Land coverage mapping method integrating data and knowledge dual drive

Abstract

The invention discloses a data and knowledge fused dual-drive land coverage mapping method, and belongs to the technical field of machine learning and land coverage mapping. The method comprises the following steps: firstly, constructing a spectral statistical feature matrix, and generating a data-driven classification result in combination with a high-confidence training sample; secondly, obtaining a year-round multi-temporal terrestrial spectrum semantic type division result, performing cumulative frequency calculation, and combining the two to generate a geographic entity classification result with semantic interpretation capability; and finally, carrying out collaborative fusion on a data-driven classification result and a geographic entity classification result, and generating an optimized annual land coverage classification map and an optimized near-real-time land coverage map in combination with a current land coverage map data product. According to the method, the advantages of data driving and knowledge driving are fully fused, and high precision, stability and semantic consistency of annual and near-real-time land coverage mapping are realized; the space continuity, the space-time generalization ability and the application reliability of a classification result can be improved, and the method has high practicability and expandability.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and land cover mapping technology, and relates to a land cover mapping method that integrates data and knowledge, specifically a method for constructing annual and dynamic land cover mapping frameworks using remote sensing data, machine learning methods and geoscientific ontology knowledge. Background Technology

[0002] With the acceleration of globalization and the continuous increase in the intensity of human activities, significant changes have occurred in land cover patterns. Processes such as deforestation, urban expansion, and agricultural intensification have profoundly impacted the structure and function of ecosystems, making land cover change one of the core issues in global environmental change research. The United Nations Sustainable Development Goals (SDGs) explicitly state that the monitoring and assessment of land change processes should be strengthened to enhance scientific decision-making in key areas such as ecological protection, resource management, and sustainable land use. Large-scale, continuous, and highly interpretable land cover mapping and change monitoring data can be applied to natural resource survey updates, urban planning and design, land policy formulation, and effect evaluation, and are of significant practical importance for achieving harmony between ecological environmental protection and human-land relations.

[0003] Currently, remote sensing technology has become a core tool for land cover monitoring. With the continuous release of multi-source high spatiotemporal resolution remote sensing images such as Sentinel-2 and Landsat, data-driven land cover classification methods have developed rapidly. These methods mainly rely on large-scale manual samples and extract remote sensing features using machine learning or deep learning models. They have advantages such as high automation and good accuracy, and have been widely used in the rapid generation of annual and even monthly land cover maps. For example, Chinese invention patent (202510792058.2) provides a land cover classification method based on time-series data. It calculates the normalized difference vegetation index and tasseled cap change humidity component based on multispectral image data, and uses a time-series difference evaluation index to select the best time-series feature set. Combined with a support vector machine classifier, it generates the land cover classification results for the study area.

[0004] However, data-driven methods typically lack semantic interpretability and are easily affected by factors such as sample distribution and geographical differences, resulting in insufficient generalization ability of classification results in complex areas. The opacity of the model also limits user intervention and adjustment. In contrast, knowledge-driven methods emphasize the introduction of geoscientific knowledge and classification rules, achieving interpretable classification of remote sensing images through spectral rules, ontology systems, or semantic logical reasoning. This method has significant advantages in controllability, interpretability, and semantic consistency, and can effectively assist in the semantic optimization and boundary correction of land cover results. For example, Chinese invention patent (application number 202411854580.0) provides a method for constructing a remote sensing intelligent interpretation sample dataset coupled with geoscientific knowledge. It uses the analytic hierarchy process (AHP) to construct a partitioning index system for interpretation samples, generating a remote sensing intelligent interpretation sample dataset coupled with geoscientific knowledge based on remote sensing image data and sample elements within the sample collection area. However, traditional knowledge-driven methods generally rely on manually constructed rules, making it difficult to efficiently adapt to multi-temporal remote sensing data and lacking the ability to quickly respond to dynamic changes in large areas.

