Forest pest and disease identification method based on data analysis

Abstract

The invention relates to the technical field of data analysis, and further relates to a forest pest and disease identification method based on data analysis, and the method comprises the steps: 1, obtaining an aviation orthoimage of a target forest region and a trapping point count in a set time interval, and associating the trapping count to a sample region unit according to the recent affiliation; 2, a composite simplex complex composed of a plaque layer, a sample area layer and a trapping layer is constructed on a forest map grid, counter examples are screened out through a two-stage verifier, and then suspected pest and disease damage triggering units are output; step 3, establishing a disease and insect pest atlas database, and describing typical modes by disease type topological fingerprints; and carrying out similarity judgment on the topological evidence packet and the disease type topological fingerprint to generate a disease type, a range and a credibility level, and carrying out layering labeling on a forest map grid. According to the method, the fine-grained structure change of diseases and insect pests can be captured in an early stage, and the recognition accuracy and interpretability are improved through multi-source evidence synthesis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data analysis, and particularly relates to a forest pest identification method based on data analysis. BACKGROUND

[0002] The monitoring and identification of forest pests have always been a key link in forestry protection and ecological management. In the prior art, researchers mainly rely on remote sensing images, ground surveys and trapping devices to monitor the occurrence and spread of pests. For example, the traditional remote sensing method obtains infrared, visible light or multispectral images, and uses vegetation index or spectral characteristics to infer the distribution of pests. This method has the advantages of high efficiency and wide coverage in large-scale monitoring, but has limitations in local identification and fine-grained lesion boundary detection. In the early stage of pest occurrence, the tree crown often shows local discoloration, leaf spots or small-scale texture abnormalities. These detailed features are easily ignored or misjudged under insufficient image resolution or complex lighting conditions.

[0003] On the other hand, ground survey methods can directly observe the type and severity of pests, and have high accuracy. However, due to the wide area and complex terrain of forest areas, manual survey is time-consuming and labor-intensive, and cannot achieve high-frequency and full-coverage dynamic monitoring. In addition, the results of manual survey are easily affected by the experience level of the survey personnel, and the data standardization degree is insufficient, making it difficult to form quantifiable and reproducible identification standards. The application of trapping devices provides direct information on the dynamic changes of pest populations. By counting the number of pests in the trap, the activity level of pests in the region can be determined. However, trapping data has spatial limitations and can only represent the conditions within a certain radius around the device. If the layout of the trapping points in the forest area is sparse or uneven, it often leads to insufficient data support, making it difficult to effectively integrate with remote sensing or ground images. In addition, trapping counts are greatly affected by weather, season and environmental factors, and have certain volatility. If only a single count curve is used for judgment, false positives or false negatives may occur. SUMMARY

[0004] The main purpose of the present application is to provide a forest pest identification method based on data analysis, which constructs a forest atlas grid by acquiring aerial orthographic images and trap counts, and forms a composite simplex complex of patch layer, sample area layer and trap layer on this basis, extracts topological features such as connected weakening, hole expansion and expansion persistence combined with scale sequence and guard window; when the key features coexist in the same time window, a backfill bridge edge is established, and further through the abnormal vortex core and trajectory chain, the diffusion process of pests and diseases is revealed, and finally the topological evidence package and the disease type topological fingerprint are matched, and the disease type, range and confidence level are output. The beneficial effects of the method are that it can capture the fine-grained structural changes of pests and diseases at an early stage, improve the identification accuracy and explainability by using multi-source evidence synthesis, at the same time, the stability of the results is guaranteed through neighborhood consistency and cross-time window review, and clear distinction can be achieved among different disease types, thereby providing reliable basis for forest pest monitoring, early warning and precise prevention and control.

[0005] In order to solve the above problems, the technical scheme of the present application is as follows: A forest pest identification method based on data analysis, the method comprising: Step 1: acquiring aerial orthographic images of the target forest area and trap point counts in a set time interval, generating a forest atlas grid according to a fixed grid, dividing sample area units, and generating patch segments in the sample area units, and associating the trap counts to the sample area units according to the nearest attribution; Step 2: constructing a composite simplex complex composed of patch layer, sample area layer and trap layer on the forest atlas grid, forming scale sequences according to color, texture and count and setting guard windows; recording the persistence events of connectedness, hole and expansion along the scale sequence, obtaining the persistence pairs; when patch hole persistence, trap rising and sample area connected weakening coexist in the same guard window, backfill bridge edges are established in the sample area layer; abnormal vortex cores are extracted in space-time according to this, and abnormal vortex trajectory chains are grown, topological evidence packages are synthesized according to the asymmetric weights of connected weakening, hole expansion and trap rising, and after being screened out by two-level verifiers, suspected pest trigger units are output; Step 3: establishing a pest atlas library to describe typical patterns with disease type topological fingerprints; similarity decision is made between the topological evidence package and the disease type topological fingerprint, and disease type, range and confidence level are generated and layered annotated on the forest atlas grid.

[0006] Further, in step 1, an aerial orthophoto of the target forest area and a set of daily counts of the trapping points in the past 14 days are obtained; the aerial orthophoto includes red, green and blue three channels, and the spatial resolution is not less than 20 cm; the target forest area is divided into a plurality of sample unit cells according to a fixed grid, and the grid side length is 10 m; the superpixel segmentation method is used to generate patch fragments in each sample unit cell, and the number of patch fragments in each sample unit cell is limited to between 20 and 60; the trapping count is attributed to the sample unit cell by the nearest trapping point within a distance of 80 m, and if there is no trapping point falling within the range, it is marked as a trapping-free sample unit cell; after geometric registration and color white balance, a forest atlas grid is formed, which includes sample unit cells, patch fragments and trapping point attribution.

[0007] Further, the complex simplex complex in step 2 is composed of three layers of patch layer, sample area layer and trapping layer; the patch layer takes patch fragments as nodes, establishes adjacent edges between mutually contacting patch fragments, and establishes triangular units in the three-patch-fragment common boundary area; the sample area layer takes sample unit cells as nodes, establishes adjacent edges between four adjacent sample unit cells, and establishes four-edge units for any four square regions enclosed by the grid; the trapping layer takes trapping points as nodes, establishes cross-layer association edges with the sample unit cells to which the trapping points belong, and establishes cluster units for the sample unit cell set belonging to the same trapping point; a cross-layer mapping table is established, which records the one-to-one or one-to-many correspondence relationship from the patch layer to the sample area layer and from the sample area layer to the trapping layer.

[0008] Further, the process of forming a scale sequence and setting a guard window in step 2 includes: constructing a scale sequence in the patch layer, the sample area layer and the trapping layer respectively and managing the advance with a unified guard window; the scale sequence of the patch layer is dispersed into 256 levels from low to high according to the green channel mean value of the patch fragment; the scale sequence of the sample area layer is dispersed into 100 levels from low to high according to the texture roughness level of the sample unit cell, and the texture roughness is obtained by gray level co-occurrence matrix statistics; the scale sequence of the trapping layer is dispersed into 20 levels from low to high according to the quantile level of the past 14-day cumulative count; the patch layer, the sample area layer and the trapping layer uniformly adopt a guard window with a length of 5 levels, and only merging, backtracking and bridging operations within the guard window are allowed during scale advance.

