LULC dataset fusion method based on Bayesian-fuzzy probability prediction
Through the Bayesian-fuzzy probability prediction method, the reliability and subjectivity problems in LULC data set fusion are solved, and higher consistency and accuracy are achieved, and no additional auxiliary data is required. It makes full use of fine-grained information and is applicable on different source data sets.
Patent Information
- Application Number
- CN202210791734.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-05
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing LULC dataset fusion method does not have reliability when the pixel space consistency is insufficient, it is susceptible to subjective factors, and cannot make full use of fine-grained information. It also requires the introduction of additional auxiliary data and samples, resulting in error and non-existence. Determinism.
By using Bayesian-fuzzy probability prediction, the target LULC classification system is constructed, the spatial resolution and classification system are unified, the prior and posterior probability are calculated, and the fuzzy probability fusion method with weighted average is used to avoid the introduction of LULC samples and auxiliary. Data, using the maximum fuzzy membership strategy, achieve objective quantification of the fusion results.
Improve the consistency and accuracy of the fusion results, avoid the influence of subjective factors, and make full use of the fine-grained information of all source data sets, which is more applicable and does not require additional auxiliary data.
Smart Images

Figure CN115100526B_ABST
Abstract
Description
Technical Field
[0001] The invention provides a LULC data set fusion method based on Bayesian-fuzzy probability prediction, belonging to the technical field of geographic information processing. Background Art
[0002] A fusion method based on spatial consistency and GWR designed by Linda et al. is from a journal article: See, L., et al., Building a hybrid land cover map with crowdsourcing and geographically weighted regression. ISPRS Journal of Photogrammetry and Remote Sensing, 2015.103: p.48-56.
[0003] This method mainly includes 4 steps:
[0004] (1) First, perform a spatial consistency check on the source dataset;
[0005] (2) For pixels that are consistent in the source dataset, the attribute with the highest consistency is selected and assigned to the pixel;
[0006] (3) For pixels with a consistency of 0 in the source dataset, select some LULC samples closest to the pixel and use GWR to perform regression analysis to determine the probability that the pixel attributes at this location in each source dataset are correct;
[0007] (4) The attribute with the highest probability is assigned to the pixel to form the fusion result.
[0008] This technology has the following disadvantages:
[0009] (1) When the pixel spatial consistency is 0, the probability that the attributes in each source data set are correct may be very low (for example, less than 10%). Even if the attribute with the highest probability is selected, the result is still unreliable.
[0010] (2) LULC samples need to be introduced in the fusion process. Due to the uncertainty and quality of the sampling process of LULC samples, errors and uncertainties are introduced into the fusion results.
[0011] (3) Only the land types that are available in all source datasets can be integrated, and the detailed land type information that is available in only a few datasets cannot be fully utilized.
[0012] (4) In the fusion results, all pixel attributes are selected from the pixel attributes of the location in the source dataset, which does not reflect the advantages of fusion and cannot effectively correct the common errors in the source dataset.
[0013] An evidence fusion method designed by Bai Yan et al. that combines expert scoring, semantic relevance and fuzzy logic is from the journal article: Bai Yan, Feng Min. Research on fusion and evaluation of multi-source land cover data at global scale[J]. Acta Geographica Sinica, 2018, 73(11): 2223-35.
[0014] This method mainly includes 4 steps:
[0015] (1) First, the definition of each land type in each source dataset is compared with the definition of forest, and their relevance scores with forest are determined based on semantic relevance (expert scoring); the forest relevance scores and canopy cover data (ranging from 0 to 100%) of all source datasets are arithmetic averaged. If the arithmetic mean is greater than 0.3, the corresponding pixel attribute is assigned as forest, otherwise it is assigned as non-forest.
[0016] (2) For forest pixels, the correlation scores (expert scoring) between land types and leaf types in the source dataset are determined based on semantic relevance; the leaf type correlation scores and leaf type data (values ranging from 0 to 100%) of all source datasets are averaged, and the leaf type with the largest arithmetic mean is assigned to the corresponding pixel. The correlation scores (expert scoring) between land types and leaf phenology in the source dataset are determined based on semantic relevance; the leaf phenology correlation scores and leaf phenology data (values ranging from 0 to 100%) of all source datasets are averaged, and the leaf phenology with the largest arithmetic mean is assigned to the corresponding pixel.
