A method for excavating rare alpine vegetation species existence form and growth area

Abstract

The application discloses a method for mining the existing form and growth area of rare alpine vegetation species, comprising the following steps: step S1, establishing a spatial data set containing rare alpine vegetation species, and obtaining a spatial feature set; step S2, screening core features in the spatial data set, and dividing sub-regions based on the core features; the core features are rare alpine vegetation species; step S3, mining frequent core patterns in each sub-region obtained through division; the core pattern is a subset in the spatial feature set, wherein the core pattern must contain the core features; step S4, collecting the frequent core patterns in all sub-regions, and establishing a query dictionary; and step S5, querying the mining result of the existing form and growth area of rare alpine vegetation species based on the established query dictionary. The application solves the problems that there is no mining method for the existing form and growth area of rare alpine vegetation species in the prior art, and the traditional distance threshold measurement method has defects.

Description

Technical Field

[0001] This invention belongs to the field of regional spatial data mining technology, and in particular relates to a method for mining the existence forms and growth areas of rare alpine vegetation species. Background Technology

[0002] With the rapid development of location-based services and spatial data acquisition technologies, data mining has expanded from traditional transactional databases to spatial databases. How to extract potentially interesting knowledge from massive amounts of high-dimensional spatial data and use it to guide decision-making has become crucial. Spatial isotope pattern mining is a very important research direction in the field of spatial data mining. A spatial isotope pattern refers to a set of spatial features that can discover implicit relationships between different spatial features, which often have strong geographical relevance. Examples include many communities around schools, pharmacies around hospitals, and parking lots around shopping malls. Due to the heterogeneity of spatial data distribution, some spatial isotope patterns may only be distributed in certain specific areas. Regional isotope pattern mining is a branch of spatial isotope pattern mining, used to discover isotope patterns that do not frequently co-occur globally but frequently co-occur within local areas. For example, in the field of public health, the overall incidence of malignant tumors in a certain region may not seem high, but the incidence is particularly high in areas with heavy industrial pollution. In the field of urban safety, high crime rates in cities are mainly concentrated near bars in the city center, while this phenomenon is not observed in other areas. This implicit regional correlation is called a regional frequent spatial isomorphism pattern, which represents a subset of spatial features that frequently appear together in certain sub-regions of the study area. Regional isomorphism pattern mining can uncover the correlations between spatial features implicit in local areas, more objectively reflecting the patterns exhibited by these spatial features in local areas. This has guiding significance for areas including urban planning, environmental protection, and social governance.

[0003] Based on existing research and extensive surveys, the causes of regionalization or the formation of regional isomorphic patterns can be broadly categorized into three types: 1) naturally formed geographical factors, such as mountains and rivers; 2) formed by human intervention, such as nature reserves and ecological bases; and 3) clustered due to the correlation between features, such as dependency relationships (vines clinging to trees to obtain sunlight). Existing regionalization methods do not fully consider these three factors. The main existing methods are clustering-based, rule-based, and network-constraint-based (e.g., prior information such as road networks and maps). Clustering and rule-based methods primarily consider the proximity relationships between instances, arguing that geographically proximate instances have stronger correlations, which is effective in uncovering regional isomorphic patterns formed by geographical factors. Network-constraint-based methods require specifying prior information, such as land cover and jurisdiction information, which is helpful for uncovering regional isomorphic patterns, but lacks research specifically on regional isomorphic patterns formed by the correlation and constraints between features.

[0004] Regional isotope patterns formed by the correlation between features are of significant research value in discovering the existence forms and growth areas of rare alpine vegetation species. Therefore, this invention proposes a regional kernel pattern mining method for discovering the existence forms and growth areas of rare alpine vegetation. Regional kernel patterns can be used to discover spatial isotope relationships in sub-local regions caused by the presence of kernel features. For example, a group with cancer is usually a rare group in reality, so using cancer as a rare resource can lead to the inference of pollution sources that cause cancer. At the same time, due to the dual heterogeneity of spatial data in terms of spatial distribution and feature instance quantity distribution, some spatial features themselves exhibit local distribution and are rare features. In practical applications, discovering kernel patterns containing rare features is of greater research significance and application value than searching for regional isotope patterns with low attention across the entire scope.

[0005] Furthermore, existing research often uses Euclidean distance to measure the proximity of two spatial instances, but this quantitative method has some drawbacks. Instance density varies across different spatial datasets, and even within the same dataset, different spatial features can lead to significant differences in instance density. Therefore, selecting a suitable distance threshold requires a thorough understanding of the dataset and prior knowledge of the relevant domain, which is clearly difficult to achieve. In practical applications, whether in ecological conservation or urban planning, instances with similar characteristics often compete with each other (e.g., the "two tigers cannot share one mountain" phenomenon in ecology, and two similar supermarkets in urban planning often being several blocks apart rather than adjacent). Therefore, this invention proposes a kernel proximity attribution metric to replace traditional distance threshold-based methods. Using a kernel proximity attribution metric not only replaces traditional distance threshold methods but also addresses, to some extent, the complex coupling relationships between spatial instances. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying the existence forms and growth areas of rare alpine vegetation species, in order to solve the problems of the lack of a method for identifying the existence forms and growth areas of rare alpine vegetation species in the technology, and the defects of traditional distance threshold measurement methods.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for identifying the existence forms and growth areas of rare alpine vegetation species, comprising the following steps: Step S1: Establish a spatial dataset containing rare alpine vegetation species and obtain the spatial feature set F; Step S2: Filter kernel features in the spatial dataset and divide sub-regions based on the distribution influence of the kernel features; the kernel features are rare alpine vegetation species. Step S3: Mine frequent kernel patterns in each of the divided sub-regions; the kernel pattern c is a k-order subset of the spatial feature set F, where c must contain kernel feature cf (cf ∈ F); the kernel pattern refers to a spatial isotopic pattern containing rare alpine vegetation species. Step S4: Collect frequent kernel patterns in all sub-regions and establish a query dictionary for the existence forms and growth areas of alpine vegetation species; Step S5: Based on the established query dictionary, query the existence forms and growth areas of rare alpine vegetation species.