[0005] Therefore, how to collaboratively integrate multi-temporal features of remote sensing, data-driven models, and ontological knowledge rules to construct a land cover classification and change monitoring system with a high degree of automation, semantic consistency, and near real-time response capability remains a key issue that urgently needs to be addressed in the field of remote sensing intelligent interpretation and land resource monitoring. Summary of the Invention

[0006] To address the problems existing in current technologies, this invention provides a land cover mapping method that integrates data and knowledge-driven approaches. This invention utilizes multi-temporal remote sensing imagery, machine learning models, and ontology semantic rules to achieve annual and near real-time high-precision mapping and change identification of land cover, improving the semantic consistency and interpretability of classification results. This meets the dual requirements of real-time performance and accuracy for remote sensing classification results in ecological environment monitoring and resource management.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A land cover mapping method that integrates data and knowledge-driven approaches includes the following steps:

[0009] Step S1: Data-driven classification. A spectral statistical feature matrix is ​​constructed based on multi-temporal Sentinel-2 surface reflectance image data. A random forest model is used, combined with high-confidence training samples, to generate data-driven classification results: initial annual land cover classification mapping and initial near-real-time land cover mapping; specifically:

[0010] Step S11: Obtain multi-temporal Sentinel-2 surface reflectance image data of the target area throughout the year, preprocess it to obtain target images that can be used for feature construction, and obtain preprocessed multi-temporal Sentinel-2 surface reflectance image data of the target area throughout the year; the preprocessing includes cloud detection, image screening and radiometric correction.

[0011] Step S12: Extract multispectral band information, including blue light, green light, red light, red edge band, near-infrared band, and shortwave infrared band, from the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11, and calculate the derived spectral index; calculate the statistical characteristic values ​​of the time series data of the above spectral bands and derived spectral indices, including the decimal places, 20th percentile places, ..., 90th percentile places, to construct a multi-temporal characteristic matrix characterizing the seasonal changes of land cover;

[0012] Step S13: Introduce a high-confidence land type sample set to obtain an initial sample set with high label reliability; compare the similarity and difference of each feature in the multi-temporal feature matrix of the mapping year and the initial sample set collection year, and filter samples with high spectral similarity and low difference through similarity threshold and difference threshold for sample transfer to construct a training dataset of the mapping year.

[0013] Step S14: Based on the training dataset of the mapping year constructed in step S13, a machine learning classifier is used to train the multi-temporal feature matrix of the mapping year obtained in step S12 to obtain a land cover classification model suitable for annual mapping; the land cover classification model is used to perform full-pixel classification of the multi-temporal feature matrix of the mapping year to output the initial annual land cover classification map.

[0014] Step S15: Based on the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11, obtain the latest available single-scene image to construct a simplified feature set, combine it with the training dataset of the mapping year constructed in Step S13, train a near-real-time land cover mapping classification model, and generate the initial near-real-time land cover map closest to the current time.

[0015] Furthermore, in step S12, the derived spectral indices include the normalized vegetation index, the normalized combustion index, and the normalized water index.

[0016] Furthermore, in step S13, the high-confidence land type sample set is obtained based on remote sensing images with higher spatial resolution;

[0017] Furthermore, in step S13, the similarity threshold is greater than 0.9, and the difference threshold is less than 0.1.

[0018] Furthermore, in step S15, the simplified feature set includes multispectral band information of a single scene image, including blue light, green light, red light, red edge band, near-infrared band, and short-wave infrared band, as well as derived spectral indices such as normalized vegetation index, normalized combustion index, and normalized water index calculated using the multispectral bands of the single scene image.

