify the accuracy of the U-Net deep learning model's predictions of farmland and non-farmland.

[0053] Step S52: Input global NDVI time-series image data and global six types of arable land datasets into the trained U-Net deep learning model to perform global arable land and non-arable land identification calculations, and finally obtain global arable land mapping results.

[0054] Finally, based on the 181 areal samples corresponding to The accuracy of global farmland mapping results was validated using NDVI time-series imagery data from six global farmland datasets within the grid. The results are as follows: Figure 2 As shown, the accuracy verification results demonstrate that the global arable land mapping results of this invention have the best OA value of 89.8%, proving its reliability in accurately identifying the location of global arable land. Furthermore, this invention achieved an F1-score of 0.852, the highest among all data products. The overall accuracies of the ESA World Cover, ESRI Land Cover, GLC_FCS30, FROM-GLC, GLAD, and GlobeLand30 global arable land datasets are 87.5%, 84.8%, 82.3%, 81%, 86.8%, and 85%, respectively; their F1-scores are 0.81, 0.77, 0.78, 0.72, 0.79, and 0.80, respectively.

[0055] Example 2: In this embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments 1.

[0056] Example 3: In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps in the above-described method embodiments 1.

[0057] Example 4: In this embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments 1.

[0058] It should be noted that the embodiments described in this invention are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A global cultivated land mapping method based on coupled prior knowledge and weakly supervised learning strategy, characterized in that, Includes the following steps: Step S1: Acquire NDVI time-series image data; Step S2: Obtain global climate data, global DEM data, and six types of global arable land datasets; Step S3: Based on NDVI time-series imagery data, global climate data, global DEM data, and six global farmland datasets, the overlapping areas of the six global farmland datasets are divided into N initial selections. The grid, as the overall grid set, is based on the initial selection. The grid is determined using a simulated annealing algorithm to identify the final optimal selection grid set. Specifically, this involves considering all initial selections... Filtering within a grid to select the set total number of elements. The grid forms a selection grid set. The historical best grid objective function value and the historical best selection grid set are initialized. The minimum objective function value of the current selection grid set is then calculated. Based on the minimum objective function value The final optimal filter grid set is determined; the filter grids in the final optimal filter grid set are the optimal filters. Grid, determining each optimal filter The final point samples of the grid are generated, and each optimal filter is selected. The final point sample set of the grid; Step S4: Based on each optimal filter The final point samples of the grid, along with NDVI time-series imagery data and six types of global farmland data, were used to train and validate the selected machine learning models. Based on the validation results, the optimal model that meets the requirements was selected. The grid serves as a set of areal samples, and the farmland prediction results and non-farmland prediction results of the areal sample set are used as labels. Step S5: Train the U-Net deep learning model based on the label and areal sample set. After the U-Net deep learning model is trained, input the global NDVI time-series image data and six global farmland datasets to obtain the global farmland mapping results. Calculate the minimum objective function value of the current filter grid set. Based on the following formula: ； ； in, Represents continuous variables within the total grid set; Indicates the category number of a continuous variable; Indicates the number of categories for a continuous variable; Indicates the initial selection in the total grid set The first in the grid Six continuous variables, specifically including six types of global arable land datasets, in which arable land area accounts for respectively The grid area ratio, NDVI time-series image data, and global DEM data; Indicates filtering within the grid set. The grid number; Filtering in a grid collection Number of grids; and This indicates that the range of each continuous variable within the total grid set is equal. When divided equally, the first [unit] in the total grid set The first type of continuous variable range The quantile and the first quantile quantiles, when hour, The first in the total grid set The minimum value of the range of a continuous variable, when hour, The first in the total grid set The maximum value within the range of a continuous variable; express Located in the interval [ Internal filtering Number of grids; Represents discrete variables. This indicates the total number of climate categories. This indicates the first corresponding global climate data. Climate type; Indicates that it is classified as the first Screening of climate types Number of grids; Indicates having the first Preliminary selection of climate types The percentage of a grid in the total grid set; This indicates determining the first [number] in the filter grid set. Filter Grid, and calculate the filter. Grid and Filter: Other Remaining Filters in the Grid Set The distance between grids; Indicates the first Filter Grid and Filter: Other Remaining Filters in the Grid Set Minimum distance between grids This represents the area of ​​the total grid set; For the preliminary selection Representativeness bias of grid continuous variables For the preliminary selection Representativeness bias of grid discrete variables For the preliminary selection Grid nearest neighbor index.

2. The global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategy according to claim 1, characterized in that, Step S3 is based on the following steps: Step S31: Divide the overlapping areas of six global arable land datasets into N initial selections. Grid, for each initial selection The grid is rated on a "non-consistency" level, which includes "very high", "high", "medium" and "low"; Step S32, based on the initial selection The mesh is determined using simulated annealing algorithm to identify the final optimal selection mesh set. The selection meshes in the final optimal selection mesh set are the optimal selection meshes. Grid; Step S33: Use the simulated annealing algorithm to determine each optimal filter in the optimal filter grid set. The final point samples in the grid are divided into "arable land" and "non-arable land" to obtain point-based arable land sample sets and point-based non-arable land sample sets. The point-based arable land sample sets and point-based non-arable land sample sets constitute the final point sample set.