[0009] Further, in step 2, the scale sequence of the patch layer, the sample area layer and the trapping layer is synchronously advanced along the scale sequence respectively, the generation and merging events of connected components are recorded, and the connected persistent pair is obtained; the hole appearance and disappearance events on the patch layer and the sample area layer are recorded simultaneously, and the hole persistent pair is obtained; the continuous expansion events of the cluster unit on the trapping layer are recorded, and the expansion persistent pair is obtained; a persistent event list is established for each sample unit cell.

[0010] Further, in step 2, when the same sample unit meets the following three conditions, a backfill bridge edge is introduced in the sample layer to stabilize the topological evidence chain: the first condition is that the patch layer has a hole persistence pair and the persistence level is not lower than 30; the second condition is that the trapping layer appears monotone rising in the adjacent two days and the quantile level is promoted by not less than 2 levels; the third condition is that the sample layer appears a connected weakening event in the current guard window; when the three conditions are met, a backfill bridge edge is established between the node where the connected weakening occurs in the sample layer and the nearest node that is connected and stable in the same guard window, and the backfill bridge edge is retained to the next guard window.

[0011] Further, in step 2, at the end of each guard window, an abnormal vortex core determination is performed on all sample units, and if the following three abnormal determination criteria are met, it is marked as an abnormal vortex core: the first abnormal determination criterion is that the number of holes corresponding to the hole persistence pair of the patch layer is dominant in the neighborhood of the sample unit; the second abnormal determination criterion is that the connectivity of the sample layer decreases by at least one level compared with the last guard window; the third abnormal determination criterion is that the quantile level of the trapping layer is located in the upper half interval of the highest 20 percent of the region; the abnormal vortex core record includes the core sample unit identifier, the corresponding hole index and the trapping quantile level; the trajectory growth is performed in both space and scale directions from the abnormal vortex core as the starting point, and the trajectory growth strategy is: preferentially expanding along the 8-neighborhood in the direction of continued connectivity decline, and secondly expanding along the scale sequence to a higher level; the single expansion step is one neighborhood unit or one scale level; when two trajectories meet within 3 steps, they are merged into one trajectory chain, and if the difference in persistence total of the two trajectories exceeds 10 levels, the one with larger persistence total is retained and the smaller one is marked as a subordinate trajectory.

[0012] Further, in step 2, for each abnormal vortex trajectory chain, three layers of evidence are summarized and synthesized according to asymmetric weights, the connectivity weakening evidence weight is 50, the hole expansion evidence weight is 30, and the trapping rising evidence weight is 20; when the synthesized score is not less than 70 and at least contains two types of evidence, a topological evidence package is generated; the topological evidence package contains the evidence type list, the support layer distribution, the spatial coverage range, the guard window span, the synthesized score and the backfill bridge edge usage times; the topological evidence package of each sample unit is executed by two-level verifiers; the first-level robust consistency verifier requires the simultaneous occurrence of connectivity weakening and hole expansion in the same direction in the last two guard windows; the second-level neighborhood counterexample exclusion verifier requires that the number of dominant features in the 8-neighborhood opposite to the topological evidence package is not more than 1; the sample units that pass through the two-level verifiers are output as suspected pest trigger units, and the corresponding topological evidence package is output together.

[0013] Further, in step 3, a disease and pest atlas library is established, the disease and pest atlas library describes a typical mode by a disease type topological fingerprint, and at least contains three types of disease type topological fingerprints of leaf spot type, branch boring type and canopy wilting type; the disease type topological fingerprint is composed of mandatory items and optional items, the mandatory items at least contain one of connectivity weakening dominance or hole expansion dominance, and the optional items at least contain one of trap rising support or backfill bridge edge multiple triggering; the topological evidence package output in step 2 is compared with the disease and pest atlas library for similarity decision, when all mandatory items and at least one optional item of the corresponding disease type topological fingerprint are satisfied, it is determined that the disease type is matched; the matched disease type is divided into high, medium and low three grades according to the synthetic score, a disease and pest recognition result is generated and is displayed in a layered manner on the forest atlas grid; when adjacent guard windows give different disease types for the same area unit, a safe retreat rule is executed, the one with higher credibility level is used as the reference, and the other result is recorded as a review pending.

[0014] The forest disease and pest recognition method based on data analysis has the following beneficial effects: the method can simultaneously fuse aerial orthographic images and trap counting information in the forest atlas grid, construct a composite simplex complex composed of patch layers, sample area layers and trap layers, and extract persistent features such as connectivity weakening, hole expansion and abnormal vortex through topological data analysis. Compared with the existing method which relies on a single spectral index or an empirical threshold, the beneficial effects of the method lie in significantly improving the accuracy and stability of disease and pest recognition. First, the scale advancing range is limited by the guard window mechanism, the random fluctuation is constrained locally, the determination of persistent events is more stable, and the misjudgment caused by short-term light or counting anomaly is effectively avoided. Second, the backfill bridge edge strategy is used to heal the adjacent units when structure damage and trap rising coexist, solving the problem of false fragmentation caused by shadows, roads or local occlusion, and ensuring the integrity of the spatial continuity of the disease spot. Third, the spatial and scale two-way tracking of the abnormal vortex core and the trajectory chain presents the dynamic process of disease and pest spread in an interpretable trajectory manner, providing a basis for revealing the evolution path of the disease spot. In addition, the evidence synthesis method with asymmetric weights strengthens the dominant position of structural evidence, while taking into account the biological support of trap data, so that the judgment result not only has physical verifiability, but also has ecological rationality. Finally, through the matching of the disease type topological fingerprint, clear distinction can be made between the leaf spot type, branch boring type and canopy wilting type, and fine recognition of different disease and pest types can be realized. The output of the method not only contains the hierarchical and spatial range of the credibility level, but also can execute safe retreat in conflict situations, ensuring the consistency and traceability of the results across the guard windows, thereby providing reliable technical support for the early warning and precise prevention and control of forest diseases and pests. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1A method flowchart of a forest pest identification method based on data analysis provided for an embodiment of the present application is shown in the figure. Figure 2 A hole persistence event statistics analysis diagram provided for an embodiment of the present application is shown in the figure. Figure 3 A backfill bridge edge usage times and guard window span relationship diagram provided for an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0016] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0017] REFERENCE Figure 1 A forest pest identification method based on data analysis, the method comprising: A forest pest identification method based on data analysis, the method comprising: Step 1: Obtain the aerial orthographic image of the target forest area and the trap point count in the set time interval, generate a forest atlas grid according to the fixed grid, divide the sample unit, and generate a patch segment in the sample unit, and associate the trap count to the sample unit according to the nearest attribution; In the specific implementation process, first, the aerial orthographic image of the target forest area is obtained, which contains three channels of red, green and blue, and the ground resolution is not less than 20 cm. Read the geographic reference information of the image, check whether the coordinates and projection are consistent, and avoid deviation in subsequent grid positioning. The reason for requiring not less than 20 cm is that the crown boundary, branch texture and forest road linear structure are not easy to distinguish at a higher resolution, which may cause the patch segment boundary to be excessively smoothed. Obtain the daily count and plane coordinates of the trap points in the set time interval. Check whether the coordinates are consistent with the image coordinate system; if they are not consistent, only perform coordinate system conversion without changing the original count sequence. Keep the time stamp of the daily count, which is required in step 2 to derive the quantile level in time sequence, so it is not combined into a single cumulative value here. Remove the count sequence with serious missing rows (for example, more than 20 missing days in the total number of days in the time interval), to avoid misleading the nearest attribution sample unit in the subsequent step. For the image, check the cloud and shadow coverage; if the cloud and shadow coverage is more than 20% of the area, it is recommended to replace the image with a better time phase to avoid misclassification of the patch segment under a large area of shadow.

