A swarm adaptive segmentation method based on DBSCAN 3D parameter space

The DBSCAN three-dimensional parameter space-based adaptive insect swarm segmentation method solves the problem of imprecise insect swarm segmentation in insect radar monitoring data, realizes fine segmentation and behavioral analysis of insect swarms, and improves the recognition rate.

CN117493919BActive Publication Date: 2026-11-17BEIJING INST OF TECH +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311539004.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-17
Publication Date
2026-11-17
Estimated Expiration
2043-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively segment insect swarms in insect radar monitoring data. Conventional clustering algorithms struggle to handle complex insect swarms that are diverse in shape, prone to sticking together, and have large density differences, resulting in imprecise segmentation.

Method used

The DBSCAN three-dimensional parameter space-based adaptive segmentation method for insect swarms is adopted. By setting the MinPts value and calculating the Eps value, clustering is performed in combination with the DBSCAN algorithm, and edge sparse target removal, swarm fusion and edge expansion optimization are carried out to achieve fine segmentation of the insect swarm.

Benefits of technology

It achieves refined and accurate segmentation of migratory insect swarms in terms of time, altitude, and weight, improving the depth of swarm behavior analysis, with high recognition rate and good segmentation effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117493919B_ABST
    Figure CN117493919B_ABST
Patent Text Reader

Abstract

This invention discloses an adaptive insect swarm segmentation method based on the DBSCAN three-dimensional parameter space. Based on the migration characteristics of insect swarms, this invention sets an empirical parameter MinPts according to the number of migrating insects per minute per day. Then, Eps is calculated based on the density change curve corresponding to MinPts, and the DBSCAN algorithm is used to achieve swarm clustering. Finally, optimization operations such as edge sparsity target removal, swarm merging, and swarm edge expansion are performed on the preliminary clustering results, resulting in good segmentation performance and high recognition rate. Using this invention, precise and accurate segmentation of migrating insect swarms in terms of time, altitude, and weight dimensions can be achieved, contributing to refined analysis of migratory behavior and protecting my country's food security.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of insect radar data processing technology, specifically to an adaptive insect swarm segmentation method based on the DBSCAN three-dimensional parameter space. Background Technology

[0002] Insect radar is an effective tool for studying migratory insects. However, traditional insect radars, limited by their resolution and performance, cannot precisely observe the exact location of each insect, hindering in-depth analysis and processing of the observation data. The Ku-band high-resolution fully polarimetric radar developed by the Radar Technology Research Institute of Beijing Institute of Technology has a range resolution of 0.2 meters, enabling precise measurement of the location of each insect. Simultaneously, employing a mature insect biological parameter inversion algorithm, it can accurately retrieve insect biological parameters such as body length, weight, and wingbeat frequency. The Radar Technology Research Institute of Beijing Institute of Technology has deployed this radar in Dongying and conducted long-term operational monitoring, accumulating a large amount of observational data.

[0003] When processing operational radar monitoring data, it was found that insects often take off, land, and cruise in swarms. Furthermore, many insect swarms exhibit aggregation based on weight. Using insect swarms as the research subject and analyzing insect migration patterns can lead to a deeper understanding of the mechanisms of insect migration, enabling the control of migratory insect pests and protecting my country's food security.

[0004] Currently, there is no method for insect swarm segmentation. However, the daily spatiotemporal distribution of insect swarms is complex, with diverse shapes, a tendency to aggregate, and significant density differences between swarms. Conventional clustering algorithms struggle to achieve fine-grained segmentation of these complex swarms. Therefore, there is an urgent need to develop a fine-grained insect swarm segmentation method based on high-resolution, fully polarized operational monitoring data. Summary of the Invention

[0005] In view of this, the present invention provides an adaptive segmentation method for insect swarms based on the DBSCAN three-dimensional parameter space, which can accurately and finely segment migratory insect swarms in terms of time, altitude, and weight, and helps to realize the behavioral analysis of migratory insect swarms of different weights.

[0006] The swarm adaptive segmentation method based on DBSCAN three-dimensional parameter space of the present invention includes:

[0007] Step 1, set the MinPts value:

[0008] Calculate the number of migratory insects per minute (dense) within the observation period, and set the MinPts value based on the dense value; the larger the dense value, the larger the set MinPts value.