[0008] Further, step S1 specifically includes: Establish a spatial dataset, where each vegetation type in the dataset is combined into a spatial feature set. , where n represents the number of feature categories; each feature category contains a set of feature instances distributed at different locations in space; all feature instances constitute the spatial instance set. The i-th spatial instance is represented as , It is described by a triple <feature category, instance number, spatial location>, where the instance number represents the ordinal number of the spatial instance among all instances of the feature category, and the spatial location refers to the spatial location of the spatial instance; the spatial instance... This refers to a single, actually growing alpine plant.

[0009] Furthermore, the method for selecting kernel features in step S2 is as follows: Use code or online tools to generate box plots of all features in a spatial dataset on the horizontal and vertical distributions. In the box plots, features with narrower quartile intervals or larger median deviations are considered kernel features.

[0010] Furthermore, the sub-region division method in step S2 is as follows: First, calculate the partition distance threshold pd. Then, based on each kernel instance and the partition distance threshold pd, add all nearby instances to the set it forms. Determine if the set contains other feature instances. If it does not contain other feature instances, it means that there are only kernel instances in the current region, and the set will be discarded. Finally, merge the sets containing the same instances to obtain the final sub-region partitioning result. The partition distance threshold pd is set to 1.5. rd to 2 Between rd, where rd is the relative distance between kernel instances:

[0011] Among them, dist_min(co i ) Computational kernel instance The minimum competition distance, where i is the i-th core instance and m represents the total number of all core instances; The nuclear example is an instance of a rare alpine vegetation species.

[0012] Further, step S3 includes: Step S31: Construct the storage structure Core_Hash; Step S32: Mining frequent core patterns based on the Core_Hash structure; The Core_Hash storage structure is constructed as follows: For each non-core instance, iterate through all core instances and determine whether the current non-core instance satisfies the proximity relationship with a core instance. If it does, add the corresponding information to Core_Hash. If the key of Core_Hash already contains the current core instance, directly add the non-core instances adjacent to the current core instance to the corresponding feature table. Finally, the Core_Hash structure is obtained. The nuclear example is an example of a rare alpine vegetation species, and the non-nuclear example is an example of a non-rare alpine vegetation species.

[0013] Furthermore, the method for determining whether a non-kernel instance satisfies the proximity relationship with a kernel instance is as follows: Given a maximum number of kernel neighbors x, for a non-kernel instance, the non-kernel instance can be assigned to a maximum of x nearest kernel instances.

[0014] Further, step S32 specifically includes: First, based on Core_Hash, we obtain the second-order frequent kernel patterns. Then, we combine the second-order frequent kernel patterns to generate third-order candidate kernel patterns. Based on Core_Hash, we calculate the kernel participation degree of all candidate kernel patterns. We find the frequent kernel patterns among the candidate kernel patterns and repeat the process of "generating k+1 order candidate kernel patterns from k-order frequent kernel patterns and then judging whether the k+1 order candidate kernel patterns are frequent" until the generated candidate kernel patterns are empty. Finally, we obtain all the frequent kernel patterns. The method for determining whether the kernel mode is frequent is as follows: The nuclear participation degree of the nuclear mode was obtained through calculation. Given a minimum kernel participation threshold min_cprev, if CPI(R, c)≥min_cprev, then kernel pattern c is said to be a frequent kernel pattern in region R, and region R is called the frequent region of kernel pattern c.

[0015] Furthermore, step S4 specifically includes: Step S41: For all the collected frequent kernel patterns, establish a query dictionary structure {kernel pattern: {all frequent sub-regions}}; Step S42: For all the sub-regions divided in step S2, establish another query dictionary structure {sub-region: {all frequent kernel patterns}}.

[0016] The beneficial effects of this invention are: First, this invention reveals the existence forms of rare alpine vegetation species and discovers their frequently occurring areas by mining regional kernel patterns, filling the gaps in existing methods. Second, this invention provides mining methods under two proximity relation measurement modes, which can be adapted to spatial datasets in various application backgrounds. Third, this invention has made technical optimizations to both the region division stage and the pattern mining stage, significantly improving mining efficiency. Finally, specific embodiments show that the mining effect of this invention is superior to existing methods, not only discovering more existence forms of rare alpine vegetation species, but also finding their more comprehensive growth areas, achieving excellent results while maintaining efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the existence forms of rare alpine vegetation species and the method for identifying their growth areas, according to an embodiment of the present invention.

[0019] Figure 2 This is an example diagram of a spatial instance containing a local region according to an embodiment of the present invention.

[0020] Figure 3 These are box plots of the spatial distribution of different spatial features in the horizontal and vertical directions in real and synthetic datasets, according to embodiments of the present invention. (a) is a box plot of the spatial distribution of the Gaoligong Mountain vegetation dataset, and (b) is a box plot of the spatial distribution of a synthetic dataset.

[0021] Figure 4 This embodiment of the invention is based on the synthetic dataset. Figure 3 The spatial distribution scatter plot of rare resource characteristics selected by box plot.

[0022] Figure 5 These are example diagrams of region partitioning based on clustering and scanning iteration methods according to embodiments of the present invention, wherein (a) is an example diagram of region partitioning results based on clustering and (b) is an example diagram of region partitioning results based on scanning iteration.

[0023] Figure 6 These are examples of region partitioning based on kernel features according to embodiments of the present invention, wherein (a) is an example of region partitioning criteria based on kernel features, and (b) is an example of region partitioning results based on kernel features.

[0024] Figure 7 This is an example diagram of the storage structure of Core_Hash in an embodiment of the present invention.

[0025] Figure 8 This is an example diagram of the spatial proximity measurement method between instances according to an embodiment of the present invention, wherein (a) is the effect diagram using distance threshold measurement, and (b) is the effect diagram using kernel proximity attribution measurement.

