A remote sensing land cover data fusion method based on category accuracy geographical weighting

Abstract

This invention discloses a remote sensing land cover data fusion method based on category-accuracy geographic weighting, comprising: preprocessing multi-source land cover products to unify spatial reference and classification systems; constructing a validation sample set using stratified random sampling combined with visual interpretation; providing two methods for determining the number of neighborhood samples N, with an adaptive method dynamically adjusting N based on sample density and product category consistency, and adjusting the neighborhood radius to adapt to the sample distribution; calculating pixel scales GWUA and GWPA based on spatial weights, calculating the fusion probability using GWUA, and arbitrating the average GWPA when balancing. This invention achieves spatially adaptive estimation of accuracy, improving the accuracy and stability of the fusion results, and providing reliable data support for ecological monitoring and land management.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information processing technology, and specifically to a method for fusion of remote sensing land cover data based on category-accuracy geographic weighting. Background Technology

[0002] Land cover data is crucial foundational data for ecological environment monitoring, climate change research, land resource management, and Earth system science research. With the development of remote sensing technology, land cover mapping products have evolved from 100-meter to 10-meter resolutions, continuously improving the finer representation of surface information. In recent years, with the development of high-resolution remote sensing technology, big data processing technology, and machine learning technology, various high-resolution remote sensing land cover products at global and national scales with 30-meter and 10-meter resolutions have emerged, including CLCD, FROM-GLC30, GLC_FCS30, Globeland30, FROM-GLC10, WorldCover10, ESRI-LC10, and Dynamic World. However, due to differences in data sources, training samples, classification algorithms, and mapping strategies, the classification accuracy of different land cover products often varies significantly across different regions and land cover types.

[0003] To fully utilize the advantages of various remote sensing land cover products and reduce the impact of errors from single products on certain land types, it is necessary to conduct research on multi-source remote sensing land cover data fusion. By integrating information from multiple products, more reliable land cover data can be generated, thereby improving the overall accuracy and stability of land cover data. Currently, the commonly used land cover product fusion method is the voting method. Its basic idea is to statistically analyze the classification results of multiple land cover products at the same pixel or sample point and determine the final category through majority voting. Existing research shows that this method can improve the reliability of classification results to a certain extent.

[0004] However, the error distribution of remote sensing land cover products is influenced by factors such as topography, climate, and the spatial distribution of training samples, typically exhibiting significant spatial non-stationary characteristics. Existing voting fusion methods operate at a global scale, failing to fully consider the strong spatial heterogeneity of land cover data and the uneven distribution of conflicting categories among products. They neither specifically adapt to the accuracy requirements of different regions nor effectively address the issues of varying product accuracy and uneven distribution of category conflicts across regions. This results in fusion results with insufficient spatial rationality and significant accuracy deviations in some areas, further impacting the practicality and reliability of the fused data and making it difficult to meet the needs of refined, high-precision land cover monitoring and applications. Furthermore, existing voting-based fusion methods typically treat all input products as equally reliable sources, failing to fully consider the differences in classification accuracy across different land cover categories. During the voting process, when multiple categories receive the same number of votes, existing technologies lack scientific criteria for regionalized category accuracy, easily leading to errors in land cover classification and further reducing the spatial consistency and practical reliability of the fusion results. Summary of the Invention

[0005] To address the problems of existing multi-source remote sensing land cover product fusion methods, such as failure to consider spatial non-stationarity of accuracy, insufficient rationality of global voting, unreasonable allocation of credibility under equal weighting, and lack of discrimination mechanism in the case of a tie, the purpose of this invention is to provide a remote sensing land cover data fusion method based on geographical weighting of category accuracy. This method introduces the concept of geographical weighting, dynamically estimates the local category accuracy weights based on the verification samples around the pixels to be fused, and makes refined fusion decisions at the local scale, thereby generating land cover classification results that are more consistent with spatial characteristics and improving the accuracy and stability of the fused data.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] A remote sensing land cover data fusion method based on category-accuracy geographic weighting, the method comprising the following steps:

[0008] S1, acquire multi-source remote sensing land cover product data, preprocess and perform consistency transformation on the product data to ensure that the multi-source remote sensing land cover product data maintains consistency in spatial reference and classification system;