[0018] Firstly, the image is spliced to make the seam consistent, and the displacement and brightness mutation at the overlapping joint are eliminated. When splicing, the high-contrast details in the overlapping area are preferred as the alignment basis, because the repeating features in the rich detail area are more stable, which can reduce the misalignment caused by local deformation. The image is cropped to a buffer zone extending 20 meters outside the target forest area boundary. The advantage of setting a buffer is that when the subsequent fixed grid is divided by 10 meters, the boundary grid will not have gaps caused by tight cropping, thus maintaining the continuity and traceability of the sample unit number. If there is a slight difference in ground resolution in different areas of the image, the image is resampled to the same resolution to avoid the same size sample unit containing different number of pixels in different positions, which will affect the stable control of the number of patch fragments.

[0019] The color white balance based on overall statistics is adopted. Specifically, the average brightness of the red, green and blue channels in the entire image is calculated, and the average brightness of the three channels is pulled to the same level. This can reduce the overall color bias caused by the difference in color temperature of the light source during shooting, so that the green color and soil color in different areas have more comparable relative brightness relationship. It is particularly important for patch fragment generation, because superpixel segmentation relies on color and local contrast, and color bias will cause different color clusters of the same tree crown, which will be fragmented. Then generate shadow mask and highlight mask, which specifically include: the shadow area usually has significantly lower brightness and reduced color saturation, and the highlight area (such as gravel road or bare soil) has abnormally high brightness. Mark the shadow and highlight areas, which are only used for subsequent patch fragment boundary constraint and quality marking, and do not change the pixel value at this stage. The reason is that forcibly increasing the brightness of the shadow will introduce noise texture, which will interfere with superpixel aggregation; through the mask, it can avoid excessive segmentation in these areas.

[0020] Within the image space range, a fixed square grid with a side length of 10 meters is established, and sample units are generated row by row and column by column and assigned unique numbers. The advantage of the square grid is that it is naturally aligned with the image pixel grid, reducing coordinate conversion errors, and facilitating the implementation of eight-neighbor consistency check and cross-unit trajectory growth in subsequent step 2. For edge sample units that cross the forest boundary, if the effective forest pixel ratio is less than 50, they are marked as low-effective sample units but the numbers are still retained. This strategy is better than directly discarding, because retaining the number can ensure the integrity of the adjacency relationship in subsequent spatial operations, and the low-effective mark reminds the user to interpret the segmentation results of the unit carefully. Record the pixel count, shadow mask coverage rate and highlight mask coverage rate for each sample unit as the basis for subsequent quality control. When the shadow or highlight coverage rate is higher than 50, only the target number of patch fragments is reduced, and the unit is not directly skipped, so as to avoid forming spatial voids.

[0021] Each sample unit is segmented into superpixels, and the number of target patches is controlled between 20 and 60. The reason for setting the upper and lower limits is that less than 20 will cause different tree crowns to be merged, losing the crown boundary information; more than 60 will generate a large number of fragments in the texture-rich area, increasing the noise density of the subsequent topological events. The specific target number is adaptively determined according to the texture complexity within the unit. The texture complexity can be reflected by the contrast or energy index obtained by the common co-occurrence matrix statistics. When the texture complexity is low (for example, pure canopy and no road insertion), the value close to 20 is taken; when the texture complexity is high (for example, forest land interspersed with forest roads and empty land), the value close to 60 is taken. This adaptation can match the average size of the patch fragment to the scene complexity without changing the interval constraint, reducing over-segmentation or under-segmentation. In the segmentation process, the boundary fitting and small piece merging strategy is enabled. The boundary fitting makes the patch fragment edge close to the real tree crown or road edge by means of color gradient; for small fragments with small area and long shape, if their color is close to the adjacent large patch, they are merged into the adjacent patch. The benefit of this processing is to reduce unstable small fragments, which are easy to introduce false connectivity or false holes in the subsequent connectivity event statistics. If the shadow mask coverage is high, the target number of patch fragments is reduced in the shadow area, and the shape continuity is maintained as the priority. The reason is that the color similarity of the shadow area is poor, and forcibly maintaining a high target number will cut the same tree crown into unnecessary multiple pieces, which will damage the reliability of the subsequent connectivity measurement. After the patch fragments are generated, the number of fragments in each sample unit, the number of pixels of each patch fragment, the average color, and the circumscribed boundary are recorded. For the units with the number of fragments less than 20, re-segmentation is triggered to express the details adequately; for the units with the number of fragments more than 60, adjacent fragments are merged until they return to the interval. The rollback and merging are only performed within the current sample unit, ensuring the independence and repeatability between different units.

[0022] The center of the sample unit is taken as the representative position, and the planar distance between all trapping points is calculated. The nearest trapping point within 80 meters is selected to establish the attribution relationship. When there is no trapping point that meets the conditions, the sample unit is marked as a non-trapping sample unit. The reason for choosing 80 meters is that the representative of common trapping devices under the small-scale diffusion conditions in the forest is mainly within 100 meters. Taking 80 meters can balance the representativeness and spatial resolution; too large a radius will mistakenly attribute sample areas that cross forest roads or forest type mutations to the same trapping point, and too small a radius will result in a large number of units without data support. Spatial indexing is used to accelerate the nearest attribution, so that the global calculation can be completed within an acceptable time. For overlapping or dense trapping points, the nearest attribution naturally forms their respective influence domains, which can avoid assigning a sample unit to multiple counting sources at the same time, thereby maintaining the one-to-one correspondence between the subsequent quantile level and the sample unit. For the counting sequence of the trapping point with consecutive missing data, the original missing label is retained without filling the values. This can explicitly distinguish between "true low activity" and "unobserved" in the subsequent step 2 scale promotion, reducing the risk of misjudging missing data as low activity. A time sequence reference pointer is added to the attributed sample unit, pointing to the daily count within the specified time interval. No averaging or smoothing is performed here to avoid making any transformations that may mask short-term upward trends in step 1, providing complete input for step 2 to calculate the past 14-day cumulative and quantile level.

[0023] A quality label is written for each sample unit, including shadow mask coverage, highlight mask coverage, whether the number of patch fragments is between 20 and 60, and whether the nearest attribution of the trapping point is successfully established. The quality label is used for subsequent interpretation and review of the recognition result. A mapping relationship between the unique number of the sample unit and the geographical location, patch fragment list, and trapping point identifier is established. This mapping relationship ensures that the same input data can reproduce the same forest atlas grid results in different devices and different implementations, which is a prerequisite for subsequent topological data analysis across time.

[0024] After completing the above process, a forest atlas grid is formed, containing three types of basic elements: sample units, patch fragments, and trapping point attribution, as well as their quality labels. This data structure is directly used as the input of step 2 to construct a compound simplex complex, set the scale sequence and guard window, and count connected, hole, and expansion events in the scale promotion. Since step 1 does not perform any smoothing or filling on the count sequence, it can preserve the details of short-term increases, providing a true basis for triggering backfill bridge edges within the same guard window.