3. The global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategy according to claim 2, characterized in that, The preliminary selection The final optimal mesh set is determined using the simulated annealing algorithm based on the following steps: Step S321, in all initial selections Within the grid, select the initial selection of the corresponding set number of filters for each level of "inconsistency". The grid uses a sequence of "inconsistency" levels, with the number of corresponding selections decreasing sequentially from high to low. A total of four "inconsistency" levels filter out the set number of items to be selected. The grid forms the selection grid set; the historical best grid objective function value and the historical best selection grid set are initialized; Step S322: Calculate the minimum objective function value of the current filter grid set. If the minimum objective function value of the pre-selected grid set If the objective function value is less than the historical best grid value, update the historical best grid objective function value to the minimum objective function value of the current selected grid set. The minimum objective function value The corresponding set of filter grids is used as the historical best set of filter grids; Step S323: Repeat steps S321 to S322 until the historical optimal filtering grid set stops updating. The final historical optimal filtering grid set is taken as the final optimal filtering grid set, and the filtering grids in the final optimal filtering grid set are the optimal filtering grids. Grid.

4. The global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategy according to claim 3, characterized in that, The simulated annealing algorithm is used to determine each optimal filter in the optimal filter grid set. The final point samples in the grid are based on the following steps: Step S331: In any optimal selection Random selection within the grid A set of point samples is used as the point sample set. The objective function value of the historical optimal point sample and the historical optimal point sample set are initialized. Step 332: Calculate the minimum objective function value corresponding to the point sample set. If the minimum objective function value corresponding to the point sample set If the objective function value of the current point sample is less than the historical best objective function value, update the historical best objective function value to the minimum objective function value of the current point sample set. The minimum objective function value The corresponding point sample set serves as the historical optimal point sample set; Step S333: Repeat steps S331 to S332 until the historical optimal state sample set no longer changes. The final historical optimal state sample set is used as the optimal selection. The final optimal state sample set of the grid.

5. The global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategy according to claim 4, characterized in that, The minimum objective function value corresponding to the point sample set is calculated. Based on the following formula: ； ； in, For the continuity deviation of point samples, The nearest neighbor index for point samples; Indicates optimal selection Continuous variables in the grid; Indicates the category number of a continuous variable; Indicates the number of categories for a continuous variable; Represents the first point of the point sample in the point sample set. The continuous variables include the proportion of cultivated land area to the grid area in six global cultivated land datasets and the NDVI time-series image data values. This indicates the sequence number of each point sample; Indicates the number of point samples in the point sample set; and This indicates that the optimal selection will be performed. The first point sample of each grid The range of continuous variables is When dividing equally, the optimal selection The first grid The first type of continuous variable range The quantile and the first quantile quantiles; express Located in the interval [ The number of point samples within; Indicates an optimal screening The first determined in the grid A number of point samples are selected, and the relationship between these point samples and the optimal selection criteria is calculated. The distance between other remaining point samples in the grid; Indicates optimal selection The first determined in the grid A point sample and the optimal screening The minimum distance between other remaining point samples in the grid. Indicates the optimal selection of point samples The area of ​​the grid.

6. The global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategy according to claim 5, characterized in that, Step S4 is based on the following steps: Step S41: Divide the final point sample set into a point sample training dataset and a point sample validation dataset; Step S42: Select four machine learning models: Random Forest, Gradient Boosting Decision Tree, Multilayer Perceptron, and Support Vector Machine, based on optimal selection. The point-like training sample set within the grid is used to optimally select each... Four corresponding machine learning models are trained within the grid; each is optimally selected. NDVI time-series imagery data and global farmland data products within the grid serve as input features for various machine learning models, which are then used to train these models. Each model is then optimized. Four machine learning models trained within the grid are used to predict arable land and non-arable land. The logistic regression algorithm is used to integrate the arable land and non-arable land prediction results of the four machine learning models. Step S43, Optimal Selection The point-based validation dataset within the grid, containing corresponding NDVI time-series imagery and global farmland data, is input into each pre-trained machine learning model. A logistic regression algorithm is then used to integrate the farmland and non-farmland prediction results from the four models. The integrated farmland and non-farmland prediction results are then validated, and the optimal selection is made, retaining those with an overall accuracy greater than 95% and an F1 score greater than 0.

95. The grid serves as a isometric sample set, and the predicted farmland and non-farmland land values ​​from the isometric sample set are used as labels.

7. The global farmland mapping method based on coupled prior knowledge and weakly supervised learning strategy according to claim 6, characterized in that, Step S5 is based on the following steps: Step S51: Based on the label, the corresponding area sample set U-Net deep learning model was trained using NDVI time-series imagery data within the grid and six global farmland datasets. Step S52: Input global NDVI time-series image data and global six types of arable land datasets into the trained U-Net deep learning model to perform global arable land and non-arable land identification calculations, and finally obtain global arable land mapping results.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images