[0009] S2, based on the comprehensive area ratio of each category in each land cover product as the stratification basis, constructs a unified real category verification sample set based on stratified random sampling and high-resolution image visual interpretation;

[0010] S3. Based on the spatial distribution density of the validation samples in the study area, the number of local neighborhood samples N corresponding to each pixel is adaptively determined. For each pixel to be fused, the spatial neighborhood radius is dynamically adjusted so that the neighborhood contains exactly N validation samples. The spatial neighborhood radius corresponding to different pixels adapts to the sample distribution and the degree of class conflict. N is a natural number. For each land cover product, spatial weights are constructed based on the spatial distance between the sample and the pixel. The closer the distance, the greater the weight. The spatial weights are used to perform weighted calculations on the validation samples in the neighborhood to obtain the geographic weighted user accuracy (GWUA) and geographic weighted producer accuracy (GWPA) of each product at the pixel scale.

[0011] S4. Calculate the fusion probability of each category under the pixel based on the geographic weighted user precision of each product at the pixel to be fused, combined with the product classification results.

[0012] S5. Select the category with the highest fusion probability as the candidate category of the pixel. If multiple categories have the same maximum fusion probability, for each candidate category, calculate the average of the geographic weighted producer precision of all land cover products that classify the pixel as that category, and determine the category with the largest average as the final fusion category of the pixel.

[0013] Step S1 further includes:

[0014] Acquire data from multiple remote sensing land cover products, perform unified projection conversion on different land cover products, and make each product have the same spatial reference system;

[0015] For land cover products with different spatial resolutions, the highest resolution is used as the benchmark, and the maximum area weighting method is used to resample each land cover product.

[0016] Based on the scope of the study area, each land cover product is spatially cut, and the area outside the study area is masked to ensure that different land cover products have a one-to-one correspondence in the same spatial location.

[0017] A unified classification system is established and a classification mapping relationship is constructed, so that the categories of various land cover products are uniformly classified into this unified classification system.

[0018] Step S2 further includes:

[0019] Based on the comprehensive area ratio of various land cover types in the study area, stratified random sampling was used to generate candidate sample points in the study area, so that the samples are representative and spatially balanced among different land types.

[0020] High-resolution satellite imagery is used to visually interpret the locations of each candidate sample point. Historical imagery, multi-temporal information, and ground texture are combined to determine the true ground cover type. Sample points with blurred images, mixed ground covers, or ambiguity are removed. Multiple rounds of verification and interpretation are performed on samples of key ground cover types or transitional areas to generate the final true category verification sample dataset.

[0021] Furthermore, in step S3, the process of adaptively determining the number of local neighborhood samples N corresponding to each pixel based on the spatial distribution density of the validation samples in the study area includes the following steps:

[0022] An initial statistical window with a fixed radius is constructed centered on each pixel to be fused. The number of validation samples within the statistical window is calculated, and the local sample density is determined.

[0023] The study area is divided into several density levels according to local sample density. Each density level has a preset number of basic neighborhood samples. The higher the local sample density, the smaller the corresponding number of basic neighborhood samples.

[0024] The number of basic neighborhood samples is assigned according to the density zone to which the pixel belongs. Then, the number of basic neighborhood samples is adaptively adjusted according to the degree of inconsistency between the categories of multi-source land cover products within the local window. The higher the degree of inconsistency between products, the greater the adjustment of the number of basic neighborhood samples. Thus, the final adaptive number of local neighborhood samples N for each pixel is obtained.

[0025] Further, in step S3, the formula for calculating the geographic weighted user precision (GWUA) is:

[0026] ;

[0027] in, Indicates the position of pixel i. Geographically weighted user precision for land cover types Let J be the spatial weight of sample j relative to pixel i within the neighborhood window, and J be the total number of samples identified as type i within the neighborhood window. The number of validation samples; Let be a binary indicator function, if the j-th element in the neighborhood is identified as category... The sample, whose true category is category If the value is 1, then the value is 1; otherwise, the value is 0. Spatial weight The calculation formula is:

[0028] ;

[0029] In the formula, The spatial distance between sample j and current pixel i; Let be the bandwidth at pixel i, defined as the distance within the neighborhood that contains exactly a predetermined number of N samples.