[0009] Step 2, calculate the Eps value:

[0010] S21, normalize the time, height and weight of the observed insect data to obtain normalized observation data;

[0011] S22. For each observation data point after normalization, calculate the distance to the MinPts-1 nearest point to that observation data point, and reorder them to obtain the density change curve of the observation dataset.

[0012] S23, perform first-order difference on the density change curve, then smooth it, and then perform maximization to obtain the density change curve difference curve.

[0013] S24, calculate the mean μ and standard deviation σ of the data in the horizontal axis of the density change curve difference curve from 5% to 95%, and truncate the difference curve based on the threshold μ+3σ;

[0014] S25, select multiple ordinate values ​​as Eps from the truncated curve; where, the larger the threshold μ+3σ, the ordinate value that is earlier in the position is selected as Eps.

[0015] Step 3: Use the DBSCAN algorithm to cluster the normalized dataset based on MinPts and different Eps;

[0016] Step four: Optimize the clustering results from step three:

[0017] S41, Remove sparse targets at the edges: Calculate the degree of intersection p between the lowest density cluster and other clusters in the clustering results of step three. If there is p > 0.1, delete the lowest density cluster.

[0018] S42, Swarm merging: If the degree of intersection p of two subgroups is greater than 0.25 and the difference in the normalized average weight of the two subgroups does not exceed 0.2, then the two subgroups are merged.

[0019] S43, expands the edge of the swarm;

[0020] This completes the division of the insect swarm.

[0021] Ideally, if density > 10, the dataset is considered a normal dataset, and the MinPts value ranges from 3 to 8; if density ≤ 10, the dataset is considered a small dataset, and the MinPts value is 4.

[0022] A better approach is to select the MinPts value based on Table 1:

[0023] Table 1 MinPts Values

[0024] Number of migratory insects per minute MinPts value 10 < density ≤ 33 3 33 < dense ≤ 44 4 44 < density ≤ 55 5 55 < density ≤ 75 6 75 < density ≤ 85 7 dense>85 8 .

[0025] For a typical dataset, the following formula is used to calculate the coordinates y1 to y4 of the four positions respectively.

[0026] y1=-822572.1077042sin(x1-π)+41137.31469561(x1-10) 2 -4113730.937462 (3)

[0027] y2=-1071252.087596sin(x1-π)+53574.23663182(x1-10) 2 -5357422.994716 (4)

[0028] y3=-1252777.210748sin(x1-π)+62652.87926073(x1-10) 2 -6265287.124613 (5)

[0029] y4=-1529545.587819sin(x1-π)+76494.60189513(x1-10) 2 -7649459.226047 (6)

[0030] Where x1 is the threshold μ+3σ; the position coordinates y1~y4 are Eps;

[0031] For small datasets, set 2 Eps; where the position coordinate y1′ is a fixed value of 0.36; the position coordinate y2′ is calculated using the following formula:

[0032] y2′=-861553.1657650366sin(x1-π)+43087.89774577574(x1-10) 2 -4308788.999986974 (7)

[0033] Where x1 is the threshold μ+3σ; the position coordinates y1′ and y2′ are Eps.

[0034] Preferably, in step S41, if density > 43, the intersection degree p between the second lowest density cluster and other clusters in the clustering results of step three is further calculated. If the intersection degree with other clusters exceeds 0.25, the second lowest density cluster is deleted.

[0035] Better, determine whether it is a columnar swarm based on the density of the central part of the swarm. For columnar swarms 1 and 2, if the following conditions are met: (1) the degree of intersection between swarm 1 and swarm 2 is greater than 0.1; (2) the height of 90% of the targets in swarm 1 is greater than that in swarm 2; (3) the maximum, minimum, upper and lower quartiles and median of the weight of swarm 1 and 2 do not differ by more than 0.1; (4) the time mean of the smaller swarm falls within the time range of the larger swarm; then swarm 1 and 2 are merged.

[0036] In S42, if density ≤ 10, then the degree of cluster intersection is higher. The two subgroups were then merged.

[0037] In a better approach, in S43, starting from the edge of the swarm, we sequentially search for ungrouped points within the range of i×mean(Eps) from the edge of the swarm, i=1,2,3, and classify them into the swarm closest to them.

[0038] Preferably, in step two, the maximum-minimum normalization method is used for normalization.