[0026] Figure 9 These are spatial distribution maps of the real dataset and the synthetic dataset in the embodiments of the present invention, wherein (a) is the spatial distribution map of the vegetation data of Gaoligong Mountain, and (b) is the spatial distribution map of the synthetic data.

[0027] Figure 10 This is a display image of the existence forms and growth areas of rare alpine vegetation species mined by two comparative methods on the Gaoligong Mountain vegetation dataset in an embodiment of the present invention. (a) shows the existence forms and growth areas of rare alpine vegetation species mined by the Multi-level method, and (b) shows the existence forms and growth areas of rare alpine vegetation species mined by the NNMulti-level method.

[0028] Figure 11 This is a comparison map of frequent regions with the same pattern mined by the proposed method and the comparison method in a synthetic dataset according to an embodiment of the present invention. (a) is the frequent region identified by the comparison method, and (b) is the frequent region identified by the proposed method.

[0029] Figure 12 This is a diagram showing the existence forms and growth areas of rare alpine vegetation species mined from the Gaoligong Mountain vegetation dataset using the method proposed in this invention.

[0030] Figure 13 This is a comparison chart of the time impact of different data scales on method efficiency in an embodiment of the present invention, where (a) represents the impact of the number of instances on method efficiency, and (b) represents the impact of the number of core instances on method efficiency.

[0031] Figure 14 This is a comparison chart of the impact of the number of features and the minimum frequency threshold on the method efficiency in an embodiment of the present invention, where (a) shows the impact of the number of features on the method efficiency and (b) shows the impact of the minimum frequency threshold on the method efficiency. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] like Figure 1 As shown, this invention provides a method for identifying the existence forms and growth areas of rare alpine vegetation species, with the specific steps as follows: Step S1: Data Processing. Given a dataset of alpine vegetation distribution, each vegetation type in the dataset is called a feature category, forming a spatial feature set. , where n represents the number of feature categories. Each feature category contains a set of feature entities (instances) distributed at different locations in space. All feature instances constitute the spatial instance set. Where n represents the number of feature categories. The i-th spatial instance is represented as This is described by a triple <feature category, instance number, spatial location>. Here, the instance number represents the ordinal number of the instance within all instances of its feature category, and the spatial location refers to the spatial location of the feature instance. Specifically, in this invention, the spatial dataset is an alpine vegetation dataset, where the spatial feature set F represents the set of all alpine vegetation categories, and the spatial instance... This refers to a single, actually growing alpine plant.

[0034] Taking traditional spatial datasets as an example, such as Figure 2 The given example contains 6 feature categories, a spatial feature category set. Among them, the feature category "hospital" has 5 spatial instances, corresponding to the spatial instance set. For example For example, it is represented in the dataset as<H,1,[x, y]> , where [x, y] represents its spatial coordinates. If the distance between two instances is less than a given threshold, they are connected by a solid line to indicate that the two instances are adjacent.

[0035] To facilitate understanding of subsequent steps, the embodiments of this invention provide the following related concepts and definitions: Definition 1: Core Pattern. A core feature (in this invention, the core feature is generally defined as a rare alpine vegetation species) cf (cf ∈ F) is specified. The core pattern c is defined as a k-order subset of the feature set F, where c must contain the core feature cf (in this invention, a core pattern refers to a spatial isomorphic pattern containing a rare alpine vegetation species, i.e., the existence form of the rare alpine vegetation species). A pattern within a region R defined by the influence of a core feature is called a regional core pattern within that region R (in this invention, region R refers to the growth area of ​​the rare alpine vegetation species' existence form).

[0036] Definition 2: Kernel Proximity Relationship. If a kernel instance co i (In this invention, it refers to an example of a rare alpine vegetation species) and a non-nucleate example. j (In this invention, it refers to an example of a non-rare alpine vegetation species) If they are adjacent to each other, they are said to satisfy the nuclear proximity relationship. There is no nuclear proximity relationship between nuclear instances or between non-nuclear instances.

[0037] Definition 3: Kernel row instances and kernel table instances. In a region R partitioned by the influence of kernel features, given a kernel pattern c, if the instance set... The following conditions must be met: 1) I contains all features in c, and a subset of I does not contain all features in c; 2) There exists a kernel instance co. i If I satisfies the kernel proximity relation with all remaining instances in it, then I is called a kernel row instance of kernel pattern c in region R. The set of all kernel row instances constitutes the kernel table instance of kernel pattern c, denoted as CT(Q, c).

[0038] For example, in Figure 2 In the traditional spatial dataset shown, hospitals are assumed to be the kernel feature. In region 1, the kernel table instances of the kernel pattern {hospital, flower shop, pharmacy} are {{H1, F1, P1}, {H1, F1, P2}, {H2, F2, P3}, {H3, F3, P4}}}.

[0039] Definition 4: Kernel Participation Instance. Given a kernel pattern c and a feature f i ∈c, feature f i Instances that do not repeat in the kernel table are called features f in kernel pattern c. i The kernel participation instance is denoted as CP(f) i | c). If within a region R divided by the influence of kernel features, the feature f of kernel pattern c i The kernel participation instance is the feature f in this region. i An instance that does not appear repeatedly in the kernel table is represented as CP(R, f). i | c).

[0040] Definition 5: Participation rate and participation degree of kernel patterns. In a region R partitioned by the influence of kernel features, given a kernel pattern c and a feature f... i ∈c, feature f i The nuclear participation rate in c is defined as CPR(R, c, f). i The specific formula is as follows:

[0041] in It is feature f i The total number of instances in region R.

[0042] The kernel participation degree of a k-order kernel pattern c is defined as the minimum kernel participation rate of all features in c, denoted as CPI(R,c), and the specific formula is as follows:

[0043] In the formula, i represents the i-th feature.

[0044] Given a minimum kernel participation threshold min_cprev, if CPI(R, c)≥min_cprev, then kernel pattern c is said to be a frequent kernel pattern in region R, and region R is called the frequent region of kernel pattern c.