[0030] Further, in step S3, the formula for calculating the Geographically Weighted Producer Precision (GWPA) is:

[0031] ;

[0032] in, Indicates the position of pixel i. Geographically weighted producer precision for land cover types The spatial weights of sample j within the neighborhood window. Type within the neighborhood window The actual number of samples; Let be a binary indicator function, if the j-th true class in the neighborhood is a class The samples were correctly classified into categories. If the result is positive, then take 1; otherwise, take 0.

[0033] Further, in step S4, for each pixel, the geographic weighted user precision of each land cover product at that location is extracted, and combined with its classification results, the fusion probability of different categories is calculated in a weighted manner:

[0034] ;

[0035] In the formula, For cell i to belong to category The probability is given by C, where C is the total number of categories and M is the total number of land cover products. For land cover product m, the category at pixel i The corresponding geographic weighted user precision, This represents the classification result of product m at pixel i. For the characteristic function, when equals category The value is 1 if it is true, and 0 otherwise.

[0036] Furthermore, in step S5, for pixel i, the fused category for:

[0037] ;

[0038] The category with the highest fusion probability is selected as the candidate classification result for pixel i.

[0039] Furthermore, in step S5, when multiple categories of pixel i have the same maximum fusion probability, geographic weighted producer precision (GWPA) is introduced for arbitration. Among all candidate categories, the category with the largest average geographic weighted producer precision at pixel i is selected. As the final fusion category for pixel i:

[0040] ;

[0041] In the formula, For pixel i, the land cover product m pairs of categories Geographically weighted producer precision, Let i represent the set of candidate categories that have the same maximum fusion probability at pixel i. This represents the classification result of product m at pixel i. To classify the pixel at pixel i as a category The number of products covered by land.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] First, the remote sensing land cover data fusion method based on category precision and geographic weighting of the present invention introduces geographic weighted user precision as category weight at the pixel scale to achieve spatial adaptive adjustment of fusion weight. This effectively characterizes the spatial non-stationarity of land cover classification precision and fully adapts to the characteristics of strong spatial heterogeneity of land cover data and uneven distribution of category conflicts among products. This allows different land cover products to make higher contributions in their advantageous categories and advantageous areas, avoiding the irrationality of traditional global equal-weighted voting methods that ignore regional and land category precision differences, and improving the accuracy, spatial consistency and regional adaptability of the fusion results.

[0044] Second, the remote sensing land cover data fusion method based on category precision and geographic weighting of the present invention introduces geographic weighted producer precision as a discrimination criterion to construct a conflict arbitration mechanism to discriminate candidate categories in the case where multiple categories may have equal fusion probabilities during the fusion process, thereby improving the reliability of land cover classification results.

[0045] Third, the remote sensing land cover data fusion method based on category precision geographic weighting of the present invention can effectively integrate the advantageous information of multi-source land cover product information, effectively reduce the classification error of single products, and generate high-precision and high-stability land cover data. It can provide reliable data support for applications such as ecological environment monitoring and land resource management, and has important application and promotion value. Attached Figure Description

[0046] Figure 1 This is a flowchart of the remote sensing land cover data fusion method based on category precision geographic weighting of the present invention;

[0047] Figure 2The diagram shows the spatial distribution of four land cover products in the study area, where (a) corresponds to FROMC-GLC10; (b) corresponds to ESRI-LC10; (c) corresponds to WorldCover10; and (d) corresponds to Dynamic World.

[0048] Figure 3 This is a schematic diagram of the spatial distribution of land cover results obtained after the integration of the study areas. Detailed Implementation

[0049] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0050] This invention discloses a remote sensing land cover data fusion method based on category-accuracy geographic weighting, the method comprising the following steps:

[0051] S1, acquire multi-source remote sensing land cover product data, preprocess and perform consistency transformation on the product data to ensure that the multi-source remote sensing land cover product data maintains consistency in spatial reference and classification system;

[0052] S2, based on the comprehensive area ratio of each category in each land cover product as the stratification basis, constructs a unified real category verification sample set based on stratified random sampling and high-resolution image visual interpretation;

[0053] S3. Based on the spatial distribution density of the validation samples in the study area, the number of local neighborhood samples N corresponding to each pixel is adaptively determined. For each pixel to be fused, the spatial neighborhood radius is dynamically adjusted so that the neighborhood contains exactly N validation samples. The spatial neighborhood radius corresponding to different pixels adapts to the sample distribution and the degree of class conflict. N is a natural number. For each land cover product, spatial weights are constructed based on the spatial distance between the sample and the pixel. The closer the distance, the greater the weight. The spatial weights are used to perform weighted calculations on the validation samples in the neighborhood to obtain the geographic weighted user accuracy (GWUA) and geographic weighted producer accuracy (GWPA) of each product at the pixel scale.