[0039] Beneficial effects:

[0040] This invention is based on the characteristics of insect swarm migration. It sets an empirical parameter MinPts according to the number of migrating insects per unit minute per day. Then, it calculates Eps based on the density change curve corresponding to MinPts, and uses the DBSCAN algorithm to achieve insect swarm clustering. Finally, it performs optimization operations such as edge sparse target removal, insect swarm merging, and insect swarm edge expansion on the preliminary clustering results, and obtains good segmentation effect and high recognition rate. Attached Figure Description

[0041] Figure 1 shows the density variation curves and their difference plots: Figure 1(a) density variation curve; Figure 1(b) density variation curve difference plot (before smoothing); Figure 1(c) density variation curve difference plot (after smoothing).

[0042] Figure 2 shows the density change curve and its difference plot under threshold extraction: Figure 2(a) is the difference plot of the density change curve (after smoothing); Figure 2(b) is the density change curve.

[0043] Figure 3 shows the Eps values ​​and fitting formulas: Figure 3(a) shows the Eps values ​​and fitting formulas for a normal dataset; Figure 3(b) shows the Eps values ​​and fitting formulas for a small dataset.

[0044] Figure 4 shows the results of each step of the grouping process: Figure 4(a) is the original scatter plot of the insect time, height and weight distribution on July 10, 2021; Figure 4(b) is the preliminary grouping result; Figure 4(c) is the result of removing sparse targets at the edge; Figure 4(d) is the result of insect fusion; Figure 4(e) is the final result.

[0045] Figure 5 This is a flowchart of the present invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] This invention provides an adaptive swarm segmentation method based on the DBSCAN three-dimensional parameter space.

[0048] The DBSCAN algorithm is a classic density-based clustering algorithm that can cluster into clusters of various shapes, making it well-suited for extracting swarms of different shapes. This algorithm estimates the density of a region by calculating the number of points (including itself) within a given radius Eps of a given point. A point is called a core point if its number of points is not less than a set MinPts. A point within the Eps neighborhood of a core point but with a neighborhood number less than MinPts is called an edge point. Points with a neighborhood number less than MinPts and not within the Eps neighborhood of any other core point are called noise points. By traversing all points in the dataset, all core points, edge points, and noise points are identified. Adjacent core points and edge points are considered to belong to the same cluster, thus completing the clustering segmentation.

[0049] This invention first sets MinPts based on the insect migration rate per minute, then plots a density change curve based on MinPts, calculates the change threshold by performing first-order difference on the density change curve, calculates Eps according to an empirical formula, and uses the DBSCAN algorithm to complete preliminary clustering. Finally, it optimizes the segmentation results through sparse target removal and insect swarm fusion. The flowchart of this invention is shown below. Figure 5 As shown, the details are as follows:

[0050] Step 1: Set MinPts.

[0051] Assuming the number of insects observed in a day is n, and the number of minutes with insects present from the start time to the end time of observation is m, then the number of migrating insects per unit minute is:

[0052]

[0053] The higher the number of migrating insects per minute, the higher the insect density for that day.

[0054] When density > 10, the dataset is considered a normal dataset; otherwise, it is considered a small dataset. For normal datasets, the number of migrating insects observed per minute varies greatly across different dates. High density often represents a large data volume, and the larger the data volume, the larger the MinPts needs to measure the density level of each point. Therefore, as density increases, MinPts should also increase, but MinPts cannot increase indefinitely. A MinPts value that is too large may cause some small insect populations to be overlooked and is not conducive to fine-grained segmentation of insect population edges. For small datasets, on the one hand, a smaller MinPts is needed to measure the density level of each point; on the other hand, using too small a MinPts will result in very small segmented insect populations, leading to fragmentation.

[0055] Based on the above principles and combined with the results of 30 days of data debugging, this embodiment has developed the following empirical value MinPts table:

[0056] Table 1 MinPts Values

[0057]

[0058]

[0059] For small datasets, set MinPts to 4.

[0060] Step 2: Calculate the Eps value.

[0061] Migratory insect swarms exhibit clustering characteristics across time, altitude, and weight dimensions. This invention performs fine-grained swarm segmentation based on these three dimensions. Due to the significant differences in the values ​​for time, altitude, and weight, this invention first normalizes the three-dimensional data. In this embodiment, maximum-minimum normalization is used:

[0062]

[0063] W i W represents a certain parameter. min W max These are the minimum and maximum values ​​of the parameter for that day, respectively.