[0045] Step S2: Given a kernel feature, divide the region into sub-regions based on the distribution influence of the kernel feature. The specific steps are as follows: Step S21: Selection of kernel features. To quickly filter out rare vegetation species or those exhibiting localized distributions from a high-altitude vegetation distribution dataset, this embodiment of the invention provides a rapid filtering method based on a box plot of the dataset's spatial distribution. Its specific description is as follows: Given a spatial dataset, use code or online tools to generate boxplots of all features in the dataset along their horizontal and vertical distributions. For example... Figure 3 As shown, two spatial datasets are presented (one real dataset and one synthetic dataset; for details and their spatial distribution, see [link to documentation]). Figure 9 Box plots showing the spatial distribution of each feature in the horizontal and vertical directions. (Feature spatial distribution box plots) Figure 1 To a certain extent, this reflects the dispersion of feature instances in the global distribution. Features with narrower quartile intervals or larger median deviations are rare features or features with local distributions. Figure 3 In (a), most features show narrow boxplots in both the horizontal and vertical directions with large median deviations, indicating that most features in this dataset exhibit localized distributions. Meanwhile... Figure 3In (b), the box plots of features B and J are relatively narrow, and the median deviation of feature J in the horizontal distribution box plot is relatively large. This indicates that features B and J may exhibit local distribution and belong to rare features, which are more suitable as kernel features. Figure 4 This shows a scatter plot illustrating the spatial distribution of features A, B, and J in the synthetic dataset. Figure 4 It is clear from the data that feature A is globally distributed, while features B and J are locally distributed and are likely rare features, which confirms the correctness and feasibility of the proposed scheme.

[0046] Step S22: Divide the region into sub-regions based on the distribution influence of the selected kernel features. This embodiment of the invention considers multiple region division strategies simultaneously, and by comparing their differences in effectiveness, finally formulates an innovative partitioning strategy.

[0047] First, this invention performs region partitioning considering only the distribution of kernel instances. Two methods readily come to mind for partitioning sub-regions considering only kernel instance distribution: one is to cluster kernel instances, and the other is to expand adjacent spatial instances using a scanning iteration method, ultimately obtaining the expanded sub-regions. Figure 5 This demonstrates a simple example of dividing sub-regions using the two methods described above (yellow features represent the specified kernel features), where... Figure 5 (a) shows the region partitioning results obtained using a kernel instance-based clustering method. Figure 5 (b) shows the region partitioning result obtained using the scanning iteration method. From Figure 5 The partitioning results show that both methods only focus on the spatial distribution of kernel instances, ignoring the influence of other non-kernel instances adjacent to kernel instances. This results in a sub-region containing far more kernel instances than non-kernel instances, and in extreme cases, a sub-region containing only kernel instances. Clearly, the sub-regions obtained in this situation are unreasonable, and it is difficult to find effective kernel patterns within such sub-regions.

[0048] Therefore, this invention incorporates the influence of non-nuclear instances while considering the distribution of kernel instances, thus designing a region partitioning method based on the influence of kernel features. This method not only considers the distribution of kernel instances but also their impact on non-nuclear instances. Specifically, a kernel feature (rare alpine vegetation species) is specified, and its kernel instances are represented as co. i (0 < i < m), where m represents the total number of kernel instances for this kernel feature. A partition distance threshold pd is calculated based on the spatial distribution of kernel instances. The definition and calculation process of pd are as follows: First, calculate the relative distance rd between core instances. The minimum contention distance of a core instance is defined as the distance to its nearest neighbor core instance, while the relative distance rd between core instances is defined as the mean of the minimum contention distances of all core instances, that is:

[0049] Among them, dist_min(co i ) Computational kernel instance The minimum competition distance is given by , where i is the i-th core instance and m represents the total number of core instances. The relative density rd among core instances reflects the average density of the core instance distribution, or the range of competition influence. In practical applications, this can be understood as an average of one core instance appearing every rd interval.

[0050] Then, the partition distance threshold pd is obtained based on the relative distance rd between kernel instances. Considering the expansion and merging of feature instances during partitioning, a kernel instance needs to be merged with its neighboring kernel instances and non-kernel instances. This requires the partition distance threshold pd to be slightly larger than the relative distance rd between kernel instances. Therefore, pd can be set to 1.5 during partitioning. rd to 2 Between rd, that is:

[0051] If pd is less than 1.5 rd might result in sub-regions being divided too small, while if pd is greater than 2 However, rd may lead to the sub-regions being divided too large.

[0052] After obtaining the partition distance threshold pd, under the constraint of pd, taking a kernel instance as the starting partition point, the partitioning strategy designed in this embodiment of the invention is as follows: 1) If a core instance and all its neighboring (proximity under the partition distance pd constraint) instances are core instances, then the core instance will not be used as the starting partition point. If a core instance has neighboring non-core instances, then the core instance and its neighboring non-core instances can be used together as the starting partition. (It is guaranteed that partitions containing only core instances will not occur.) 2) A core instance in a partition cannot be used to add only kernel instances that are adjacent to the core instance, but it can be used to add all instances that are adjacent to it except under this condition. (This ensures that regions containing only core instances will not be added during region expansion.) 3) A non-core instance in a partition can only pull all adjacent core instances into that partition, while a non-core instance cannot pull adjacent non-core instances into that partition. (This ensures that non-core instances affected by a core instance are placed in the same partition as that core instance.) 4) If any unpartitioned kernel instances exist, repeat the above steps using those kernel instances as the new partitioning starting point until all sub-regions are partitioned. (Completeness) Figure 6 (a) Demonstrates the detailed steps of the partitioning strategy, where the yellow features are the specified kernel features. Figure 6 In (a), the dashed area represents the invalid area, the blue area represents the valid starting sub-region, and the red and black areas represent the sub-regions after one expansion and two expansions. Figure 6 The left figure in (a) corresponds to partitioning strategy 1), where a starting sub-region cannot be formed when there is only one kernel instance or only a single feature instance. This is to avoid invalid partitions where the number of kernel instances is much greater than the number of other feature instances or where there is only a single feature. Figure 6 (a) corresponds to partitioning strategy 2). The kernel instance indicated by the arrow will not be added to the kernel instance that is only adjacent to the kernel instance, ensuring that the region expansion will not add a region that only contains kernel instances. Figure 6 (a) The right figure corresponds to partitioning strategy 3). Non-core instances within a partition can add all adjacent core instances to the partition, while non-core instances cannot be added to sub-partitions through proximity relationships with non-core instances. Figure 6 (b) presents the region partitioning results of this strategy, in comparison. Figure 5 Based on the partitioning results of clustering and scanning iteration, the new region partitioning strategy designed in this invention can avoid erroneous partitioning results caused by focusing only on the distribution of kernel instances, and the partitioned sub-regions are more likely to have kernel patterns.