[0054] S4. Calculate the fusion probability of each category under the pixel based on the geographic weighted user precision of each product at the pixel to be fused, combined with the product classification results.

[0055] S5. Select the category with the highest fusion probability as the candidate category of the pixel. If multiple categories have the same maximum fusion probability, for each candidate category, calculate the average of the geographic weighted producer precision of all land cover products that classify the pixel as that category, and determine the category with the largest average as the final fusion category of the pixel.

[0056] 1) Land cover product data preprocessing

[0057] Multiple remote sensing land cover product data were acquired, and data preprocessing was performed on each land cover product, including projection transformation, resampling, and cropping masking. The classification systems of different data were also converted to ensure consistency in spatial and classification systems.

[0058] Specifically, firstly, a unified projection transformation is performed on different land cover products, converting them to the same projection and coordinate system, ensuring that all products have the same spatial reference system. Then, for products with different spatial resolutions, the highest resolution is used as the benchmark, and the maximum area weighting method is employed for resampling. Finally, the land cover products are spatially clipped according to the study area, and areas outside the study area are masked. Through this process, different land cover products have a one-to-one correspondence at the same spatial location, meeting the requirements of pixel-by-pixel fusion computation.

[0059] Since different land cover products may have different classification systems, it is necessary to unify and transform the original classification systems of each product. First, a unified classification system is determined, and then a mapping relationship between the unified system and the various land cover classification systems is constructed. This process reclassifies the categories of each land cover product to the unified classification system, thereby ensuring the consistency of the classification categories of each product.

[0060] 2) Construct a validation sample set

[0061] Considering that the official accuracy verification reports for different land cover products use different sources of verification samples, resulting in a lack of comparability in the classification accuracy between different data, a unified verification dataset is constructed to evaluate the accuracy of each land cover product and is used for subsequent fusion based on category accuracy and geographic weighting.

[0062] Based on the comprehensive area proportion of each land cover product category as the stratification criterion, a stratified random sampling method was used to generate candidate sample points within the study area to ensure the representativeness and spatial balance of the samples across different land cover types. High-resolution satellite imagery was used to manually interpret the location of each sample point. During the interpretation process, historical imagery, multi-temporal information, and texture structure were comprehensively utilized to determine the true land cover type. Sample points with blurred images, mixed land cover types, or obvious ambiguities were removed. Multiple rounds of manual verification were conducted for some key categories or transitional areas to generate the final validation sample dataset.

[0063] 3) Calculate geographic-weighted user precision and geographic-weighted producer precision.

[0064] Existing fusion methods employ fixed neighborhood or simple adaptive neighborhood strategies, which fail to consider the differences in the complexity of land features and the degree of conflict among multiple product categories within local areas. This makes it difficult to guarantee the reliability of local accuracy estimation under different spatial patterns, thus affecting the rationality and stability of the weighted fusion results. To address the shortcomings of traditional accuracy evaluation methods that ignore spatial heterogeneity, this invention introduces the concept of geographic weighting, calculating local accuracy indices at the pixel scale, including geographic weighted user accuracy (GWUA) and geographic weighted producer accuracy (GWPA).

[0065] Based on the verification sample points selected by stratified random sampling in step 2), this invention provides two methods for determining the number of neighborhood samples N and the spatial neighborhood radius (bandwidth) to adapt to the land cover characteristics of different study areas:

[0066] The first method is the fixed sample number method: a fixed number of neighborhood samples N can be set according to the total sample size, spatial distribution density or domain experience of the study area. Then, the spatial neighborhood radius is dynamically adjusted with each pixel to be fused as the center, so that the neighborhood includes exactly N validation samples, thereby ensuring the statistical stability of local accuracy estimation. This method is suitable for study areas with relatively uniform distribution of ground features and low degree of product category conflict.