[0019] Step S2: Based on the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11, the annual multi-temporal land cover spectral semantic type classification results are obtained using the Spectral Image Automatic Mapper (SIAM), and cumulative frequencies are calculated for different land cover spectral semantic types. The land cover spectral semantic types of the mapping year and the cumulative frequencies of each type are combined to generate geographic entity classification results with semantic interpretation capabilities. Specifically:

[0020] Step S21: Use the Spectral Image Auto Mapper (SIAM) to process each image in the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11. Based on the spectral reflectance ratio relationship between each spectral band and the preset threshold conditions, SIAM divides the pixels into six types of land cover spectral semantic types: strong green vegetation, weak green vegetation, blue water body, brown surface, white cloud, and shadow, to obtain the annual multi-temporal land cover spectral semantic type classification results.

[0021] Step S22: Based on the annual multi-temporal land cover spectral semantic type classification results obtained in step S21, count the number of occurrences of each land cover spectral semantic type for each pixel, and use the cloud content of the pixel as the weight to obtain the normalized annual cumulative frequency value. The annual cumulative frequency value ranges from 0 to 1, and the annual cumulative frequency value is divided into eight types, including: never, very few, occasionally, regularly, most, usually, often, and always, to obtain the annual land cover spectral semantic type cumulative frequency.

[0022] Step S23: Combining the annual multi-temporal land cover spectral semantic type classification results obtained in Step S21, and the annual land cover spectral semantic type frequency feature data obtained in Step S22, the feature frequencies of different land cover spectral semantic types for each pixel throughout the year are obtained. Based on the classification rules, each pixel is divided into six types, including terrestrial vegetation, terrestrial non-vegetation, aquatic vegetation, water body, cloud, and shadow, as geographic entity types.

[0023] Furthermore, in step S22, the criteria for classifying the cumulative frequency of the annual land use spectral semantic type are as follows: if the annual cumulative frequency value is equal to 0, it is classified as never; if the annual cumulative frequency value is greater than 0 and less than 0.1, it is classified as very rare; if the annual cumulative frequency value is greater than or equal to 0.1 and less than 0.25, it is classified as occasional; if the annual cumulative frequency value is greater than or equal to 0.25 and less than 0.5, it is classified as regular; if the annual cumulative frequency value is greater than or equal to 0.5 and less than 0.75, it is classified as most; if the annual cumulative frequency value is greater than or equal to 0.75 and less than 0.9, it is classified as usually; if the annual cumulative frequency value is greater than or equal to 0.9 and less than 1, it is classified as frequently; if the annual cumulative frequency value is equal to 1, it is classified as always.

[0024] Furthermore, in step S22, the division rules are as follows: First, the spectral semantic type of land cover and its cumulative frequency throughout the year are combined to obtain the feature frequency of different spectral semantic types of land cover for each pixel throughout the year. For example, if the spectral semantic type of land cover is "strong green vegetation" and the cumulative frequency of the spectral semantic type of land cover throughout the year is "never", the combination is "never strong green vegetation". Second, based on the combination of the feature frequencies of different spectral semantic types of land cover in the pixel throughout the year, the geographic entity type is obtained. For example, "never strong green vegetation + often brown surface" is inferred to be "terrestrial non-vegetation".

[0025] Step S3: Dual-drive collaborative fusion. The data-driven classification results obtained in Step S1 are collaboratively fused with the geographic entity classification results obtained in Step S2, and combined with existing land cover mapping data products to generate optimized annual land cover classification maps and optimized near-real-time land cover maps. Specifically:

[0026] Step S31: Based on the data-driven classification results output in step S1, the geographic entity classification results obtained in step S2, and the existing land cover mapping data products, construct an assertion rule base with each pixel as a geographic entity.

[0027] Step S32: Based on the assertion rule base constructed in step S31, perform pixel-level category correction on the data-driven classification results obtained in step S1, namely the initial annual land cover classification map and the initial near real-time land cover map, to improve semantic consistency and classification accuracy, and obtain the optimized annual land cover classification map and the optimized near real-time land cover map.