[0053] The above content describes a strategy for region partitioning. If region partitioning is performed directly based on this strategy, the method is similar to breadth-first search, which has a high time complexity, with a worst-case time complexity of O(ms). 2 ), where m is the number of core instances and s is the total number of instances. To improve partitioning efficiency, this invention optimizes the above method. Specifically, under the constraint of the partitioning distance threshold pd, for each core instance, a set is generated by finding all instances that satisfy the proximity relationship with it. Then, it is determined whether there is an intersection between these sets. If there is an intersection, these sets are merged to finally obtain the partitioning result. The optimized partitioning method has a time complexity reduced to O(ms + slog2s). The partitioning method mainly consists of the following two stages: 1) Specify the kernel feature cf, and each kernel instance and all instances under the partition distance pd form a starting partition.

[0054] 2) Determine if there are common instances in these partitions. If so, merge these partitions into a single sub-region.

[0055] The specifics are shown in Method 1.

[0056]

[0057] Method 1 describes the specific process based on the region partitioning strategy proposed in this invention. First, the partition distance threshold pd is calculated (step 2). Then, for each kernel instance, all instances adjacent to it are added to the set it forms (steps 3-6). It is then determined whether the set contains other feature instances. If no other feature instances are found, it means that there are only kernel instances in the current region, and the set will be discarded (steps 7, 8). Finally, sets containing the same instances are merged to obtain the final sub-region partitioning result (step 9).

[0058] Step S3: Mining core patterns in each of the divided sub-regions. To efficiently mine all core patterns, this invention designs an efficient storage structure, Core_Hash, based on a hash table, and an efficient mining method based on this storage structure. The specific steps are as follows: Step S31: Construct an efficient storage structure, Core_Hash. To efficiently query core-participating instances of a pattern during core pattern mining without needing to determine the proximity relationships between instances each time, this invention designs an efficient hash storage structure, represented as {rare core instances: {nearby non-core feature sets: <instance ID>}}, denoted as Core_Hash. For each core instance, a Core_Hash is constructed by searching for non-core instances that satisfy its proximity relationship. Using Core_Hash, core-participating instances of all features in the corresponding core pattern can be easily found.

[0059] like Figure 7 In this example, assuming feature A is the kernel feature, a Core_Hash is constructed based on kernel instances A1, A2, and A3. When calculating the participating instances of a kernel pattern, it is only necessary to determine whether the "nearest non-kernel feature set" in Core_Hash contains all non-kernel features of the kernel pattern. If it does, then the instances of the corresponding features in Core_Hash belong to the participating instances of that kernel feature. For example, the participating instances of feature A corresponding to kernel pattern {A, B, C, D, E} are {A1}, and the participating instances of feature C are {C1, C2, C3}. The participating instances of feature A corresponding to kernel pattern {A, B, D} are {A1, A2, A3}, and the participating instances of feature B are {B1, B2, B3, B4, B5, B6, B9}, meaning A1, A2, and A3 correspond to the union of the sets of instances of feature B in Core_Hash.

[0060] It is worth noting that when determining the proximity relationship between kernel instances and non-kernel instances, most existing methods use a distance threshold based on Euclidean distance. While using a distance threshold to measure the spatial proximity relationship between instances is simple, the spatial distribution density can vary significantly for different spatial features, and a single distance threshold cannot comprehensively consider the dependencies between instances with different features. Furthermore, instances with the same features often have competitive relationships. For example... Figure 8 As shown in (a), if only the distance threshold is considered, then B2 is affected equally by A1 and A2, but the instance that has the greatest impact on B2 should be A2. This is because the distance between B2 and A2 is significantly shorter than that between A1 and A2.

[0061] Therefore, this invention proposes a kernel proximity attribution metric to replace existing distance-threshold-based metrics: given a maximum kernel proximity attribution number x, for a non-kernel instance, it can be at most attributed to x nearest kernel instances. In other words, this non-kernel instance satisfies a spatial proximity relationship with its x nearest kernel instances. The nearest attribution metric considers not only the cohesion / support relationship between kernel instances and non-kernel instances but also emphasizes the competitive relationship between kernel instances. Figure 8 In (b), if x=1, then B2 will only be affected by A2. If x=2, then B2 will be affected by both A2 and A1.

[0062] The storage method for Core_Hash is applicable to both distance-based threshold and kernel-based nearest-home methods, differing only in the metric used when generating Core_Hash. This invention implements mining methods under both versions, which will be discussed separately in the embodiments section of this invention. Generating Core_Hash requires traversing the kernel instance set and non-kernel instance set, with a time complexity of O(m*(sm)), where s is the total number of instances and m is the number of kernel instances. The specific method is shown in Method 2.

[0063]

[0064] Method 2 describes the specific construction process of Core_Hash. For each non-core instance, all core instances are traversed (steps 2-3), and it is determined whether the current non-core instance satisfies the proximity relationship with a core instance (step 4). If it does, the corresponding information is added to Core_Hash (steps 5-6). If the key of Core_Hash already contains the current core instance, its adjacent non-core instances are directly added to their corresponding feature tables (steps 7-8). Finally, the Core_Hash structure is returned (step 9). The Metrics() function is used to determine whether the current core instance is one of the x closest core instances to the current non-core instance.