[0067] However, for study areas with significant differences in the complexity of ground features and the degree of conflict among multiple product categories, the fixed sample number scheme is difficult to adapt to local spatial heterogeneity, easily leading to local accuracy estimation bias. To address this problem, this invention proposes a second adaptive bandwidth acquisition method, which adaptively determines the number of local neighborhood samples N corresponding to each pixel, and the spatial neighborhood radius corresponding to different pixels adaptively changes with sample distribution and the degree of category conflict; specifically including the following steps:

[0068] Based on the actual spatial distribution density of the validation samples in the study area, the number N of local neighborhood samples corresponding to each pixel to be fused is adaptively determined, and the spatial neighborhood radius corresponding to different pixels adapts adaptively with the sample distribution and the degree of class conflict; specifically, the following steps are included:

[0069] An initial statistical window with a fixed radius is constructed centered on each pixel to be fused. The number of validation samples within the statistical window is calculated, and the local sample density is determined. The study area is divided into several density levels according to the local sample density. Each density level has a preset number of basic neighborhood samples. The higher the local sample density, the smaller the corresponding number of basic neighborhood samples, to adapt to the spatial differences in sample distribution. Considering the spatial heterogeneity of land cover complexity and conflicts between multi-source land cover product categories within the local area, after determining the number of basic neighborhood samples, a corresponding number of basic neighborhood samples is assigned according to the density zone to which the pixel belongs. Then, the number of basic neighborhood samples is adaptively corrected according to the degree of inconsistency between multi-source land cover products within the local window. The higher the degree of inconsistency between products, the greater the correction magnitude of the number of basic neighborhood samples. This yields the final adaptive number of local neighborhood samples N for each pixel. Finally, the spatial neighborhood radius is dynamically adjusted centered on each pixel to be fused until the neighborhood contains exactly the final determined N validation samples, ensuring the stability and rationality of the local accuracy estimation, adapting to the land cover distribution characteristics of different areas, and thus guaranteeing the statistical stability of the local accuracy estimation.

[0070] Based on this, the validation samples within the spatial neighborhood and the land cover product classification results of the corresponding pixels are extracted, and the geographic weighted user precision (GWUA) and geographic weighted producer precision (GWPA) are calculated.

[0071] The formula for calculating geographic weighted user accuracy (GWUA) is as follows:

[0072] ;

[0073] in, Represents the Cth pixel at position i. k Geographically weighted user precision for land cover types Let J be the spatial weight of sample j relative to pixel i within the neighborhood window, and let J be the total number of samples identified as type C within the neighborhood window. k The number of validation samples, correct j A binary indicator function (if the j-th neighbor is identified as category C) k The sample has a true class of class C. k If the value is 1, then take 1; otherwise, take 0.

[0074] Spatial weight Based on the distance between samples and target pixels, this method ensures that samples closer to the target pixel contribute more to the accuracy estimate, thus effectively reflecting the spatial non-stationary characteristics of classification accuracy. The calculation formula is as follows:

[0075] ;

[0076] In the formula, d i,j Let b be the spatial distance between sample j and current pixel i. i Let be the bandwidth at pixel i, which is the distance within the neighborhood that contains exactly a predetermined number of N samples.

[0077] The formula for calculating Geographically Weighted Producer Precision (GWPA) is as follows:

[0078] ;

[0079] in, Represents the Cth pixel at position i. k Geographically weighted producer precision for land cover types J' represents the spatial weight of sample j within the neighborhood window, and J' represents the spatial weight of type C within the neighborhood window. k The true number of samples, correct j A binary indicator function (if the j-th true class in the neighborhood is class C). k The sample was correctly classified as category C. k If the value is 1, then take 1; otherwise, take 0.

[0080] 4) Calculation of fusion probability based on geographic weighted user precision

[0081] After obtaining the geographic weighted classification accuracy information of each land cover product, the category fusion probability is calculated for each pixel to be fused based on the calculated geographic weighted user accuracy.