[0025] Optionally, the fixed grid can adopt an equal-area hexagonal honeycomb partition, with the side length corresponding to the 10-meter square grid according to the equal-area principle. The hexagon has a better isotropic adjacent structure, which can reduce directional bias in areas with significant terrain undulations. When using a hexagon, the subsequent eight-neighborhood check is replaced by a six-neighbor check that matches the hexagon, and the rest of the process remains unchanged. The patch segment generation can use region growing-based segmentation or watershed-based segmentation. Region growing is suitable for coniferous forests with relatively uniform texture, which can reduce boundary jaggedness; watershed is suitable for areas where roads and forest land are interlaced, which can better disconnect along strong gradient lines. Regardless of which one is used, the target number interval of 20 to 60 and the small piece merging strategy should be retained to maintain consistency with the subsequent topological event statistics.

[0026] Optionally, in forest areas with obvious shadows, a shadow reference-based white balance can be used, that is, the color mean of the open land or bare soil area in the same phase is preferentially counted as a reference, and then the entire image is adjusted. This can more accurately restore the color relationship under light differences and reduce the segmentation drift caused by excessive darkening in the shadow area. If a standard color plate is set up on site, the color plate can be used as a reference. 80 meters is the default value, which can be taken as 60 to 80 in dense coniferous forest areas to reduce the misfit across forest types, and 80 to 100 in sparse broadleaf forest areas to reduce the proportion of units without trapping samples. The trigger for adaptation can be based on the estimated average tree crown diameter of the sample unit or the classification result of the road density index. When using adaptation, the specific radius value needs to be recorded in the output to ensure the interpretability of the subsequent results. For obvious water bodies, bare land, and large roads, a non-forest mask can be generated on the image. Within the mask, no patch segment is generated, only the sample unit number is retained and marked as a non-forest unit. This can avoid generating a large number of meaningless fragments in non-target areas and reduce noise in subsequent topological event statistics. If there are multiple available images within a set time interval, the image with lower cloud shadow coverage and more uniform light can be preferentially selected; when a single image cannot cover the entire area, multiple images of the same day or adjacent days are spliced, and the color and geometry are consistent according to the above splicing and white balance process.

[0027] Step 2: Construct a composite simplex complex composed of patch layer, sample layer, and trapping layer on the forest map grid, form a scale sequence according to color, texture, and count, and set a guard window; record the persistent events of connectivity, hole, and expansion along the scale to obtain the persistent pair; when patch hole persistence, trapping rise, and sample connectivity weaken coexist in the same guard window, establish a backfill bridge edge in the sample layer; accordingly, extract abnormal vortex cores and grow abnormal vortex trajectory chains in space-time, synthesize topological evidence packages according to the asymmetric weights of connectivity weakening, hole expansion, and trapping rise, and output suspected pest trigger units after screening out counterexamples by two-level verifiers; In the implementation process, the input is the forest map grid formed in step 1, which contains sample unit, patch fragment and trap attribution and its quality label. Step 2 takes this input as the basis to complete simplex complex construction, scale sequence and guard window setting, persistent event recording, backfill bridge edge establishment, abnormal vortex core extraction, abnormal vortex trajectory chain growth, asymmetric weight evidence synthesis and two-level verification in turn, and outputs suspected pest trigger unit and topological evidence package.

[0028] Step 2 first takes patch fragments as nodes, and establishes adjacent edges for patch fragments with connected boundaries on the image. The smallest closed area surrounded by three patch fragments is recorded as a triangular unit. The reason for using triangular units is that three connected fragments can easily form the smallest ring of crown layer damage or hole, and recording this structure can sensitively capture the appearance and disappearance of holes in subsequent scale promotion. Taking sample units as nodes, adjacent edges are established between four adjacent sample units. Any four square regions surrounded by grids are recorded as quadrilateral units. Quadrilateral units fit the natural blocking of fixed grids, can stably depict the regional scale changes of crown layer continuity, and are convenient for eight-neighborhood statistics. Cross-layer association edges are established between traps and their attribution sample units; the sample unit set attributed to the same trap is recorded as a cluster unit. Cluster units depict the common influence domain of a trap on surrounding sample units, and this set will reflect continuous expansion events in subsequent counting.

[0029] The mapping table from the patch layer to the sample layer and the sample layer to the trap layer is established, allowing one-to-one and one-to-many relationships. The mapping table is used to refer back to the common sample unit within the layer, so as to judge whether the three types of evidence coexist in the same guard window.

[0030] Along the respective dimension sequences of the three layers, the birth and merging of connected components are recorded by threshold-crossing analysis to form connected persistence pairs. The birth and merging are recorded to measure the "length" of the connected structure, and the greater the length, the more the structure is not dependent on the color difference or noise. In the patch layer and the sample area layer, the appearance and disappearance of holes are monitored based on loop units to form hole persistence pairs. Recording the holes first and then recording the disappearance can accurately describe the duration of the "internal cavity caused by boundary damage", which is a common form of disease spots. In the trapping layer, the spatial coverage of the cluster unit is counted step by step, and any coverage that continuously expands in adjacent levels is recorded as continuous expansion and forms an expansion persistence pair. The continuity of expansion reflects the trend of insect activity from point to plane, which can exclude randomness better than single-day peak.

[0031] In the same guard window, if the same sample area unit meets three conditions, a backfill bridge edge is established: there is a hole persistence pair in the patch layer and the persistence level is not less than 30; there is a monotone rise in the adjacent two days in the trapping layer and the quantile level is improved by not less than 2 levels; there is a connected weakening event in the sample area layer. The purpose of setting the three common occurrences is to only perform structural healing when "structure damage, activity intensification, and regional connectivity deterioration" are all true, to avoid mistaking single-source noise as disease expansion. In the sample area layer, the node where the connected weakening occurs is connected to the nearest node that is stable in connectivity in the same guard window to establish a backfill bridge edge; the priority of the nearest neighbor is to reconnect the same crown layer that may be temporarily cut off by shadows or small roads, so that the subsequent statistics of connectivity and holes are not exaggerated by accidental occlusion. The backfill bridge edge is retained until the next guard window to observe whether it is still needed to heal; if the trigger condition is no longer met in the next guard window, it is automatically removed to avoid long-term distortion of the true topology.

[0032] At the end of each guard window, a determination is made for all sample area units to reduce false triggers caused by repeated fluctuations within the guard window. The following three conditions are met to mark an abnormal vortex core: the number of holes corresponding to the patch layer hole persistence pair is dominant within the eight-neighborhood of the sample area unit; the connectivity of the sample area layer decreases by at least one level compared to the previous guard window; the quantile level of the trapping layer is in the upper half of the top 20 percentile of the region. The design of the three conditions is to include "internal cavity expansion", "regional connectivity degradation", and "biological pressure increase" at the same time, and the simultaneous occurrence of the three is more consistent with the typical link of disease spot generation, which can significantly reduce false positives caused by random textures and accidental count peaks. Record the core sample area unit identifier, corresponding hole index, and trapping quantile level as the starting point and constraint for subsequent trajectory growth.

[0033] Starting from the abnormal vortex core, it expands in both space and scale. It expands preferentially along the eight-neighborhood in the direction of connectivity decline, and secondarily along the corresponding scale sequence to a higher level. The spatial priority ensures that the trajectory first adheres to the real damage propagation path, rather than being deviated by single-point extreme values; the secondary choice of scale increment helps to capture the "shallow to deep" aggravation process. The single expansion step is limited to one neighborhood unit or one scale level. Limiting the step size can prevent jumping over multiple potential inflection points when crossing complex boundaries, thus preserving the interpretability of the trajectory. Stop when encountering a boundary or when two consecutive expansions do not meet the priority condition. When two trajectories meet within 3 steps, they are merged into one trajectory chain; if the difference in total persistence between the two trajectories exceeds 10 levels, the trajectory with more total persistence is retained, and the smaller one is marked as a subordinate trajectory. Close-range merging reduces the risk of repeatedly labeling the same lesion boundary, and the priority of total persistence makes the more stable and wider coverage trajectory the main description of the region.