[0028] Furthermore, in step S31, the assertion rule base is set for annual land cover classification mapping and near real-time land cover mapping respectively. The land cover types involved are a total of ten categories, including cultivated land, forest, grassland, shrubland, wetland, water body, tundra, impermeable layer, bare land, and ice and snow. The basic form of the assertion rule base for annual land cover classification mapping is "If the initial annual land cover type of a 'geographic entity' is 'land cover type 1', then it should be corrected to 'land cover type 2' - provided that the 'geographic entity' is classified as 'land cover type 2' in existing land cover mapping data products, and the geographic entity classification result is 'geographic entity type 1'". The basic form of the assertion rule base for near-real-time land cover classification mapping is "If the initial near-real-time land cover type of a 'geographic entity' is 'land cover type 1', then it should be corrected to 'land cover type 2' - provided that the 'geographic entity' is classified as 'land cover type 2' in the optimized previous year's land cover classification mapping, and the land cover spectral semantic type classification result is 'land cover spectral semantic type 1'".

[0029] The beneficial effects of this invention are as follows:

[0030] (1) This invention integrates the data-driven classification results constructed by multi-temporal spectral features with the knowledge-driven information of geographic entity classification results, and uses assertion rule base to perform pixel-level correction of the data-driven classification results, which helps to improve the accuracy and semantic consistency of annual land cover classification.

[0031] (2) Based on the optimized annual land cover classification map of the previous year as a semantic prior, this invention can correct the initial classification results generated by near real-time remote sensing images, thereby realizing stable updates and near real-time mapping capabilities for dynamic land cover change mapping.

[0032] (3) By classifying geographic entities, the present invention can effectively reduce isolated patches and boundary noise and enhance spatial continuity; at the same time, it can be combined with the sample migration mechanism to further improve the spatiotemporal generalization ability of the present invention, making the present invention more practical and scalable.

[0033] In summary, this invention fully integrates the advantages of data-driven and knowledge-driven approaches, achieving high accuracy, stability, and semantic consistency in annual and near-real-time land cover mapping. By using geographic entity classification results as semantic priors and combining them with a sample transfer mechanism, this invention significantly improves the spatial continuity, spatiotemporal generalization ability, and application reliability of the classification results. The method of this invention has strong practicality and scalability, and can be widely applied to land cover monitoring tasks in multiple regions and time phases. Attached Figure Description

[0034] Figure 1This is a schematic diagram of the construction method for an annual and dynamic land cover mapping framework driven by both data and knowledge according to an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and beneficial effects of this invention clearer, specific embodiments are provided for further detailed explanation. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0036] According to one embodiment of the present invention, the present invention proposes a land cover mapping framework driven by both data and knowledge, such as... Figure 1 As shown, the process includes steps S1, S2, and S3. Each step will be described in detail below with reference to the accompanying drawings.

[0037] Step S1: Data-driven classification. A spectral statistical feature matrix is ​​constructed based on multi-temporal Sentinel-2 surface reflectance image data. A random forest model is used, combined with high-confidence training samples, to generate data-driven classification results: initial annual land cover classification mapping and initial near-real-time land cover mapping; specifically:

[0038] Step S11: Obtain multi-temporal Sentinel-2 surface reflectance image data of the target area throughout the year, preprocess it to obtain target images that can be used for feature construction, and obtain preprocessed multi-temporal Sentinel-2 surface reflectance image data of the target area throughout the year; the preprocessing includes cloud detection, image screening and radiometric correction.

[0039] Step S12: Extract multispectral band information, including blue light, green light, red light, red edge band, near-infrared band, and shortwave infrared band, from the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11, and calculate derived spectral indices such as normalized vegetation index, normalized combustion index, and normalized water index; calculate the statistical characteristic values ​​of the time series data of the above spectral bands and derived indices, including the decimals, 20ths, ..., 90ths, to construct a multi-temporal feature matrix characterizing the seasonal changes of land cover.