[0065] Step S32: Mining frequent kernel patterns based on the Core_Hash structure. After constructing the Core_Hash structure, the kernel participation instances of a kernel pattern can be quickly obtained by looking up a table, thus calculating the kernel participation degree. Finally, the frequency of a pattern is determined by comparing the kernel participation degrees of all kernel patterns with the minimum frequency threshold. However, directly calculating the kernel participation degree of all possible candidate patterns has a high time complexity. The time complexity of calculating the kernel participation degree of a kernel pattern is O(m). In the worst case where downward closure is not satisfied, it is necessary to calculate the kernel participation degree of all possible candidate kernel patterns, with a time complexity of O(m*2). k To reduce time consumption, this invention employs a step-by-step filtering method for pruning. Specifically: Lemma 1: Antimonotonicity of kernel patterns. Given a kernel pattern c, its kernel participation CPI(R, c) is antimonotonic in a given region R, that is, as the order of the kernel pattern increases, the kernel participation CPI(R, c) decreases.

[0066] Proof: In region R, assume there are kernel patterns c and kernel patterns. satisfy , , where k' < k and If a feature f i If an instance of (i < k') is included in a kernel line instance of kernel mode c, then it must also be included in kernel mode c. In the kernel instance, the converse is also true. Therefore, there must be... ,in For feature f i In nuclear mode The nuclear participation rate in the middle can be used to deduce .

[0067] Lemma 2: The A priori principle of kernel patterns. In a region R, if a kernel pattern is frequent, then all its sub-kernel patterns (subsets with kernel features) are also frequent in that region. Conversely, if a kernel pattern is infrequent, then any of its superkernel patterns (supersets with kernel features) must also be infrequent in that region. For a kernel pattern {A, B, C}, if A is the kernel, then {A, B} is one of its sub-kernel patterns, and {A, B, C, D} is one of its superkernel patterns.

[0068] Proof: Assume that a kernel mode exists within a local region R. , , ,in and Based on Lemma 1, it can be concluded that in the region R, there exists Assume that c is not a frequent kernel pattern within region R, i.e. It is obvious that... pattern It is not a frequent kernel mode within region R. Similarly, it can be proved that the supermode also satisfies this lemma.

[0069] The specific mining method is shown in Method 3.

[0070]

[0071] Method 3 describes the mining process of kernel patterns in a sub-region based on Core_Hash. First, frequent kernel patterns of order 2 are found according to Core_Hash (steps 2-5). Based on Lemma 1 and Lemma 2, frequent kernel patterns of order 2 are combined to generate candidate kernel patterns of order 3 (step 6). Frequent kernel patterns are found among the candidate kernel patterns, and the process of "generating k+1 order candidate kernel patterns from k frequent kernel patterns and then judging whether the k+1 order candidate kernel patterns are frequent" is repeated until the generated candidate kernel patterns are empty (steps 7-12). Finally, all frequent kernel patterns are returned (step 13). Assuming the kernel feature is A, the frequent kernel patterns {A,B,C} and {A,B,D} can be combined to generate the candidate kernel pattern {A,B,C,D}. (Note that the frequent kernel patterns that can be combined must contain the same subset of k-1 order kernel patterns, such as {A,B,C} and {A,B,D} both containing {A,B}).

[0072] Step S4: Collect frequent kernel patterns from all sub-regions and create two dictionaries. In the first dictionary (Dictionary 1), each frequent kernel pattern is used as the key, and the corresponding numbers of all frequent sub-regions are used as the values. By specifying a frequent kernel pattern, all its frequent sub-regions can be quickly obtained. In the second dictionary (Dictionary 2), all sub-regions are used as the keys, and all frequent kernel patterns within that region are used as the values. The specific steps are as follows: Step S41: For all the collected frequent kernel patterns, create a dictionary structure {kernel pattern: {all frequent subregions}}. For example, {kernel pattern {A, B}: {subregion 1, subregion 2}} indicates that the kernel pattern {A, B} appears frequently in subregion 1 and subregion 2.

[0073] Step S42: For all the sub-regions divided in step S2, create a dictionary structure {sub-region: {all frequent kernel patterns}}. For example, {sub-region 1: {{kernel pattern {A, B}}, {kernel pattern {A, C}}}} indicates that the kernel patterns that frequently appear in sub-region 1 are {A, B} and {A, C}.

[0074] Step S5: Obtain the mining results based on the established dictionary. The mining results can assist in data analysis and provide further decision-making assistance based on user selections. For each mined kernel pattern, it can reveal the frequent occurrence of kernel features, explain to some extent the reasons for its occurrence, and analyze the conditions for its frequent occurrence, providing suggestions for the utilization and protection of alpine vegetation species. Simultaneously, for a frequent kernel pattern, all its frequently occurring sub-regions can be viewed. By analyzing the common factors in these sub-regions, the geographical factors suitable for the generation or occurrence of this kernel pattern can be further determined.

[0075] Example To verify the correctness of this invention, this embodiment is divided into two parts: a real dataset and a synthetic dataset. This embodiment uses the Gaoligong Mountain vegetation dataset as the real-world example to verify the correctness and effectiveness of the invention. The real dataset is centrally distributed; to investigate the performance of the mining method on clustered datasets, this invention also designed a clustered synthetic dataset. Furthermore, this invention compares the mining results with two excellent regional pattern mining methods in recent years: Multi-level and NNMulti-level. All experiments in this embodiment are implemented in Java and conducted on a computer configured with a Windows 10 operating system, an Intel i5-11800H 2.30 GHz processor, and 16GB of RAM.

[0076] Table 1. Main parameters of the dataset

[0077] Table 1 describes the main parameters of the real and synthetic datasets. The Gaoligong Mountain vegetation dataset has 12 features and 13,264 spatial instances, while the synthetic dataset has 13 features and 71,606 spatial instances. Figure 9 The spatial distribution of different feature instances in the two datasets is shown respectively, where (a) is the spatial distribution map of the vegetation data of Gaoligong Mountain and (b) is the spatial distribution map of the synthetic data.