[0034] Along each abnormal vortex trajectory chain, the coverage of connectivity weakening evidence, hole expansion evidence, and trap rising evidence in three layers is counted respectively. Asymmetric weights are used: the connectivity weakening evidence weight is 50, the hole expansion evidence weight is 30, and the trap rising evidence weight is 20. The weight design reflects that structural signals are less susceptible to short-term weather and sampling accidents than count signals, so they are given higher weight in synthesis. When the synthesis score is not less than 70 and at least contains two types of evidence, a topological evidence package is generated. The double conditions ensure that the results have both strength and multi-source consistency. Record the evidence type list, support layer distribution, spatial coverage range, guard window span, synthesis score, and backfill bridge edge usage frequency. The reason for separately recording the backfill bridge edge usage frequency is that frequent bridging indicates the presence of persistent occlusion or fine cutting structure in this region, which has guiding significance in subsequent review and operation.

[0035] It is required that connectivity weakening and hole expansion in the same direction occur simultaneously within two consecutive guard windows. The design of two consecutive windows is to eliminate one-time isolated events, such as temporary traces left by a mechanical operation. It is required that the count of dominant features within the eight-neighborhood that are opposite to the topological evidence package does not exceed one. The addition of the counterexample count limit is to form a "locally consistent" judgment environment in space, avoiding the amplification of unit-level accidental abnormalities in the entire region. The sample units verified by two levels are output as suspected pest trigger units, and are output together with the corresponding topological evidence package, providing structured input for the pest map library matching in step 3.

[0036] When a sample cell has no trap attribution, the backfill bridge edge is not triggered, but the composition score can still be synthesized according to the two types of evidence, i.e. connectivity weakening and hole expansion. If such a cell passes the two-level verification, it is also output as a suspected pest trigger cell, with a label of "no counting support" for reference in subsequent review. When the quality label shows that the shadow or highlight coverage is higher than 50, the evidence package is allowed to be generated, but with a "quality concern" note attached to the output; meanwhile, the maximum number of steps for trajectory growth within this cell is limited to 1, to prevent large-area low-quality regions from dragging the trajectory off track. If the three trigger conditions are no longer met simultaneously in the subsequent guard window, the backfill bridge edge established in the previous guard window is automatically deleted. The backtracking mechanism ensures that the structure closure only exists when necessary, and does not change the true topology for a long time.

[0037] An event log is maintained for each sample cell, recording the three types of persistent pairs, bridge, core and trajectory operations in the order of guard windows, ensuring that any result can be traced back to a specific scale position and time sequence. The execution order is fixed as follows: construct the complex, set the scale and guard window, record the three types of persistent pairs, establish the backfill bridge edge, extract the abnormal vortex core, merge the trajectory growth, synthesize the evidence, two-level verification, and output. The fixed order avoids discrepancies between different implementations due to inconsistent operation sequences. The levels, percentiles and window lengths appearing in step 2 are fixed values, facilitating cross-batch data comparison; if the optional implementation is changed, the specific values should be written in the output to ensure interpretability.

[0038] Optionally, the color sequence of the patch layer can be changed to a single-channel sequence based on green saliency. This sequence enhances the separability of leaf proportion by emphasizing green and suppressing red-blue differences, suitable for evergreen forests with no obvious seasonal color change. The advantage of using this sequence is that when the lighting conditions change greatly, the tree leaf area still maintains a relatively stable ranking, thereby improving the stability of persistent pairs. The rest of the process and the length of the guard window remain unchanged. The capture of holes can use the "fill difference" strategy: fill the closed regions of the patch layer and the sample layer globally, and then subtract the original region. The difference region is the hole candidate. This strategy is more robust when there is a lot of boundary noise, as it directly measures the "internal gap" and does not rely on the continuity of the fine edges. Holes are still recorded according to their appearance and disappearance.

[0039] Optionally, in the forest transition zone, the trapping sub-rank can use the sliding upper percentile of the historical distribution in the region as the threshold. For example, dynamically determine the boundary of the "top 20 percentile" with the distribution of the last few guard windows. This approach can keep the sensitivity stable when the overall activity level moves up or down, avoiding the over-detection or missed detection caused by fixed thresholds. When there are multiple parallel tracks close to each other, a one-time match can be established between the track ends, replacing the "meet within 3 steps" with the combined condition of "end distance and direction consistency". This processing can reduce false mergers in areas with more parallel structures on roads or forest edges, and retain multiple real advancing fronts. If step 1 uses a hexagonal grid, the eight-neighborhood of this step is replaced by the six-neighborhood; accordingly, the neighborhood statistics in the anomaly vortex core and the two-level neighborhood counterexample exclusion verifier are adjusted to six-neighborhood count. The rest of the thresholds and processes do not need to be changed. When the trapping sub-rank is high in several consecutive guard windows, but there is no significant change in the hole and connectivity, the retention period of the backfill bridge edge can be extended from 1 guard window to 2 guard windows. This can preserve potential connections in the case of "predecessor pressure and subsequent structural damage", and avoid premature removal that breaks the subsequent evidence chain.

[0040] Step 3: Establish a pest map library to describe typical patterns with disease type topological fingerprints; perform similarity decision on topological evidence packages and disease type topological fingerprints to generate disease type, range and confidence level, and layerize annotations on the forest map grid.

[0041] The input of step 3 is the suspected pest trigger unit and its corresponding topological evidence package, anomaly vortex track chain and quality label output by step 2. The goal of step 3 is to: according to the disease type topological fingerprint of the pest map library, make similarity decision on each suspected pest trigger unit, generate disease type, range and confidence level, and complete layerized display on the forest map grid; according to the safe retreat rule, review and save the difference results generated by adjacent guard windows.

[0042] The pest map library stores "disease type topological fingerprints" in units of disease types. Each fingerprint contains mandatory and optional items, and records three types of information: evidence combination, temporal relationship and cross-layer coverage.

[0043] Leaf spot type: essential item is hole expansion dominated; common performance is that multiple small holes appear in the patch layer first and then in the sample area layer; optional items include a small amount of triggering of trap rising support or backfill bridge edge. The consideration of such setting is that leaf disease spots are mostly formed by the aggregation of small area necrosis on the leaf surface, and then the sparse cavities are formed in the fine granularity layer, and then the regional layer is presented. Branch boring type: essential item is connected weakening dominated; common performance is that the sample area layer first appears along the linear path of connected degradation, and the hole in the patch layer can be present or absent; optional items include multiple triggering of backfill bridge edge or trap rising support. The degradation along the linear path conforms to the spatial form of the damaged branch direction, and repeated bridging prompts the fragmentation. Crown wilting type: essential item is connected weakening dominated, and hole expansion is a secondary item; trap rising can not be obvious; common performance is that the sample area layer has a large range of connectivity decline and slow outward expansion. Such fingerprints emphasize the overall degradation of regional scale and weaken the dependence on local small holes.

[0044] Each entry of the fingerprint is derived from historical confirmed samples and expert review, and the entry is fixed once it is admitted, and it is not changed spontaneously in step 3. This ensures the reproducibility across batches and regions, and avoids changing the criteria due to the randomness of individual cases on site. For the first application in a new area, the example description and reference layer can be replaced without changing the structure of the entry, enhancing local adaptability.