[0017] (3) For non-forest pixels, the correlation score between the land type in the source data set and cultivated land (expert scoring) is determined based on semantic relevance; the cultivated land correlation scores and cultivated land probability data (ranging from 0 to 100%) of all source data sets are arithmetic averaged. If the arithmetic mean is greater than 0.43, the corresponding pixel attribute is assigned as cultivated land, otherwise it is non-cultivated land.
[0018] (4) For non-cultivated land pixels, the correlation scores (expert scoring) between the land classes in the source dataset and the remaining land classes (grassland, water body, etc.) are determined based on semantic correlation; the arithmetic mean of the correlation scores of all source datasets is calculated for the remaining land classes one by one; the land class attribute with the largest arithmetic mean is assigned to the corresponding pixel. Finally, the first-level class fusion result and the fusion result including the forest leaf attribute are formed.
[0019] This technology has the following disadvantages:
[0020] (1) Expert scoring based on semantic relevance is easily limited by the experience and knowledge level of experts (scientific researchers); and due to the lack of objective standards, the scoring results are highly subjective.
[0021] (2) During the integration process, the strategy of arithmetic mean threshold was adopted for forests and cultivated land, while the strategy of arithmetic mean was adopted for the remaining land types, which lacked unified standards and persuasiveness.
[0022] (3) In the integration process of forest and cultivated land, the arithmetic mean threshold plays an important role. The determination of the forest threshold (0.3) relies on the definition of forest cover by the International Geosphere-Biosphere Program (IGBP), and the determination of the cultivated land threshold (0.43) uses the threshold used by South Dakota State University in the United States when reclassifying the probability of cultivated land into 0 and 1, which lacks in-depth exploration and scientific trade-offs.
[0023] (4) A lot of additional auxiliary data need to be introduced during the fusion process: tree crown coverage data, cultivated land probability data, leaf shape data, and leaf phenology data. Summary of the invention
[0024] In view of the above technical problems, the present invention provides a LULC data set fusion method based on Bayesian-fuzzy probability prediction, which has achieved the following objectives:
[0025] (1) For each pixel, the attribute of the fusion result is selected from all land types with the highest fuzzy membership and the highest reliability.
[0026] (2) LULC samples are not introduced in the fusion process.
[0027] (3) Not only should the land types common to all source datasets be integrated, but also the detailed land type information that is only available in a few datasets should be fully utilized and reflected in the fusion results.
[0028] (4) The pixel attributes in the fusion result are not limited to the attributes of the pixel at that location in the source data. The fusion method can effectively correct the common errors in the source data set.
[0029] (5) Avoid the influence of subjective factors during the integration process.
[0030] (6) The same fusion strategy is adopted for all land types, and unified standards are used to make the method widely applicable.
[0031] (7) Abandon the threshold strategy and adopt the maximum fuzzy membership strategy.
[0032] (8) No additional auxiliary data is introduced in the fusion process.
[0033] For this reason, the specific technical solution of the present invention is:
[0034] The LULC dataset fusion method based on Bayesian-fuzzy probability prediction includes the following steps:
[0035] (1) Construct the target LULC classification system and classify and merge the source data set.
[0036] Construct the target LULC classification system, including primary and secondary classes. Merge the source data sets with different LULC classification systems and different numbers of LULC classes into the constructed target LULC classification system.
[0037] (2) Spatial upscaling.
[0038] For source datasets with different spatial resolutions, they are unified to the same spatial resolution through upscaling.
[0039] (3) Define prior probability.
[0040] The misclassification ratio between different land types is used to represent the misclassification probability. The relevant formula is as follows:
[0041]
[0042] P L (A i |A j )=R L (A i |A j ) (2)
[0043] Among them, A i and A j are the i-th and j-th land types respectively; R L (A i |A j ) is the source data set L. j Classification: A i The ratio of S L (A i |A j ) is L Lieutenant General A j Classification: A i Area in km 2 ; S L (A j ) is A in L j Area in km 2 ;P L (A i |A j ) is L Lieutenant General A j Classification: A i probability.