[0082] For each pixel, the geographic-weighted user precision of each land cover product at that location is extracted, and combined with its classification results, the fusion probability of different categories is calculated using a weighted method:

[0083] (4);

[0084] In the formula, P(C k ) ,j For pixel i to belong to category C k The probability, C is the total number of categories, M is the total number of land cover products, GWUA j (C k For land cover product m, category C is defined at pixel i. k The corresponding geographic-weighted user precision, x i,m For the classification result of product m at pixel i, I(x) ij =C k ) is an indicator function, when X i,m Equal to category C k The value is 1 if it is true, and 0 otherwise.

[0085] 5) Candidate category determination based on fusion probability

[0086] After obtaining the fusion probability of each category, the candidate category of each pixel is determined by comparing the fusion probability values ​​of different categories.

[0087] For pixel i, the fused category for:

[0088] (5);

[0089] The category with the highest fusion probability is selected as the candidate classification result for pixel i.

[0090] 6) Conflict arbitration based on geographically weighted producer accuracy

[0091] When multiple categories have the same maximum fusion probability at pixel i, geographic weighted producer precision (GWPA) is introduced for arbitration. Among all candidate categories, the category with the largest average geographic weighted producer precision at pixel i is selected as the final fusion category for pixel i.

[0092] (6);

[0093] In the formula, GWPA i,m (C k For pixel i, the land cover product m corresponds to category C. k Geographically weighted producer precision, S i Let x represent the set of candidate classes with the same maximum fusion probability at pixel i. i,m M represents the classification result of product m at pixel i. k To classify the pixel at location i as category C k The number of land cover products. By analyzing all predictions for category C... k The geographically weighted producer precision of the products is averaged and used as the basis for classifying the product category.

[0094] Example

[0095] This example demonstrates the fusion of remote sensing land cover data for a specific study area using category-accuracy geographic weighting, and provides detailed implementation steps. See the technical flowchart below. Figure 1 .

[0096] 1) First, acquire land cover data from four 10 m resolution remote sensing products: FROM-GLC10, WorldCover10, ESRI-LC10, and DynamicWorld, within the study area. Convert all four land cover products to the same spatial reference system and spatially clip the data. Then, spatially match the four land cover products to ensure a one-to-one correspondence at the same spatial location. Next, uniformly transform the original classification systems of the four products, mapping the land cover categories in each product to a unified target classification system (Table 1), thereby obtaining multi-source land cover data with a unified spatial reference and a unified classification system. Figure 2 ).

[0097] Table 1. Correspondence between the target classification system and the four categories of 10 m land cover product classification systems.

[0098] Target Category FROM-GLC10 ESRI-LC10 WorldCover10 Dynamic World arable land arable land crop arable land crop woodland woodland Trees Tree cover Trees grassland grassland Grass grassland Grass shrub shrub Shrubs and thickets shrub Shrubs and thickets wetlands wetlands Submerged vegetation herbaceous wetlands and mangroves Submerged vegetation water body water body water body permanent water bodies water body impermeable surface impermeable surface Built-up area building Built-up area bare land bare land bare land bare / sparse vegetation bare land Ice and snow Ice and snow Ice and snow Ice and snow Ice and snow

[0099] 2) A stratified random sampling method was used to generate candidate sample points within the study area. Stratified random sampling was performed based on the comprehensive area proportion of each of the four land cover products to ensure the representativeness and spatial distribution balance of the samples across different categories. Subsequently, each sample point was visually interpreted using high-resolution Google Earth imagery. The true land cover type of the sample point was determined by combining historical imagery, multi-temporal information, and land cover texture features. Sample points with blurred images, mixed land cover features, or obvious ambiguities were removed to obtain the final validation sample data.

[0100] The validation samples were matched with the classification results of each land cover product to construct confusion matrices for the four land cover products. The overall accuracy and Kappa coefficient were calculated to evaluate the overall classification performance of each product. Significant differences were observed among the four land cover products within the study area. Dynamic World had the highest overall accuracy (75.15%) and Kappa coefficient (0.69); FROM-GLC10 had an overall accuracy of 65.56% and a Kappa coefficient of 0.57; WorldCover10 had an overall accuracy of 65.90% and a Kappa coefficient of 0.58; and ESRI-LC10 had the lowest overall accuracy (54.17%) and Kappa coefficient of 0.46.