[0078] This invention was compared with two regional pattern mining methods proposed in recent years on two real datasets: the multi-level pattern mining method proposed by Deng et al. in 2017 and the NNMulti-level adaptive regional pattern mining method based on natural domains proposed by Liu et al. in 2021. Since both the multi-level and NNMulti-level methods mine all possible regional isotopic patterns, while the method proposed in this invention only mines frequent kernel patterns containing kernel features (rare alpine vegetation species), this embodiment only lists the frequent regional patterns containing kernel features in the mining results of the two comparative methods.

[0079] In this embodiment, the method proposed in this invention is implemented in three versions under different constraints: the Clique-core method based on clique constraints, the RCPM-based method based on kernel neighbor constraints, and the RCPM-CFI method (the method of this invention). The Clique-core method, along with the two comparison methods, employs clique constraints, and its mining results are mainly affected by different partitioning methods, making it useful for comparing the merits of different partitioning methods. The RCPM-based and RCPM-CFI methods both employ kernel neighbor constraints and are used to compare the mining results with the comparison methods. Furthermore, the RCPM-based and RCPM-CFI methods can be used to compare the impact of two proximity measurement methods—distance threshold-based and kernel neighbor-based—on the results.

[0080] Table 2. Results of regional kernel pattern mining based on kernel feature G in the Gaoligong Mountain vegetation dataset

[0081] ---- indicates no result. Tables 2 and 3 present the mining results of the proposed method on the Gaoligong Mountain vegetation dataset and the synthetic dataset, respectively. The Clique-core method's mining results represent patterns containing kernel features within frequent regional isotopic patterns under clique constraints. The RCPM-based method's mining results represent frequent regional kernel patterns under kernel proximity constraints. The RCPM-CFI method's mining results represent frequent regional kernel patterns based on kernel proximity attribution constraints, where the number after RCPM-CFI indicates the value of the kernel proximity attribution number x. The mining results consist of frequent regional (kernel) patterns and their minimum pattern participation within the region.

[0082] Table 3. Results of region kernel pattern mining based on kernel feature J in synthetic datasets

[0083] ---- indicates no mining results. Due to the excessive size of the synthetic dataset and the inefficiency of the NNMulti-level method, the method failed to obtain mining results even after running for more than 24 hours.

[0084] As can be seen from the mining results in Tables 2 and 3, under the same clique relation constraints, the mining results obtained by the Clique-core method are similar to or slightly less than those obtained by the Multi-level and NNMulti-level methods. The frequent kernel pattern results obtained by the RCPM-based method based on kernel neighbor constraints not only basically include the mining results of the Multi-level and NNMulti-level methods, but also mine more and higher-order frequent kernel patterns. The RCPM-CFI method, on the other hand, is a constraint reinforcement of the RCPM-based method, and can obtain frequent region kernel patterns with stronger dependencies. By specifying the number of x, frequent region kernel patterns with different constraint levels are obtained. As x increases, the mining results of the RCPM-CFI method become closer to those of the RCPM-based method. When x reaches a certain size, the RCPM-CFI method becomes equivalent to the RCPM-based method.

[0085] The mining results in Tables 2 and 3 show that the kernel-based frequent regional patterns mined by both the Multi-level and NNMulti-level methods exhibit high pattern participation, approaching 1. This suggests that the kernel-based frequent regional patterns mined by the two comparative methods may be concentrated in relatively small areas, leading to higher pattern participation. Figure 10 The unique nucleated frequent region patterns mined by the two comparative methods were visualized. (a) shows the existence forms and growth regions of rare alpine vegetation species mined by the Multi-level method, and (b) shows the existence forms and growth regions of rare alpine vegetation species mined by the NNMulti-level method. The results indicate that the frequent region areas of the unique frequent region patterns mined by Multi-level and NNMulti-level methods are indeed relatively small. These small sub-regions may be randomly formed due to noise and do not reflect the true patterns.

[0086] Experimental results also show that the kernel pattern mining method proposed in this invention can find more comprehensive frequent sub-regions for the patterns jointly mined by the comparative method and the kernel pattern mining method proposed in this invention. Figure 11The figures illustrate frequent sub-regions of the shared pattern {J, A} obtained using different methods. Figure (a) shows the distribution of features J and A, where the areas circled in red are frequent regions of pattern {J, A} obtained by the Multi-level method. Figure (b) shows the frequent regions of pattern {J, A} found by the kernel pattern mining method of this invention. Figure 11 It can be seen that for the region kernel pattern {J, A}, the frequent regions obtained by the Multi-level method are only a part of all frequent regions, while the frequent regions obtained by the kernel pattern mining method proposed in this invention are more comprehensive and realistic.

[0087] Furthermore, compared to existing regional pattern mining methods, the kernel pattern mining method proposed in this invention can mine more and higher-order kernel patterns. Figure 12 This paper showcases the unique frequent kernel patterns discovered by the method proposed in this invention. The left image shows the frequent regions of the kernel patterns, and the right image shows all instances of the kernel patterns within those frequent regions. As can be seen from the images, this invention can discover kernel patterns {G, B, C, F} that exhibit a star-shaped topological relationship with the kernel features. Feature G is the kernel feature (rare alpine vegetation species). It is evident that features C, B, and F frequently appear around feature G. This suggests that features C, B, and F are frequently observed due to the influence of feature G, or that feature G is suited to growing near features C, B, and F. Existing regional pattern mining methods can only discover patterns that satisfy clique constraints, ignoring some special cases in practical applications.

[0088] The embodiments focus on the impact of data scale on the efficiency of the proposed method. The NoCore_Hash method is based on kernel proximity constraints and is implemented using a traditional table instance join method (without Core_Hash). To evaluate the efficiency of the proposed Core_Hash structure, this embodiment compares the time consumption of two methods without Core_Hash (Clique-Core method and NoCore_Hash method) with the time consumption of methods improved based on the Core_Hash structure (RCPM-base method and RCPM-CFI method). Figure 13 The figures illustrate the runtime of different methods under varying data scales. Figure (a) shows the impact of the number of instances on method runtime, and Figure (b) shows the impact of the number of core instances on method runtime. Figure 13 The experimental results show that the method based on the Core_Hash structure has significantly higher running efficiency than the basic method, which fully demonstrates the high efficiency of the Core_Hash structure proposed in this invention.