[0045] For each suspected disease and pest triggering unit, first filter out all the disease types that meet the essential items completely in the disease and pest atlas library, to form a candidate disease type set. The advantage of filtering the essential items first is to quickly exclude fingerprints with inconsistent directions, such as units with strong hole expansion but almost no connected degradation, which should not be judged as branch boring type first. The candidate disease type set is usually small in size, which is convenient for subsequent detailed similarity decision.

[0046] The evidence types, guard window span, and support layer distribution of the topology evidence package are checked against the fingerprint of the candidate disease type, item by item. The order of checking is: first check if the mandatory items are fully covered, then check if at least one of the optional items is hit. The order of mandatory first and optional second can grasp the lower bound reliability of the identification, and avoid being misled by accidental coincidence of individual optional items. Check if the appearance and expansion direction of the abnormal vortex trajectory chain in the patch layer and the sample area layer are consistent with the time sequence relationship in the fingerprint. The reason for emphasizing the time sequence is that the disease patch often follows a fixed evolution path of "first local, then regional" or "first skeleton, then hole", and only relying on static evidence can easily mistake some construction or short-time shadow as a disease patch. The simultaneous occurrence of statistical evidence in the patch layer, sample area layer, and trapping layer is counted. The higher the proportion of cross-layer simultaneous occurrence, the higher the proportion of the same biological-structural event observed at different scales, rather than accidental events from different sources. The consistency score is obtained by weighted synthesis of evidence consistency, time sequence relationship, and cross-layer coverage, and is normalized to the range of 0 to 100. When weighting, the influence of structure-related items is moderately increased because structure items are more stable than count items and are not sensitive to weather and sampling interval. The score is used for sorting rather than separate determination, and the final result is still subject to the criterion of "mandatory items fully satisfied, at least one optional item hit". The highest scoring candidate disease type in the candidate disease type set is selected as the matching disease type of the suspected disease trigger unit. When the difference between the highest score and the second highest score is small and they hit different optional items, mark them as "parallel candidates" and keep both of them to further distinguish them through spatial context in subsequent range aggregation and neighborhood consistency evaluation.

[0047] The forest map grid is regionally grown according to the matching disease type in unit granularity. The starting set is all suspected disease trigger units with the same matching disease type, and it is expanded outward along the eight-neighborhood, with the expansion condition being that the adjacent unit has the same disease type and the consistency score is not lower than a certain proportion of the starting unit. Using a relative condition rather than an absolute threshold can adapt to the overall contrast difference of different stands, avoiding misjudgment of large areas as low-confidence scattered points when the overall contrast is weak. Expansion stops when the adjacent unit has a different disease type or the consistency score is significantly reduced, obtaining a connected region set. The holes inside the set are then filled once, provided that the hole area is much smaller than the surrounding area and the units inside the hole do not show opposite evidence. The role of filling is to eliminate accidental gaps caused by small shadows or road joints, making the disease patch boundary more in line with natural morphology. The count of opposite disease types in the edge band of each connected region is calculated. When the proportion of opposite count in the edge band is too high, the confidence level of the region is reduced, or the region is divided into multiple smaller sub-regions. This can inhibit the excessive aggregation of "multiple disease type interlaced bands" and avoid treating the contact surface of two different diseases as a single disease patch.

[0048] The matching results are divided into three levels: high, medium and low. High: all mandatory items are met, at least two optional items are met or one item is met and the time sequence relationship is completely consistent, and the neighborhood consistency is good. Medium: all mandatory items are met, at least one optional item is met, and there is a slight deviation in the time sequence relationship or the opposite count ratio of the edge band is moderate. Low: only the mandatory items are met, but the optional items are not hit or the neighborhood consistency is poor. This division method includes "evidence structure", "evolutionary rationality" and "spatial stability" at the same time, avoiding relying on single strong evidence to pull up the overall judgment. For each connected region, a result record is generated, including disease name, region number, range bounding box, area, center coordinates, guard window span, main supporting evidence list and credibility level. The result record is linked one-to-one with the event log in step 2, realizing the whole process traceability from the final result back to the specific guard window and specific evidence.

[0049] The sample area unit is the basic drawing unit, which is colored according to the disease type and the transparency is adjusted according to the credibility level. High credibility is more opaque and low credibility is more transparent, which is convenient for focusing on high credibility areas when browsing in a large range. In the suspected disease and pest trigger unit, representative patch segments of the patch layer are set with point marks to indicate the type of evidence. The point marks do not cover all segments, only representative positions are selected to reduce visual clutter. For each connected region, a boundary line is drawn, and the line type can be distinguished according to the difference in disease type. The boundary layer is used for overlay checking with roads, rivers and other features to assist in understanding the spatial relationship. A structured list is output simultaneously, including the core fields and quality notes of each region (such as "no counting support" and "high shadow coverage"), which is used for archiving and batch statistics.

[0050] When adjacent guard windows give different disease types for the same sample area unit, it is determined as a conflict. The result with higher credibility is saved preferentially, and the other result is marked as pending review. If the credibility levels of the two are the same, the disease type consistent with the previous confirmed result is selected preferentially to maintain the time sequence continuity; at the same time, the backfill bridge edge near the unit is marked as "review attention" to prompt step 2 to focus on checking the bridge rationality in the next round of scale promotion. The results of the retired are kept in complete records but not involved in range aggregation. If the retired disease type is supported by two consecutive guard windows, the retirement is automatically lifted and the range is updated.

[0051] When a unit is labeled as "no-trap zone unit", the "trap up support" option in the optional item is automatically set as indeterminate, and the disease type is not downgraded due to the absence of this optional item; but under the same conditions, the fingerprint that does not rely on the trap optional item is preferred. When the coverage of shadow or highlight is more than 50, the confidence level is at most upgraded to medium, and high is not given. This limitation reminds the user to be cautious about the visual evidence, while retaining the structured output for subsequent on-site review. For edge zone units with an effective forest pixel ratio of less than 50, a "edge unit" note is added to the result list after a successful match, and at least one adjacent non-edge unit is required to support the inclusion of the region during range aggregation.

[0052] Optionally, while keeping the essential and optional item framework unchanged, the disease atlas library can extend the "crown layer striping disease type". This disease type is often found along roads or forest edges, with repeated linear damage. Its fingerprint emphasizes the co-occurrence of directionally consistent connectivity degradation of the sample area layer and a small number of elongated holes in the patch layer. The addition of this disease type helps to separate non-planar disease patches from "linear disturbance" and reduces confusion with branch-feeding types. A seasonal adaptation table is added to each disease type fingerprint, recording the weight tendency of trap support and hole prominence in different seasons. For example, in early spring, trap up often leads structural damage, allowing a medium-confidence match to be completed by relying on connectivity weakening and trap up in the short term without hole expansion; while in late summer, hole expansion is required to be more obvious. Seasonal adaptation can reduce the deviation of the same fingerprint throughout the year.