[0101] 3) Based on the validation samples, calculate the geographic weighted user precision and geographic weighted producer precision for each land cover product. For any pixel in the study area, an adaptive bandwidth strategy is used to construct its spatial neighborhood. By dynamically adjusting the neighborhood radius, the neighborhood contains 500 validation samples, thereby ensuring the stability of the local precision estimation. Within the neighborhood of each pixel, the validation samples and the classification results of the four land cover products are extracted respectively. Based on formulas (1) and (3), the geographic weighted user precision GWUA and geographic weighted producer precision GWPA are calculated respectively.

[0102] 4) Perform fusion calculations based on the above accuracy information. For any pixel in the study area, extract the classification results of the four land cover products at that location, and read the geographic weighted user accuracy of each product in the corresponding category. Use formula (4) to calculate the fusion probability value of each candidate category at that pixel. Then compare the fusion probability values ​​of each candidate category and select the category with the highest probability as the candidate category of that pixel.

[0103] 5) When two or more categories have the same maximum fusion probability, geographic weighted producer accuracy is introduced for conflict arbitration. The geographic weighted producer accuracy of the corresponding category on each land cover product is read, and the average geographic weighted producer accuracy of products predicted to be of that category is calculated. The category with the largest average geographic weighted producer accuracy is selected as the final land cover category from the candidate categories (Formula 6).

[0104] 6) Using this method, all pixels within the study area are calculated one by one to generate the final land cover fusion result. Figure 3 The accuracy of the fusion results was evaluated based on validation samples. The results show that the overall accuracy of the fused land cover data reached 76.10%, and the Kappa coefficient reached 0.69. Compared to the individual land cover products before fusion, the method of this invention has improved both overall accuracy and consistency. Specifically, compared to the highest-accuracy Dynamic World product, the overall accuracy improved by approximately 0.95 percentage points, while the Kappa coefficient remained at a high level; compared to FROM-GLC10 and WorldCover10 products, the overall accuracy improved by approximately 10 percentage points respectively; and compared to the ESRI-LC10 product, the overall accuracy improved by more than 20 percentage points.

[0105] The remote sensing land cover data fusion method proposed in this invention, based on category precision and geographic weighting, introduces geographic weighted user precision at the pixel scale for category weighting and combines it with geographic weighted producer precision for conflict rejection, thereby achieving spatial adaptive fusion of multi-source land cover products and effectively improving the accuracy and stability of land cover classification results.

[0106] The preferred embodiments are described, but those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0107] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A remote sensing land cover data fusion method based on category-accuracy geographic weighting, characterized in that, The method includes the following steps: S1, acquire multi-source remote sensing land cover product data, preprocess and perform consistency transformation on the product data to ensure that the multi-source remote sensing land cover product data maintains consistency in spatial reference and classification system; S2, based on the comprehensive area ratio of each category in each land cover product as the stratification basis, constructs a unified real category verification sample set based on stratified random sampling and high-resolution image visual interpretation; S3. Based on the spatial distribution density of the validation samples in the study area, the number of local neighborhood samples N corresponding to each pixel is adaptively determined. For each pixel to be fused, the spatial neighborhood radius is dynamically adjusted so that the neighborhood contains exactly N validation samples. The spatial neighborhood radius corresponding to different pixels adapts to the sample distribution and the degree of class conflict. N is a natural number. For each land cover product, spatial weights are constructed based on the spatial distance between the sample and the pixel. The closer the distance, the greater the weight. The spatial weights are used to perform weighted calculations on the validation samples in the neighborhood to obtain the geographic weighted user accuracy (GWUA) and geographic weighted producer accuracy (GWPA) of each product at the pixel scale. S4. Calculate the fusion probability of each category under the pixel based on the geographic weighted user precision of each product at the pixel to be fused, combined with the product classification results. S5. Select the category with the highest fusion probability as the candidate category of the pixel. If multiple categories have the same maximum fusion probability, for each candidate category, calculate the average of the geographic weighted producer precision of all land cover products that classify the pixel as that category, and determine the category with the largest average as the final fusion category of the pixel.

2. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, Step S1 further includes: Acquire data from multiple remote sensing land cover products, perform unified projection conversion on different land cover products, and make each product have the same spatial reference system; For land cover products with different spatial resolutions, the highest resolution is used as the benchmark, and the maximum area weighting method is used to resample each land cover product. Based on the scope of the study area, each land cover product is spatially cut, and the area outside the study area is masked to ensure that different land cover products have a one-to-one correspondence in the same spatial location. A unified classification system is established and a classification mapping relationship is constructed, so that the categories of various land cover products are uniformly classified into this unified classification system.

3. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, Step S2 further includes: Based on the comprehensive area ratio of various land cover types in the study area, stratified random sampling was used to generate candidate sample points in the study area, so that the samples are representative and spatially balanced among different land types. High-resolution satellite imagery is used to visually interpret the locations of each candidate sample point. Historical imagery, multi-temporal information, and ground texture are combined to determine the true ground cover type. Sample points with blurred images, mixed ground covers, or ambiguity are removed. Multiple rounds of verification and interpretation are performed on samples of key ground cover types or transitional areas to generate the final true category verification sample dataset.

4. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, In step S3, the process of adaptively determining the number of local neighborhood samples N corresponding to each pixel based on the spatial distribution density of the validation samples in the study area includes the following steps: An initial statistical window with a fixed radius is constructed centered on each pixel to be fused. The number of validation samples within the statistical window is calculated, and the local sample density is determined. The study area is divided into several density levels according to local sample density. Each density level has a preset number of basic neighborhood samples. The higher the local sample density, the smaller the corresponding number of basic neighborhood samples. The number of basic neighborhood samples is assigned according to the density zone to which the pixel belongs. Then, the number of basic neighborhood samples is adaptively adjusted according to the degree of inconsistency between the categories of multi-source land cover products within the local window. The higher the degree of inconsistency between products, the greater the adjustment of the number of basic neighborhood samples. Thus, the final adaptive number of local neighborhood samples N for each pixel is obtained.

5. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, In step S3, the formula for calculating the geographic weighted user precision (GWUA) is as follows: ； in, Indicates the position of pixel i. Geographically weighted user precision for land cover types Let J be the spatial weight of sample j relative to pixel i within the neighborhood window, and J be the total number of samples identified as type i within the neighborhood window. The number of validation samples; Let be a binary indicator function, if the j-th element in the neighborhood is identified as category... The sample, whose true category is category If the value is 1, then the value is 1; otherwise, the value is 0. Spatial weight The calculation formula is: ； In the formula, The spatial distance between sample j and current pixel i; Let be the bandwidth at pixel i, defined as the distance within the neighborhood that contains exactly a predetermined number of N samples.

6. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, In step S3, the formula for calculating the geographic weighted producer precision (GWPA) is as follows: ； in, Indicates the position of pixel i. Geographically weighted producer precision for land cover types The spatial weights of sample j within the neighborhood window. Type within the neighborhood window The actual number of samples; Let be a binary indicator function, if the j-th true class in the neighborhood is a class The samples were correctly classified into categories. If the result is positive, then take 1; otherwise, take 0.

7. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, In step S4, for each pixel, the geographic weighted user precision of each land cover product at that location is extracted, and combined with its classification results, the fusion probability of different categories is calculated using a weighted method: ； In the formula, For cell i to belong to category The probability is given by C, where C is the total number of categories and M is the total number of land cover products. For land cover product m, the category at pixel i The corresponding geographic weighted user precision, This represents the classification result of product m at pixel i. For the characteristic function, when equals category The value is 1 if it is true, and 0 otherwise.

8. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, In step S5, for pixel i, the fused category for: ； The category with the highest fusion probability is selected as the candidate classification result for pixel i.

9. The remote sensing land cover data fusion method based on category-accuracy geographic weighting according to claim 1, characterized in that, In step S5, when multiple categories of pixel i have the same maximum fusion probability, geographic weighted producer precision (GWPA) is introduced for arbitration. Among all candidate categories, the category with the largest average geographic weighted producer precision at pixel i is selected. As the final fusion category for pixel i: ； In the formula, For pixel i, the land cover product m pairs of categories Geographically weighted producer precision, Let i represent the set of candidate categories that have the same maximum fusion probability at pixel i. This represents the classification result of product m at pixel i. To classify the pixel at pixel i as a category The number of products covered by land.

Images