[0044] Perform accuracy verification on the source dataset or use the accuracy verification results published by the source dataset author to obtain the error matrix, which can reflect the overall misclassification of the source dataset. Use these values to calculate formula 1 and formula 2 to determine the misclassification ratio and probability.
[0045] Afterwards, based on the misclassification probability, the prior probability of each type of land in the source data set is determined by Formula 3:
[0046]
[0047] Among them, A i and A j are the i-th and j-th land types respectively; P L (A i |A j ) is the source data set L. j Classification: A i The probability of L (A j ) is A in L j M is the number of primary classes, which is determined according to the constructed target LULC classification system; P L (A i ) is the pixel in L that is classified as A i The prior probability of .
[0048] (4) Calculation of posterior probability after introducing spatial relationship.
[0049] The neighborhood window is introduced to improve the posterior probability calculation method.
[0050] The specific approach is: take the target pixel as the center, build a neighborhood window of appropriate size; count the proportion of pixels of each type in the window; in the calculation of the posterior probability, use the proportion of the pixel instead of the prior probability. The improved posterior probability calculation formula is:
[0051]
[0052] Among them, A i and A j are the i-th and j-th land types respectively; PL′(A j |A i ) is the source dataset L and the pixel is identified as A i In this case, the pixel is actually A j The probability of , that is, the posterior probability; PL(A i |A j ) is L Lieutenant General A j Classification: A i The probability of PL(A i ) is the pixel in L that is classified as A iThe prior probability of; w is the neighborhood window; R L,W (A j ) is the A in the neighborhood window w j The proportion of quantity.
[0053] (5) Fuzzy probability fusion.
[0054] According to formula 5, the posterior probability of each source data set, pixel by pixel, and class by class is obtained. The posterior probability results of each source data set are synthesized using the weighted average method. The formula is as follows:
[0055]
[0056] Among them, A j is the jth land type; P c ′(A j ) is A on pixel c j The posterior probability of; N is the number of source data sets; P L,c (A j ) is the pixel c in the Lth source dataset. j The posterior probability of OA L is the overall accuracy of the Lth source data set, which can be calculated based on the error matrix in (3).
[0057] Afterwards, for each pixel, the comprehensive posterior probabilities of each first-level class are compared, and the attribute of the land class with the largest comprehensive posterior probability is assigned to the pixel to achieve fuzzy probability fusion:
[0058] A c ={A n |P c ′(A n )=MAX(P c ′(A i )), i = 1, 2, ..., M} (7)
[0059] Among them, A i and A j are the i-th and j-th land types respectively; P′ c (A i ) is A on pixel c i The comprehensive posterior probability of ; M is the number of primary classes, which is determined according to the constructed target LULC classification system; MAX is the maximum value function; A c It is the final land type attribute of pixel c.
[0060] (6) Secondary detailed classification.
[0061] On the basis of obtaining the first-level classification fusion results, the second-level classification is carried out according to the constructed LULC classification system.
[0062] The number of source datasets with secondary land class information is divided into the following two situations:
[0063] ① There are two or more source data sets with secondary land class information. In this case, the same fuzzy posterior probability fusion method as the primary class is used to generate the secondary class results.
[0064] ② Only one source dataset has secondary land class information. In this case, the first-class fusion results are classified into secondary categories one by one based on the only secondary land class information and the principle of spatial correlation. The specific implementation process includes two parts, taking forest land as an example: first, for the pixels that are forest land in both the source dataset and the first-class fusion results, the secondary class attributes of forest land in the source dataset are directly assigned; then, for other pixels identified in the first-class fusion results, the neighborhood search method is used to assign the secondary class attributes of the nearest source dataset forest land pixels to these pixels, thereby completing the secondary subdivision of all forest land pixels in the first-class fusion results. For other first-class classes, the secondary subdivision strategy is the same as that of forest land.