[0053] Optionally, when multiple discrete clusters of suspected diseases of the same disease type are triggered, aggregation and confidence division can be completed within each cluster first, and then at a higher level, check whether they are connected through narrow corridors. If the corridor is supported by low-confidence units, it is retained as multiple independent regions. This processing can avoid "thin and weak corridors" from incorrectly connecting independent disease patches. For units determined to be "parallel candidates", a delayed decision can be made after range aggregation is completed: if the majority of regions in the eight-neighborhood of the unit are matched to a certain disease type and the confidence level is not lower than medium, the unit is assigned to that disease type; otherwise, the parallel state is maintained and marked for review. Delayed decision making enhances robustness using spatial context. When layering is displayed, the boundary line of high-confidence regions can be thickened and the guard window span is marked at the center of gravity; medium-confidence regions are represented by dashed boundaries; low-confidence regions are only represented by semi-transparent coloring in the grid display layer. Through differential display, the user can focus on more critical areas in one glance.

[0054] A square region with length and width of 40 meters in forest area A is selected as an example region. The ground resolution of aerial orthophoto is set to 0.2 meters per pixel. A forest map grid is established according to a fixed square with a side length of 10 meters, and 16 sample unit cells are divided into 4 by 4. Two trapping points are set at plane coordinates (12, 18) and (37, 5) respectively. The time interval is set to daily count for 14 consecutive days. After geometric homogenization and color white balance of the global pixels, the calculation is performed.

[0055] Let the row and column indexes of the sample unit cell be and , respectively. and . Define the grid side length as , and take the value . The center coordinates of the sample unit cell are , where is the two-dimensional coordinate of the sample unit cell center. Take as the target unit, and get . Define the distance function as , where is the sample unit center coordinate, is the trapping point coordinate. The distances from to the two trapping points are and . The nearest attribution is , and satisfies the nearest attribution radius constraint.

[0056] Let be the count of the attribution trapping point on the day, and take the value sequence . Define the past 14-day cumulative count as , and get . Let be the 14-day cumulative set of all sample unit cells on the same day, and define the empirical quantile as , where is the set cardinality. Define the 20th quantile level as , where is the upward rounding. In this embodiment, the global set is set on the 13th day to make the of the current unit , and get . On the 14th day, the global set is set to make the of the current unit . Accordingly, the adjacent two-day monotonic increase and the cumulative quantile level increase are . Let be the number of patch fragments of the target unit, and get ; The recorded shadow coverage was 0.18 and the highlight coverage was 0.06, both below the quality concern threshold of 0.5.

[0057] Establish a composite simplex complex consisting of patch layer, plot layer and trap layer; in the patch layer, use patch segments as nodes, adjacent segments as edges, and three segments with common boundaries to build triangular units; in the plot layer, use plot units as nodes, four adjacent segments as edges, and four units as quadrilateral units; in the trap layer, use trap points as nodes, build cross-layer associated edges with their affiliated units, and record cluster units for the set of units affiliated with the same trap point. Establish a mapping table from patch layer to plot layer and from plot layer to trap layer. Let is the green channel level of the patch layer, range to ;make is the texture roughness level of the sample area, ranging from to ;make The 20-level quantile level of the trapping layer accumulated over the past 14 days, ranging from to The guard window length is defined as In this example, three consecutive guard windows are selected in the patch layer. .

[0058] make For the sample layer in the guard window The average number of connected components in the window, let For plaque layer in guard window The total area of ​​the holes in the window (in square meters) is averaged within the window, let For the trapping layer The 20-level quantile level of the day. Obtained by suprathreshold connectivity and loop detection ; According to the definition, connected persistent pairs and hole persistent pairs appear, and the holes continue to expand in two adjacent guard windows.

[0059] make The backfill bridging edge is established when the following three conditions are met at the same time: there are persistent pairs of holes in the patch layer and their level is not less than 30; and Established; the sample layer has a connectivity weakening event in the current guard window. According to the value, the hole is to The persistence level is more than 30, And the counts increased monotonically from the 13th to the 14th. This indicates that connectivity weakening is established, so At the sample area level, a backfill bridge edge will be established between the node with weakened connectivity and the nearest node with stable connectivity in the same window, and it will be retained until .make is the number of holes, let is the connectivity level of the sample area. and its eight neighborhoods, Dominant in the eight neighborhoods, Compared Descend at least 1 level, and Located in the upper half of the highest 20 percentile of the region. Satisfy the three criteria and mark is the abnormal vortex core. The spatial step is defined as 1 eight-neighborhood unit and the scale step is 1 level. The starting point is . Expand along the eight neighborhoods in the direction of decreasing connectivity. and 1 step each; then follow the green level by Advance to 1 step each, and get two branches of length 2. The third step is to expand to And it meets the branch from the other side within 3 steps and merges into one abnormal vortex trajectory chain. is the total persistence of the trajectory chain, which is measured as the sum of the persistence levels of each segment. In this example, .

[0060] The connectivity weakening index, hole expansion index and trapping rise index are defined as , , Let the composite score for ,in denote the normalized intensities of connectivity weakening, hole expansion, and trapping rise, respectively. Substituting the values ​​into Considering the abnormal vortex trajectory chain in Keep going and , the cross-window stability gain Count, define , then the cross-window synthesis score is .because And at least two of the three types of evidence are non-zero, generate a topological evidence package. Record the list of evidence types, three-layer distribution, and coverage as , the guard window span is to , the number of times the backfill bridging edge is used is 1.

[0061] make To guard the window The weakening of internal connectivity and the expansion of pores appear in the same direction, indicating that is the counterexample count of the eight-neighborhood. and All satisfied , secondary verification gets . Accordingly, output and its expansion unit is a suspected plant disease trigger unit.

[0062] The plant disease atlas library contains three types of disease topological fingerprints: leaf spot type, branch boring type, and canopy wilting type. According to the essential items, the leaf spot type takes hole expansion dominance as the essential item, the branch boring type takes connected weakening dominance and degradation along the linear path as the essential item, and the canopy wilting type takes regional scale connected decline as the essential item. Since the hole expansion is strong and dominant in the patch layer in this example, the leaf spot type and the canopy wilting type are included in the candidate disease type set. Define the evidence consistency score , the timing consistency score and the cross-layer coverage score , with a value range of 0 to 1. Let the comprehensive score be . For the leaf spot type, according to the essential items and optional items, we get: evidence consistency (hole expansion dominance meets, optional item hits trap rising support), timing consistency (the order of the patch layer and the sample area layer basically meets the requirements), cross-layer coverage (patch layer and sample area layer are significant, and trap layer is a high quantile support). Accordingly, we get . For the canopy wilting type, the regional scale connectedness continues to decline, but the hole expansion weight is low, giving , we get . Therefore, the leaf spot type is selected as the matching disease type. Take all suspected plant disease trigger units that match the leaf spot type as the starting set. Let be the relative expansion threshold, defined as , where is the highest comprehensive score in the starting set. In this example , we get . For the eight-neighborhood comprehensive score not less than and the same disease type, perform region growing, and finally get a connected region composed of 5 sample units, with an area of .

[0063] Let be the proportion of opposite disease types in the region edge band, calculated as . According to the credible level division rule: essential items meet, at least one optional item hits, timing consistency, , the confidence level of the region is determined to be high. Checking the conflict records of adjacent guard windows, no case of the same region unit being judged as different disease types in adjacent guard windows is found, and there is no need to trigger the retreat. If a parallel alternative case occurs, the decision is made according to the higher confidence level or the disease type consistent with the last confirmed result. The output disease type name is leaf spot type, the suspected range covers 5 sample region units, and the highest unit is , the guard window span is to , and the main supporting evidence is the coexistence of hole expansion and connectivity weakening, high position of the trapping subposition, and rising of the adjacent two days, with 1 backfill bridge edge. Three types of layers are generated based on the forest atlas grid as the base map: the grid display layer is colored according to the disease type and the transparency is adjusted according to the confidence level, the tree node label layer is placed at the position of the representative plaque segment, and the disease spot boundary layer draws the outer polygon of the connected region.