[0040] Step S13: Introduce a high-confidence land cover sample set. The high-confidence land cover sample set is manually interpreted based on remote sensing images with higher spatial resolution to obtain an initial sample set with high label reliability. Compare the similarity and difference of each feature in the multi-temporal feature matrix of the mapping year and the initial sample set collection year. Select samples with high spectral similarity and low difference through a specific threshold for sample transfer to construct a training dataset for the mapping year.

[0041] In the implementation case, the spectral similarity threshold was greater than 0.9, and the difference threshold was less than 0.1.

[0042] Step S14: Based on the training dataset of the mapping year constructed in step S13, a machine learning classifier is used to train the multi-temporal feature matrix of the mapping year obtained in step S12 to obtain a land cover classification model suitable for annual mapping; the land cover classification model is used to perform full-pixel classification of the multi-temporal feature matrix of the mapping year to output the initial annual land cover classification map.

[0043] Step S15: Based on the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11, obtain the latest available single-scene image to construct a simplified feature set, combine it with the training dataset of the mapping year constructed in Step S13, train a near-real-time land cover mapping classification model, and generate the initial near-real-time land cover map closest to the current time.

[0044] In this embodiment, the simplified feature set includes multispectral band information of a single scene image, including blue light, green light, red light, red edge band, near-infrared band, and short-wave infrared band, as well as derived spectral indices such as normalized vegetation index, normalized combustion index, and normalized water index obtained by calculating using the multispectral bands of a single scene image.

[0045] Step S2: Based on the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11, the annual multi-temporal land cover spectral semantic type classification results are obtained using the Spectral Image Automatic Mapper (SIAM), and cumulative frequencies are calculated for different land cover spectral semantic types. The land cover spectral semantic types of the mapping year and the cumulative frequencies of each type are combined to generate geographic entity classification results with semantic interpretation capabilities. Specifically:

[0046] Step S21: Use the Spectral Image Auto Mapper (SIAM) to process each image in the preprocessed annual multi-temporal Sentinel-2 surface reflectance image data obtained in Step S11. Based on the spectral reflectance ratio relationship between each spectral band and the preset threshold conditions, SIAM divides the pixels into six types of land cover spectral semantic types: strong green vegetation, weak green vegetation, blue water body, brown surface, white cloud, and shadow, to obtain the annual multi-temporal land cover spectral semantic type classification results.

[0047] Step S22: Based on the annual multi-temporal land cover spectral semantic type classification results obtained in step S21, count the number of occurrences of each land cover spectral semantic type for each pixel, and use the cloud content of the pixel as the weight to obtain the normalized annual cumulative frequency value. The annual cumulative frequency value ranges from 0 to 1, and the annual cumulative frequency value is divided into eight types, including: never, very few, occasionally, regularly, most, usually, often, and always, to obtain the annual land cover spectral semantic type cumulative frequency.

[0048] In this embodiment, the cumulative frequency classification criteria for land use spectral semantic types throughout the year are as follows: if the annual cumulative frequency value is equal to 0, it is classified as never; if the annual cumulative frequency value is greater than 0 and less than 0.1, it is classified as very rare; if the annual cumulative frequency value is greater than or equal to 0.1 and less than 0.25, it is classified as occasional; if the annual cumulative frequency value is greater than or equal to 0.25 and less than 0.5, it is classified as regular; if the annual cumulative frequency value is greater than or equal to 0.5 and less than 0.75, it is classified as most; if the annual cumulative frequency value is greater than or equal to 0.75 and less than 0.9, it is classified as usually; if the annual cumulative frequency value is greater than or equal to 0.9 and less than 1, it is classified as frequently; if the annual cumulative frequency value is equal to 1, it is classified as always.

[0049] Step S23: Combining the annual multi-temporal land cover spectral semantic type classification results obtained in Step S21 with the annual land cover spectral semantic type frequency feature data obtained in Step S22, the feature frequencies of different land cover spectral semantic types for each pixel throughout the year are obtained, and each pixel is divided into six types, including terrestrial vegetation, terrestrial non-vegetation, aquatic vegetation, water body, cloud, and shadow, as geographic entity types.