[0065] The technical solution of the present invention has the following beneficial effects:
[0066] (1) The present invention constructs a LULC dataset fusion method based on Bayesian-fuzzy probability prediction. A case study is conducted in the Indochina Peninsula as the study area. The results show that compared with the source dataset, the average overall consistency of the fusion results is improved by 6.42% to 13.61%, the overall accuracy is improved by 4.84% to 7.11%, and the Kappa coefficient is improved by 4.98% to 7.60%. The case study shows the reliability and superiority of the present invention, which can meet the LULC dataset fusion needs of researchers and managers.
[0067] (2) The present invention is based on the source data set itself, and the fusion process does not require the introduction of additional auxiliary data. This effectively avoids the difficulty of obtaining high-quality and available auxiliary data (such as LULC sample data, forest coverage data, cultivated land probability data, leaf shape data, leaf phenology data) in the fusion of high-timeliness (such as 2020) and high-spatial resolution (such as 10m, 30m) LULC data sets.
[0068] (3) In the implementation process of the present invention, the determination of the prior probability, the posterior probability, the comprehensive posterior probability and the fusion result are all achieved through objective quantification, and there is no need to set a threshold, which effectively avoids the influence of subjective factors.
[0069] (4) The present invention not only integrates the land categories that all source data sets have, but also can make full use of the fine land category information of a few source data sets; on the other hand, the present invention does not require the source data sets to have the same spatial resolution, nor does it require them to have the same LULC classification system and number of categories. Therefore, the present invention has lower requirements for source data and has a wider applicability. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0071] The specific technical solution of the present invention is explained in conjunction with embodiments.
[0072] In this example, the Indochina Peninsula is taken as the research area, and LSV10, GLC_FCS30, ESRI10, and Globeland30 are selected as source datasets. Among them, only GLC_FCS30 has fine land classification information.
[0073] like Figure 1 As shown, the embodiment mainly includes 6 steps:
[0074] (1) Construct a target LULC classification system.
[0075] The target LULC classification system (Table 1) constructed in this embodiment first includes 9 first-level categories shared by the four source data sets: cultivated land, forest land, grassland, shrub, wetland, water area, construction land, bare land, and permanent ice and snow. Considering the important value of cultivated land and forest land in scientific research and social applications, the target LULC classification system constructed in this embodiment further subdivides cultivated land into 2 second-level categories: dry land and paddy field; and subdivides forest land into 6 second-level categories: evergreen broad-leaved forest, deciduous broad-leaved forest, evergreen coniferous forest, deciduous coniferous forest, mixed forest, and fruit tree forest (orchard).
[0076] Table 1 Target LULC classification system constructed in this example
[0077]
[0078]
[0079] (2) Spatial upscaling and LULC classification system integration.
[0080] Among the four source data sets, the spatial resolution of LSV10 and ESRI10 is 10 meters, and the spatial resolution of GLC_FCS30 and Globeland30 is 30 meters. It is necessary to perform spatial upscaling operation on LSV10 and ESRI10 to change their spatial resolution to 30 meters so that the spatial resolutions of all four source data sets are consistent; the four source data sets have different LULC classification systems, and they are unified and merged into the target LULC classification system constructed in this embodiment.
[0081] (3) Define prior probability.
[0082] In the development of the source dataset, the original authors conducted accuracy verification. When the source dataset was released, it contained the results of stratified sampling and cross-validation based on the global scope. The error matrix can reflect the overall misclassification of the dataset. We used these values published by the original authors of the four datasets for the prior probability calculation formula to determine the misclassification ratio and misclassification probability. Then, based on the misclassification probability, we determined the prior probability of each type of land in the source dataset.
[0083] (4) Calculation of posterior probability after introducing spatial relationship.
[0084] This embodiment uses 9×9 square neighborhood windows and uses the improved posterior probability formula to calculate the posterior probability.
[0085] (5) Fuzzy probability fusion.
[0086] In (4), the posterior probabilities of each source dataset, each pixel, and each first-level class are calculated. The posterior probability results of each source dataset are integrated using the weighted average method. After that, for each pixel, the comprehensive posterior probabilities of each first-level class are compared, and the attribute of the land class with the largest comprehensive posterior probability is assigned to the pixel to obtain the first-level class fusion result.
[0087] (6) Secondary detailed classification.