[0064] Given a minimum value (MinPts), this invention uses observation data from July 19, 2021 as an example. It calculates and sorts the distance from each point in the dataset to its nearest point at its MinPts-1 (here, we take 2, excluding itself) digit, and plots the result in Figure 1(a). This figure can be interpreted as a trend in the density of the dataset. As the horizontal axis increases, the distance from each point to its nearest MinPts-1 digit increases, indicating that the density near that point is decreasing. Performing a first-order difference and maximization yields the density change rate curve, as shown in Figure 1(b). Due to the complex density structure of the dataset, the curve exhibits spikes as shown in the red circles. Therefore, this invention smooths the first-order difference before maximization, resulting in a relatively smooth curve, as shown in Figure 1(c).

[0065] As can be seen from the difference plot, the rate of change in dataset density first decreases, then remains at a long minimum, and finally increases rapidly. Clearly, the long minimum period is the most common occurrence.

[0066] Therefore, the goal is to find an adaptive threshold on the smoothed difference curve, and values ​​below this threshold are considered to be "the density change rate is at a minimum over a long period".

[0067] By testing data over multiple days, it was found that the difference curve was at a relatively low level between 5% and 95% on the horizontal axis. However, due to the asymmetry of the difference curve, the value at 95% was larger than that at 5%. Therefore, the mean μ and standard deviation σ of this portion were calculated, and a threshold μ+3σ was set to extract the difference curve. This threshold ensures that most difference values ​​between 5% and 95% are less than it, and the slowly changing density portion defined by it is generally positioned higher. The density corresponding to this portion is higher and should be included in the clustering. Figures 2(a) and (b) show the density change curves extracted using the threshold, where red represents the slowly changing portion and green dots represent the portions at 5% and 95% on the horizontal axis.

[0068] Then, for a typical dataset, given MinPts, suitable ordinates were selected as Eps from the slowly changing portion of the density change curve. Since different parts of a typical dataset exhibit significant density variations, multiple density gradients are needed to ensure the segmentation of all swarms. That is, a fixed MinPts corresponds to multiple Eps to achieve different density gradients. Because the threshold μ+3σ represents the rate of density change in the dataset, a larger threshold indicates a faster rate of density decrease. Therefore, it is even more important to select the ordinates from the earlier positions in the slowly changing portion of the density change curve as Eps, as the density in this region will be relatively higher, and it should be more likely to be segmented into swarms in the dataset; that is, the position parameter should be smaller. However, it is obviously not advisable to decrease it indefinitely; ultimately, a portion of the points in the dataset will need to be segmented into swarms.

[0069] Based on the above principles, after debugging data for several days, 23 days of cases with relatively complex density distributions were finally accumulated. These cases required four density gradients to ensure the accuracy of clustering. If the dataset density distribution is relatively simple, the rationality of clustering can be improved later through some optimization measures. This invention statistically determined the position of the required EPS in the slowly changing part and the threshold of the dataset for that day.

[0070] Table 2

[0071] date Position 1 Position 2 Position 3 Position 4 threshold 2021.6.15 0.516 0.645 0.775 0.931 0.000070182 2021.6.20 0.5206 0.652 0.7834 0.9411 0.00006817 2021.6.21 0.482 0.602 0.722 0.866 0.00026765 2021.6.27 0.455 0.5674 0.6798 0.8147 0.00037684 2021.7.09 0.5068 0.6336 0.7604 0.9125 0.00015143 2021.7.14 0.5134 0.6422 0.771 0.9256 0.00013658 2021.7.18 0.511 0.638 0.766 0.92 0.00014824 2021.7.19 0.471 0.588 0.704 0.844 0.00034652 2021.7.22 0.454 0.562 0.67 0.799 0.00051562 2021.7.23 0.365 0.445 0.52 0.6 0.0012 2021.7.24 0.462 0.576 0.689 0.825 0.00042037 2021.7.25 0.448 0.556 0.664 0.793 0.00055647 2021.7.26 0.364 0.433 0.5025 0.586 0.0028 2021.8.19 0.495 0.62 0.743 0.891 0.00020985 2021.8.29 0.5154 0.6464 0.7742 0.9296 0.00010289 2021.9.09 0.481 0.6 0.719 0.861 0.00032508 2021.9.12 0.37 0.45 0.54 0.65 0.0015 2021.9.15 0.4645 0.5776 0.6907 0.8264 0.0005948 2021.9.16 0.449 0.559 0.668 0.800 0.00060779 2021.9.17 0.502 0.628 0.754 0.904 0.00018766 2021.9.28 0.438 0.534 0.648 0.775 0.00065578 2021.10.03 0.508 0.636 0.763 0.916 0.00016469 2021.10.09 0.511 0.647 0.766 0.919 0.00013893