[0089] In addition, this embodiment also tested the impact of other factors on the efficiency of the method. Figure 14 The running times of various methods are shown under different numbers of features and different minimum frequency thresholds, where (a) shows the impact of the number of features on the method efficiency, and (b) shows the impact of the minimum frequency threshold on the method efficiency. Experimental results show that the kernel pattern mining method proposed in this invention still has the best efficiency performance. Among them, the Clique-core method is the most sensitive to changes in the number of features, and its time consumption increases exponentially with the increase in the number of features. The NoCore_Hash method is the second most effective, and its time consumption is positively linearly correlated with the number of features. The time consumption of the kernel pattern mining method proposed in this invention is less affected by changes in the number of features. The same applies to changes in the frequency threshold.

[0090] Based on all the above embodiments, it is sufficient to prove that the kernel pattern mining method proposed in this invention has outstanding performance in both mining results and method efficiency, and is far superior to existing methods.

[0091] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for identifying the existence forms and growth areas of rare alpine vegetation species, characterized in that, Includes the following steps: Step S1: Establish a spatial dataset containing rare alpine vegetation species and obtain the spatial feature set F; Step S2: Filter kernel features in the spatial dataset and divide sub-regions based on the distribution influence of the kernel features; the kernel features are rare alpine vegetation species. Step S3: Mine frequent kernel patterns in each of the divided sub-regions; the kernel pattern c is a k-order subset of the spatial feature set F, where c must contain kernel feature cf, cf ∈ F; the kernel pattern refers to a spatial isotopic pattern containing rare alpine vegetation species. Step S4: Collect frequent kernel patterns in all sub-regions and establish a query dictionary for the existence forms and growth areas of alpine vegetation species; Step S5: Based on the established query dictionary, query the existence forms and growth areas of rare alpine vegetation species. Step S3 includes: Step S31: Construct the storage structure Core_Hash; Step S32: Mining frequent core patterns based on the Core_Hash structure; The Core_Hash storage structure is constructed as follows: For each non-core instance, iterate through all core instances and determine whether the current non-core instance satisfies the proximity relationship with a core instance. If it does, add the corresponding information to Core_Hash. If the key of Core_Hash already contains the current core instance, directly add the non-core instances adjacent to the current core instance to the corresponding feature table. Finally, the Core_Hash structure is obtained. The nuclear instance is an example of a rare alpine vegetation species, and the non-nuclear instance is an example of a non-rare alpine vegetation species. The method for determining whether a non-kernel instance satisfies the proximity relationship with a kernel instance is as follows: Given a maximum number of kernel neighbor associations x, for a non-kernel instance, the non-kernel instance and the x nearest kernel instances satisfy a spatial proximity relationship; if the number of kernel instances exceeds x, the proximity relationship is not satisfied.

2. The method for identifying the existence forms and growth areas of rare alpine vegetation species according to claim 1, characterized in that, Step S1 specifically involves: Establish a spatial dataset, where each vegetation type in the dataset is combined into a spatial feature set. , where n represents the number of feature categories; each feature category contains a set of feature instances distributed at different locations in space; all feature instances constitute the spatial instance set. The i-th spatial instance is represented as , It is described by a triple <feature category, instance number, spatial location>, where the instance number represents the ordinal number of the spatial instance among all instances of the feature category, and the spatial location refers to the spatial location of the spatial instance; the spatial instance... This refers to a single, actually growing alpine plant.

3. The method for identifying the existence forms and growth areas of rare alpine vegetation species according to claim 1, characterized in that, The method for selecting kernel features in step S2 is as follows: Generate box plots of all features in the spatial dataset on the horizontal and vertical distributions. In the box plots, features with narrower quartile intervals or larger median deviations are kernel features.

4. The method for identifying the existence forms and growth areas of rare alpine vegetation species according to claim 1, characterized in that, The sub-region division method in step S2 is as follows: First, calculate the partition distance threshold pd. Then, based on each core instance and the partition distance threshold pd, add all instances that are adjacent to each core instance to the set it forms. Determine whether the set contains other feature instances. If it does not contain other feature instances, it means that there are only core instances in the current region, and the set will be discarded. Finally, merge the sets containing the same instances to obtain the final sub-region partitioning result. The partition distance threshold pd is set to 1.

5. rd to 2 Between rd, where rd is the relative distance between kernel instances: ； Among them, dist_min(co i ) Computational kernel instance The minimum competition distance is defined as the distance to the nearest neighboring core instance. The nuclear example is an instance of a rare alpine vegetation species.

5. The method for identifying the existence forms and growth areas of rare alpine vegetation species according to claim 1, characterized in that, Step S32 specifically involves: First, based on Core_Hash, frequent second-order core patterns are obtained. These frequent second-order core patterns are then combined to generate third-order candidate core patterns. Based on Core_Hash, the core participation degree of all candidate core patterns is calculated. Frequent core patterns are then identified from the candidate core patterns. This process of "generating k+1 order candidate core patterns from k-order frequent core patterns and then determining whether the k+1 order candidate core patterns are frequent" is repeated until all generated candidate core patterns are empty. Finally, all frequent core patterns are obtained. The method for determining whether the kernel mode is frequent is as follows: The nuclear participation degree of the nuclear mode was obtained through calculation. Given a minimum kernel participation threshold min_cprev, if CPI(R, c)≥min_cprev, then kernel pattern c is said to be a frequent kernel pattern in region R, and region R is called the frequent region of kernel pattern c.

6. The method for identifying the existence forms and growth areas of rare alpine vegetation species according to claim 1, characterized in that, Step S4 specifically includes: Step S41: For all the collected frequent kernel patterns, establish a query dictionary structure {kernel pattern: {all frequent sub-regions}}; Step S42: For all the sub-regions divided in step S2, establish another query dictionary structure {sub-region:{all frequent kernel patterns}}.

Images