[0088] On the basis of obtaining the first-class fusion results, the cultivated land and forest land are further classified at the second level. In this embodiment, only GLC_FCS30 has fine second-class land information. Therefore, firstly, for the pixels that are cultivated land in both GLC_FCS30 and the first-class fusion results, the second-class attributes of GLC_FCS30 are directly assigned; then, for other cultivated land pixels identified in the first-class fusion results, the neighborhood search method is used to assign the second-class attributes of the nearest GLC_FCS30 cultivated land pixels to these pixels, thereby completing the second-level subdivision of all cultivated land pixels in the first-class fusion results. For forest land, its second-level subdivision strategy is the same as that of cultivated land.
Claims
1. LULC dataset fusion method based on Bayesian-fuzzy probability prediction, It is characterized in that The following steps are involved: (1) Construct the target LULC classification system and classify and merge the source data set; Construct a target LULC classification system, including primary and secondary classes; merge source data sets with different LULC classification systems and different numbers of LULC classes into the constructed target LULC classification system; (2) Spatial upscaling: For source datasets with different spatial resolutions, they are unified to the same spatial resolution through upscaling; (3) Define prior probability; The specific steps include: The misclassification ratio between different land types is used to represent the misclassification probability; the relevant formula is as follows: (1) (2) in, and They are , No. Farming type; For the source dataset Lieutenant General Classification proportion; for Lieutenant General Classification Area in km 2 ; for middle Area in km 2 ; for Lieutenant General Classification probability; Verify the accuracy of the source dataset or use the accuracy verification results published by the source dataset author to obtain the error matrix, which can reflect the overall misclassification of the source dataset; use these values in the calculation of formula 1 and formula 2 to determine the misclassification ratio and probability; Afterwards, based on the misclassification probability, the prior probability of each type of land in the source data set is determined by Formula 3: (3) in, and They are , No. Farming type; For the source dataset Lieutenant General Classification probability; for middle The area share of is the number of primary classes, which is determined according to the target LULC classification system constructed; for The pixels are classified into The prior probability of (4) Calculation of posterior probability after introducing spatial relations; The specific steps include: With the target pixel as the center, a neighborhood window of appropriate size is constructed; the proportion of pixels of various types in the window is counted; in the calculation of the posterior probability, the proportion of the pixel is used instead of the prior probability; the improved posterior probability calculation formula is: (5) in, and They are , No. Farming type; For the source dataset Identify the pixels as In this case, the pixel is actually The probability of , that is, the posterior probability; for Lieutenant General Classification probability; for The pixels are classified into The prior probability of is the neighborhood window; Neighborhood window middle The proportion of quantity; (5) Fuzzy probability fusion; According to formula 5, the posterior probability of each source data set, pixel by pixel, and class by class is obtained; the posterior probability results of each source data set are integrated using the weighted average method; the formula is as follows: (6) in, It is Farming type; For pixels c superior The posterior probability of is the number of source data sets; For the Set of pixels in the source dataset c superior The posterior probability of For the The overall accuracy of the source data set can be calculated based on the error matrix in (3); Afterwards, for each pixel, the comprehensive posterior probabilities of each first-level class are compared, and the attribute of the land class with the largest comprehensive posterior probability is assigned to the pixel to achieve fuzzy probability fusion: (7) in, and They are , No. Farming type; For pixels c superior The combined posterior probability of is the number of primary classes, which is determined according to the target LULC classification system constructed; is the maximum value function; It is a pixel c The final land type attributes; (6) Secondary fine classification: Based on the first-level classification results, secondary classification is performed according to the constructed LULC classification system.
2. The LULC data set fusion method based on Bayesian-fuzzy probability prediction according to claim 1, It is characterized in that In step (6), the number of source datasets with secondary land category information is divided into the following two cases: 1 There are two or more source data sets with secondary land class information; in this case, the same fuzzy posterior probability fusion method as the primary class is used to generate the secondary class results; 2Only one source dataset has secondary land category information; In this case, the secondary classification of the first-level classification results is carried out land-class by land-class based on the only secondary land-class information and the principle of spatial correlation.
Citation Information
Patent Citations
Quality control method and device for geographic national condition monitoring data and terminal equipment
CN109657907A
Surface covering product integration or fusion method based on EAGLE matrix
CN110033044A