[0072] This invention proposes that there may be a certain functional relationship between position parameters and thresholds. Therefore, this invention uses positions 1, 2, 3, and 4 to fit empirical formulas with the thresholds respectively:

[0073] y1=-822572.1077042sin(x1-π)+41137.31469561(x1-10) 2 -4113730.937462 (3)

[0074] y2=-1071252.087596sin(x1-π)+53574.23663182(x1-10) 2 -5357422.994716 (4)

[0075] y3=-1252777.210748sin(x1-π)+62652.87926073x1-10) 2 -6265287.124613 (5)

[0076] y4=-1529545.587819sin(x1-π)+76494.60189513(x1-10) 2 -7649459.226047 (6)

[0077] Where x1 represents the threshold, y i These are the position parameters. The fitted curve and the original point position are shown in Figure 3(a). The position coordinates y1 to y4 are the corresponding 4 Eps.

[0078] Small datasets have a small total number of samples and a relatively simple density distribution. This invention selects two density gradients and statistically analyzes the positions of the required Eps for four datasets in the slowly changing regions, as well as the threshold for the dataset on that day.

[0079] Table 3

[0080] date Position 1 Position 2 threshold 2021.8.11 0.3441 0.6453 0.00071089 2021.8.12 0.3871 0.7233 0.00027011 2021.8.13 0.3536 0.6368 0.0008468 2021.8.16 0.35 0.55 0.0017

[0081] Position 2 also yielded the same empirical formula for density:

[0082] y2′=-861553.1657650366sin(x1-π)+43087.89774577574(x1-10) 2 -4308788.999986974 (7)

[0083] The fitted image is shown in Figure 3(b). Position 1 is around 0.36, therefore y1′ is set to 0.36. The position coordinates y1′ and y2′ are the corresponding two Eps.

[0084] Combining steps one and two, calculate the parameters MinPts and Eps required by the DBSCAN algorithm.

[0085] Step 3, cluster optimization.

[0086] The DBSCAN algorithm is used to cluster the normalized dataset based on MinPts and different Eps, while removing excessively small clusters. However, the clustering results still have some issues and room for optimization. First, the initial clustering produces many sparse colonies, which may be independent of a low-density region or adjacent to dense colonies. In the latter case, because the density of the sparse colonies is too low, they cannot be considered a single colony relative to the surrounding dense colonies. Therefore, this invention defines such sparse colonies as marginal sparse targets, and their existence affects the clustering quality. Furthermore, some groups are divided into one, two, or more groups due to uneven internal density distribution. This also affects the clustering quality. Finally, the edges of some colonies after initial clustering are relatively rough and not rounded enough.

[0087] Therefore, modules such as edge sparse target removal, swarm fusion, and swarm expansion were designed to optimize the initial swarming results. Before that, the degree of intersection p between swarms was defined as follows:

[0088] Using swarm 1 as a reference, calculate the k-value of a point Q in swarm 1. hb Neighborhood (i.e., the k nearest to this point) hb Take a point, If this k hb If more than 10% of the points Q do not belong to swarm 1, and some of these non-swarm 1 points belong to swarm 2, then Q is considered an "intersection point" between swarm 1 and swarm 2. Count the number n "intersection points" between swarm 1 and swarm 2. 12 Calculate the degree of overlap p between swarms 1 and 2.

[0089]

[0090] Where m1 and m2 are the numbers of swarm 1 and swarm 2, respectively.

[0091] For typical datasets, sparse targets at the edges are generally generated by the lowest density clustering. The significance of this density gradient is to detect insect populations far from high-density regions; these populations should have little or no overlap with other high-density populations. Therefore, the degree of overlap, p, between the lowest density clustering result and the surrounding high-density populations is calculated. If p > 0.1, the result is deleted. This preserves some low-density populations far from high-density populations.