[0050] In this embodiment, the division rules are as follows: First, the spectral semantic type of land cover and its cumulative frequency throughout the year are combined to obtain the feature frequency of different spectral semantic types of land cover for each pixel throughout the year. For example, if the spectral semantic type of land cover is "strong green vegetation" and the cumulative frequency of the spectral semantic type of land cover throughout the year is "never", the combination is "never strong green vegetation". Second, based on the combination of the feature frequencies of different spectral semantic types of land cover in the pixel throughout the year, the geographic entity type is obtained. For example, "never strong green vegetation + often brown surface" is inferred to be "terrestrial non-vegetation".

[0051] Step S3: Dual-drive collaborative fusion. The data-driven classification results obtained in Step S1 are collaboratively fused with the geographic entity classification results obtained in Step S2, and combined with existing land cover mapping data products to generate optimized annual land cover classification maps and optimized near-real-time land cover maps. Specifically:

[0052] In implementation cases, existing land cover mapping data products include GlobeLand30 and DynamicWorld;

[0053] Step S31: Based on the data-driven classification results output in step S1, the geographic entity classification results obtained in step S2, and the existing land cover mapping data products, construct an assertion rule base with each pixel as a geographic entity.

[0054] In this embodiment, the assertion rule base is set separately for annual land cover classification mapping and near-real-time land cover mapping, involving ten land cover types, including cultivated land, forest, grassland, shrubs, wetlands, water bodies, tundra, impermeable layer, bare land, and snow and ice. The basic form of the assertion rule base for annual land cover classification mapping is "If the initial annual land cover type of a certain 'geographic entity' is 'land cover type 1', then it should be corrected to 'land cover type 2', provided that the 'geographic entity' is classified as 'land cover type 2' in existing land cover mapping data products, and the geographic entity classification result is 'geographic entity type 1'." The basic form of the assertion rule base for near-real-time land cover classification mapping is "If the initial near-real-time land cover type of a 'geographic entity' is 'Land Cover Type 1', then it should be corrected to 'Land Cover Type 2', provided that the 'geographic entity' was classified as 'Land Cover Type 2' in the optimized land cover classification map of the previous year, and the spectral semantic type classification result of the land cover is 'Spectral Semantic Type 1'". Assertions 1-2 are provided as examples here.

[0055] Assertion 1: If the initial annual land cover type of a pixel is cultivated land, it should be corrected to bare land - provided that the pixel is classified as bare land in both Dynamic World and GlobeLand30, and its geographic entity classification result is "terrestrial non-vegetation".

[0056] Assertion 2: If the initial near-real-time land cover type of a pixel is forest, it should be corrected to water body—provided that the pixel is classified as water body in Dynamic World or GlobeLand30 and its geographic entity classification result is "water body".

[0057] Step S32: Based on the assertion rule base constructed in step S31, perform pixel-level category correction on the data-driven classification results obtained in step S1, namely the initial annual land cover classification map and the initial near-real-time land cover map, to improve semantic consistency and classification accuracy, and obtain the optimized annual land cover classification map and the optimized near-real-time land cover map.

[0058] Based on land cover verification samples obtained through visual interpretation, the overall accuracy of the semantically optimized annual land cover map is significantly improved. Compared to the initial annual land cover classification map, the F1-scores for cultivated land, artificial land surface, and shrubland increased by 0.08, 0.04, and 0.12, respectively, representing an overall accuracy improvement of approximately 5%. Z-score statistical tests show that the optimized annual land cover classification map has a Z-score of 3.69 compared to the initial annual land cover classification map, indicating that the optimization is statistically significant. Validation was performed on 12 Sentinel-2 images from 2023. The optimized dynamic land cover classification map shows an overall accuracy improvement of approximately 20% compared to the initial dynamic land cover classification map, with a Z-score of 12.28, demonstrating the high reliability and application value of the optimization strategy in dynamic mapping.