[0092] Furthermore, for high-density datasets with density > 43, the second-lowest density gradient clustering also generates some marginally sparse targets. These targets generally maintain a high degree of intersection with multiple high-density swarms. Therefore, the intersection degree between the second-lowest density clustering results and the surrounding high-density swarms is calculated, and the overall inter-swarm intersection degree is statistically analyzed to remove marginally sparse targets. In practical engineering, when the intersection degree between a swarm and multiple swarms exceeds 0.25, the swarm is removed. Some swarms are too fragmented, more like appendages of surrounding swarms. Therefore, based on the inter-swarm intersection degree, this invention proposes a swarm fusion strategy.

[0093] For a normal dataset, this invention calculates the degree of overlap between insect groups. When p > 0.25 and the difference in the normalized average weight of two insect groups does not exceed 0.2, the two insect groups are merged. In the later testing process, columnar insect groups were observed on some days, which were divided into multiple groups due to differences in internal density. These columnar insect groups extended from low altitude to high altitude in a certain period of time. Merging multiple small columnar insect groups into a large columnar insect group may be beneficial for exploring the migration / landing mechanism of insects. For columnar insect groups 1 and 2, the following conditions must be met: (1) The degree of overlap between insect group 1 and insect group 2 is greater than 0.1. (2) 90% of the targets in insect group 1 are at a higher altitude than insect group 2. (3) The maximum, minimum, upper and lower quartiles and median weights of insect groups 1 and 2 do not differ by more than 0.1. (4) The time mean of the small insect group falls within the time range of the large insect group. Then insect groups 1 and 2 are merged.

[0094] Initial clustering of small datasets often results in fragmented swarms, with most swarms being relatively small. To address this issue, this invention relaxes the criteria for swarm fusion, allowing for fusion based on the degree of overlap between swarms. The two swarms merged at that time.

[0095] The current segmentation results have relatively coarse edges, so an expansion operation is performed on the swarm edges. Starting from the swarm edge point, we find unsegmented points that are less than a certain upper limit away from the edge point and classify them into the swarms closest to them. This is equivalent to drawing a circle with the edge point as the center and a radius, which makes the swarm edges smoother.

[0096] In practical applications, this invention finds ungrouped points within a range of i×mean(Eps), (i=1,2,3) from the edge of the swarm, and ultimately achieves good results.

[0097] Example:

[0098] To verify the implementation steps and effectiveness of the algorithm, the clustering process is illustrated using operational monitoring data from July 10, 2021. First, the insect time-altitude-weight distribution for that day is shown in Figure 4(a). The number of migrating insects per minute that day was 57. Following step one, the three-dimensional data was normalized, and MinPts was set to 6. Next, following step two, the distance from each point to its fifth nearest point was calculated, and the corresponding Eps values ​​were calculated using the density empirical formula: 0.01358, 0.01670, 0.02071, and 0.02761. Then, multiple clustering operations were performed based on the MinPts and Eps parameters to obtain preliminary clustering results, as shown in Figure 4(b). Next, clustering optimization was performed according to step three. Figure 4(c) shows the result of removing sparse targets at the edges, Figure 4(d) shows the result of swarm fusion, and Figure 4(e) shows the final result after swarm expansion.

[0099] There is no clear boundary or definition for a swarm. This invention aims to identify swarms as clusters or clumps, with a swarm density significantly higher than the density of surrounding data points. Densities within the swarm can vary, but the differences should not be excessive. Finally, the vast majority of high-density areas in the dataset should be identified as swarms.

[0100] Based on this standard, this invention can use CloudCompare software to manually label the insect colony in the dataset. Then, the cluster purity is used: (w1,w2...w k This represents clusters of insects, including ungrouped insects, c1, c2...c j This represents the correctly labeled insect colonies and the unlabeled adult insect colonies, where N represents the total number of insects on that day. The quality of each day's grouping is evaluated.

[0101] Cluster purity greater than 80% is defined as good results. Taking data from 29 days of observation in October 2021 as an example, the results showed that 24 days were good, accounting for 82.8%.

[0102] In summary, the above are merely preferred embodiments of the present invention and are 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 should be included within the scope of protection of the present invention.