[0059] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A land cover mapping method driven by fusion of data and knowledge, characterized in that, The land cover mapping method comprises the following steps: Step S1, data-driven classification, constructing a spectral statistical feature matrix based on multi-temporal Sentinel-2 ground reflectivity image data, using a random forest model and combining high-confidence training samples to generate data-driven classification results: initial annual land cover classification mapping and initial near-real-time land cover mapping; Step S2, based on the pre-processed multi-temporal Sentinel-2 ground reflectivity image data obtained in step S11, using a spectral image automatic mapper (SIAM) to obtain multi-temporal ground object spectral semantic type classification results, and calculating the cumulative frequency of different ground object spectral semantic types; combining the ground object spectral semantic types in the mapping year and the cumulative frequencies of each type to generate a geographical entity classification result with semantic interpretation ability; Step S3, dual-driven collaborative fusion, the data-driven classification result obtained in step S1 is collaboratively fused with the geographical entity classification result obtained in step S2, and combined with the existing land cover mapping data product to generate an optimized annual land cover classification mapping and an optimized near-real-time land cover mapping.

2. The land cover mapping method of claim 1, wherein, The step S1 is specifically: Step S11, obtaining multi-temporal Sentinel-2 ground reflectivity image data of the target region, and pre-processing the same to obtain multi-temporal Sentinel-2 ground reflectivity image data; the pre-processing includes cloud detection, image screening, and radiation correction; Step S12, extracting multi-spectral band information including blue light, green light, red light, red edge band, near-infrared band, and short-wave infrared band from the pre-processed multi-temporal Sentinel-2 ground reflectivity image data obtained in step S11, and calculating derived spectral indices; calculating the statistical characteristic values of the time series data of each spectral band and derived spectral index to construct a multi-temporal feature matrix representing the seasonal changes of ground objects; Step S13, introducing a high-confidence ground class sample set to obtain an initial sample set with high label reliability; comparing the similarities and differences of each feature in the multi-temporal feature matrix of the mapping year and the initial sample set collection year, and filtering samples with high spectral similarity and low difference through similarity threshold and difference threshold to migrate samples and construct a training data set for the mapping year; Step S14, based on the training data set for the mapping year constructed in step S13, using a machine learning classifier to train the multi-temporal feature matrix of the mapping year obtained in step S12 to obtain a land cover classification model suitable for annual mapping; using the land cover classification model to perform full-pixel classification on the multi-temporal feature matrix of the mapping year to output an initial annual land cover classification mapping; Step S15, based on the pre-processed annual multi-temporal Sentinel-2 surface reflectance image data obtained in step S11, obtain the current latest available single scene image to construct a simplified feature set, combine it with the training data set of the mapping year constructed in step S13, train the near real-time land cover mapping classification model, and generate the initial near real-time land cover mapping closest to the current time.

3. The land cover mapping method of claim 2, wherein, In step S12, the derived spectral index includes normalized vegetation index, normalized burning index, and normalized water index.

4. The land cover mapping method of claim 2, wherein, In step S13, the high-confidence class sample set is obtained from higher spatial resolution remote sensing images; the similarity threshold is greater than 0.9, and the difference threshold is less than 0.

1.

5. The land cover mapping method of claim 2, wherein, In step S15, the simplified feature set includes multi-spectral band information of the single scene image, including blue light, green light, red light, red edge band, near-infrared band, and short-wave infrared band, and derived spectral indices such as normalized vegetation index, normalized burning index, and normalized water index calculated using multi-spectral bands of the single scene image.