Claims

1. A swarm adaptive segmentation method based on DBSCAN three-dimensional parameter space, characterized in that, include: Step 1, set the MinPts value: Calculate the number of migratory insects (dense) per minute within the observation period, and set the MinPts value based on the dense value; The larger the dense value, the larger the set MinPts value; Step 2, calculate the Eps value: S21, normalize the time, height and weight of the observed insect data to obtain normalized observation data; S22. For each observation data point after normalization, calculate the distance to the MinPts-1 nearest point to that observation data point, and reorder them to obtain the density change curve of the observation dataset. S23, perform first-order difference on the density change curve, then smooth it, and then perform maximization to obtain the density change curve difference curve. S24, calculate the mean μ and standard deviation σ of the data in the horizontal axis of the density change curve difference curve from 5% to 95%, and truncate the difference curve based on the threshold μ+3σ; S25, select multiple ordinate values ​​as Eps from the truncated curve; where, the larger the threshold μ+3σ, the ordinate value that is earlier in the position is selected as Eps. Step 3: Use the DBSCAN algorithm to cluster the normalized dataset based on MinPts and different Eps; Step four: Optimize the clustering results from step three: S41, Remove sparse targets at the edges: Calculate the degree of intersection p between the lowest density cluster and other clusters in the clustering results of step three. If there is p > 0.1, delete the lowest density cluster. S42, Swarm merging: If the degree of intersection p of two subgroups is greater than 0.25 and the difference in the normalized average weight of the two subgroups does not exceed 0.2, then the two subgroups are merged. S43, expands the edge of the swarm; This completes the division of the insect swarm.

2. The method as described in claim 1, characterized in that, If density > 10, the dataset is considered a normal dataset, and the MinPts value ranges from 3 to 8; if density ≤ 10, the dataset is considered a small dataset, and the MinPts value is 4.

3. The method as described in claim 2, characterized in that, Based on Table 1, the MinPts value is determined as follows: Table 1 MinPts values Number of migratory insects per minute MinPts value 10 < density ≤ 33 3 33 < dense ≤ 44 4 44 < density ≤ 55 5 55 < density ≤ 75 6 75 < density ≤ 85 7 dense>85 8 。 4. The method as described in claim 2 or 3, characterized in that, For a typical dataset, the following formula is used to calculate the four position coordinates y1 to y4 respectively; y1=-822572.1077042sin(x1-π)+41137.31469561(x1-10) 2 -4113730.937462 (3) y2=-1071252.087596sin(x1-π)+53574.23663182(x1-10) 2 -5357422.994716 (4) y3=-1252777.210748sin(x1-π)+62652.87926073(x1-10) 2 -6265287.124613 (5) y4=-1529545.587819sin(x1-π)+76494.60189513(x1-10) 2 -7649459.226047 (6) Where x1 is the threshold μ+3σ; the position coordinates y1~y4 are Eps; For small datasets, set 2 Eps; where the position coordinate y1′ is a fixed value of 0.36; the position coordinate y2′ is calculated using the following formula: y′2=-861553.1657650366sin(x1-π)+43087.89774577574(x1-10) 2 -4308788.999986974 (7) Where x1 is the threshold μ+3σ; the position coordinates y1′ and y2′ are Eps.

5. The method as described in claim 1, characterized in that, In step S41, if density > 43, the intersection degree p between the second lowest density cluster and other clusters in the clustering results of step three is further calculated. If the intersection degree with other clusters exceeds 0.25, the second lowest density cluster is deleted.

6. The method as described in claim 1 or 5, characterized in that, Based on the density of the central part of the swarm, determine whether it is a columnar swarm. For columnar swarms 1 and 2, if the following conditions are met: (1) the degree of intersection between swarm 1 and swarm 2 is greater than 0.1; (2) the height of 90% of the targets in swarm 1 is greater than that in swarm 2; (3) the maximum, minimum, upper and lower quartiles and median of the weight of swarm 1 and 2 do not differ by more than 0.1; (4) the time mean of the smaller swarm falls within the time range of the larger swarm; then merge swarm 1 and 2.

7. The method as described in claim 1 or 5, characterized in that, In S42, if density ≤ 10, then the degree of cluster intersection is... The two subgroups were then merged.

8. The method as described in claim 1, characterized in that, In S43, starting from the edge of the swarm, we search for ungrouped points within the range of i×mean(Eps) from the edge of the swarm, i=1,2,3, and classify them into the swarm closest to it.

9. The method as described in claim 1, characterized in that, In step two, the minimum-maximum normalization method is used for normalization.

Citation Information

Patent Citations

  • High-definition multidirectional coordination insect migration radar measuring instrument

    CN106501802A

  • Image characteristic density peak searching-based image segmentation method, system and medium

    CN108629783A