6. The land cover mapping method of claim 2, wherein, Step S2 is specifically: Step S21, using a spectral image automatic mapper SIAM to process each image in the pre-processed annual multi-temporal Sentinel-2 surface reflectance image data obtained in step S11, the spectral image automatic mapper SIAM divides the pixels into six types of spectral semantic types of ground objects, i.e. strong green vegetation, weak green vegetation, blue water, brown ground, white cloud, and shadow, according to the spectral reflectance relationship between the spectral bands and the preset threshold conditions, and obtains the annual multi-temporal ground class spectral semantic type division result; Step S22, based on the annual multi-temporal ground class spectral semantic type division result obtained in step S21, counting the number of occurrences of each type of ground object spectral semantic type for each pixel, and taking the cloud content of the pixel as a weight, obtaining a normalized annual cumulative frequency value, the value range of the annual cumulative frequency value is 0~1, and the annual cumulative frequency value is divided into eight types, including never, rarely, occasionally, regularly, mostly, usually, often, and always, obtaining the annual ground class spectral semantic type cumulative frequency; Step S23, combining the annual multi-temporal ground class spectral semantic type division result obtained in step S21 and the annual ground class spectral semantic type frequency feature data obtained in step S22, obtaining the feature frequency of different ground class spectral semantic types of each pixel, and dividing each pixel into six types according to the division rule, including land vegetation, land non-vegetation, aquatic vegetation, water, cloud, and shadow, as geographic entity types.

7. The fused data and knowledge dual-driven land cover mapping method according to claim 6, wherein, The division standard of the annual land class spectral semantic type cumulative frequency in the step S22 is as follows: if the annual cumulative frequency value is equal to 0, it is divided into never; if the annual cumulative frequency value is greater than 0 and less than 0.1, it is divided into rarely; if the annual cumulative frequency value is greater than or equal to 0.1 and less than 0.25, it is divided into occasionally; if the annual cumulative frequency value is greater than or equal to 0.25 and less than 0.5, it is divided into regularly; if the annual cumulative frequency value is greater than or equal to 0.5 and less than 0.75, it is divided into mostly; if the annual cumulative frequency value is greater than or equal to 0.75 and less than 0.9, it is divided into usually; if the annual cumulative frequency value is greater than or equal to 0.9 and less than 1, it is divided into often; and if the annual cumulative frequency value is equal to 1, it is divided into always.

8. The fused data and knowledge dual-driven land cover mapping method of claim 6, wherein, In the step S22, the division rule is as follows: first, the land class spectral semantic type and the annual land class spectral semantic type cumulative frequency are combined to obtain the feature frequency of different land class spectral semantic types of each pixel.

9. The fused data and knowledge dual-driven land cover mapping method of claim 6, wherein, The step S3 is specifically: In the step S31, the assertion rule base is constructed based on the data-driven classification result output in the step S1, the geographic entity classification result obtained in the step S2, and the current existing land cover mapping data product, taking each pixel as a geographic entity. In the step S32, pixel-level class correction is performed on the data-driven classification result obtained in the step S1, i.e., the initial annual land cover classification mapping and the initial near real-time land cover mapping, according to the assertion rule base constructed in the step S31, so as to improve the semantic consistency and classification accuracy, and obtain the optimized annual land cover classification mapping and the optimized near real-time land cover mapping.

10. The land cover mapping method of claim 9, wherein, In the step S31, the assertion rule base is respectively set for the annual land cover classification mapping and the near real-time land cover classification mapping, and the land cover types involved are ten types, including cultivated land, forest, grassland, shrub, wetland, water body, tundra, impervious layer, bare land and ice and snow. The basic form of the assertion rule base for the annual land cover classification mapping is: if the initial annual land cover type of a geographic entity is land cover type 1, it is corrected to land cover type 2; provided that the geographic entity is classified as land cover type 2 in the existing land cover mapping data product, and the geographic entity classification result is geographic entity type 1. The basic form of the assertion rule base for the near real-time land cover classification mapping is: if the initial near real-time land cover type of a geographic entity is land cover type 1, it should be corrected to land cover type 2; provided that the geographic entity is classified as land cover type 2 in the optimized last annual land cover classification mapping, and the geographic feature spectral semantic type classification result is geographic feature spectral semantic type 1.

Images