[0064] Figure 2 The dynamic change rule of hole persistence pairs in the plaque layer and sample region layer with the advancement of the guard window sequence in the process of forest pest identification is shown. The horizontal axis represents the guard window number, and the vertical axis represents the hole persistence level. The threshold line is marked at the persistence level of 30. From the experimental data, it can be observed that the hole appearance events (solid column chart) in the plaque layer show a significant upward trend in the first to fourth guard windows, with the fourth guard window reaching a peak of 85 hole events. This phenomenon reflects the rapid change in the topological structure of the disease and pest area in this period, with a large number of holes appearing inside the plaque segment. Subsequently, in the fifth to eighth guard windows, the number of hole appearance events gradually declines to 45, 35, 35, and 15, respectively, indicating that the spread of the disease and pests has entered a relatively stable stage. The change trend of hole extinction events (dashed column chart) in the sample region layer is basically consistent with that in the plaque layer, but there is a slight difference in the numerical value. The fourth guard window also reaches the highest value, and then shows a downward trend. It is worth noting that the hole extinction events in the sample region layer decline more slowly after the fifth guard window, which indicates that the topological repair at the sample region scale requires a longer time period. When the plaque layer hole persistence level exceeds the threshold of 30, the system triggers the backfill bridge edge mechanism to stabilize the topological evidence chain. As can be seen from the figure, the values of the third and fourth guard windows are significantly higher than the threshold line, meeting the triggering condition of "the existence of hole persistence pairs in the plaque layer and the persistence level not lower than 30" in the patent claim 6. This spatiotemporal distribution feature provides an important basis for the accurate identification of abnormal vortex cores and verifies the effectiveness of the compound simplex complex in multi-scale disease and pest detection.

[0065] Figure 3A dual-Y-axis design comprehensively demonstrates the inherent correlation between the number of backfill bridge edge uses (scatter plot, left Y-axis) and the guard window span (step line, right Y-axis). The horizontal axis represents the guard window number, reflecting the evolution over time. The number of backfill bridge edge uses exhibits a distinct unimodal distribution, peaking at 27 in the fifth guard window. This peak corresponds to the trigger mechanism described in patent claim 6: when the three conditions of a hole persistence level ≥30 in the patch layer, an increase in the trapping layer quantile level ≥2, and a connectivity weakening event at the plot layer are simultaneously met, the system establishes a backfill bridge edge at the plot layer. From the first to the fifth guard windows, the number of uses rapidly increases from 2 to 27, reflecting the increasing instability of the topological structure during pest and disease spread. After the fifth guard window, the number of backfill bridge edge uses decreases, reaching 5 in the ninth guard window. This indicates that with the full development of the anomalous vortex trajectory chain, the topological structure gradually stabilizes, reducing the need for human intervention. The guard window span increases gradually within the first four windows, from level 1 to level 4, and then remains at level 5. This design ensures that sufficient analysis windows are provided during the rapid spread of pests and diseases, while maintaining a fixed span during the stable period to maintain detection accuracy. The use of trend lines clearly reveals the negative correlation between the two: when the guard window span reaches its maximum value, the demand for backfill bridging edges begins to decline. This phenomenon verifies the design concept of retaining backfill bridging edges until the next guard window in this invention. By dynamically adjusting the life cycle of the bridging edges, the optimal stabilization effect of the topological evidence chain is achieved, laying a solid foundation for the accurate output of suspected pest and disease trigger units.

[0066] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying forest pests and diseases based on data analysis, characterized in that: The method includes: Step 1: Obtain aerial orthophotos of the target forest area and trap counts within a set time interval, generate a forest atlas grid based on a fixed grid, divide the sample plot into units, generate patch fragments within the sample plot units, and associate the trap counts to the sample plot units based on the most recent attribution; Step 2: Construct a composite simplex complex consisting of a patch layer, a sample layer, and a trap layer on the forest atlas grid, forming scale sequences based on color, texture, and count, and setting guard windows. Record persistent events of connectivity, holes, and expansion along the scales to obtain persistent pairs. When persistent patches and holes, increased trapping, and weakened sample connectivity co-occur within the same guard window, establish backfill bridging edges at the sample layer. Based on this, extract anomalous vortex cores in space and time and grow anomalous vortex trajectory chains. A topological evidence package is synthesized according to the asymmetric weights of weakened connectivity, hole expansion, and increased trapping. After filtering out counterexamples through a two-stage verifier, a suspected pest trigger unit is output. Step 3: Establish a pest and disease atlas library and describe typical patterns with pathotype topological fingerprints; conduct similarity judgment between the topological evidence package and the pathotype topological fingerprints, generate pathotypes, ranges and credibility levels, and layer them on the forest map grid.

2. The method for identifying forest pests and diseases based on data analysis according to claim 1, characterized in that: In step 1, obtain an aerial orthophoto of the target forest area and a set of daily counts of trapping points for the past 14 days; the aerial orthophoto includes three channels of red, green and blue, and the spatial resolution is not less than 20 cm; divide the target forest area into several sample units according to a fixed grid with a grid side length of 10 meters; use the superpixel segmentation method to generate patch fragments in each sample unit, and limit the number of patch fragments in each sample unit to between 20 and 60; establish an attribution relationship between the trapping count and the sample unit based on the nearest trapping point with a distance of no more than 80 meters. If no trapping point falls within this range, it is marked as a sample unit with no trapping; after geometric alignment and color white balance, a forest atlas grid is formed, which contains the sample units, patch fragments and trapping point attribution.

3. The method for identifying forest pests and diseases based on data analysis according to claim 2, wherein: The composite simplex complex in step 2 consists of three layers: patch layer, plot layer and trapping layer; wherein, the patch layer uses patch fragments as nodes, establishes adjacent edges between mutually contacting patch fragments, and establishes triangular units in the common boundary area of ​​three patch fragments; the plot layer uses plot units as nodes, establishes adjacent edges between four adjacent plot units, and establishes quadrilateral units for any four square areas forming a grid; the trapping layer uses trapping points as nodes, establishes cross-layer associated edges with the plot units to which they belong, and establishes cluster units for the set of plot units belonging to the same trapping point; establish a cross-layer mapping table, which records the one-to-one or one-to-many correspondence between the patch layer and the plot layer, and the plot layer and the trapping layer.

4. The method for identifying forest pests and diseases based on data analysis according to claim 3, characterized in that: The process of forming scale sequences and setting guard windows in step 2 includes: constructing scale sequences at the patch layer, sample plot layer, and trap layer respectively and advancing them with unified guard window management; the scale sequence of the patch layer is discretized into 256 levels from low to high according to the green channel mean of the patch fragment; the scale sequence of the sample plot layer is discretized into 100 levels from low to high according to the texture roughness level of the sample plot unit, and the texture roughness is obtained through gray-level co-occurrence matrix statistics; the scale sequence of the trap layer is discretized into 20 levels from low to high according to the quantile level of the cumulative counts in the past 14 days; the patch layer, sample plot layer, and trap layer uniformly adopt a guard window of 5 levels in length, and only merging, backoff, and bridging operations are allowed within the guard window during scale advancement.

5. The method for identifying forest pests and diseases based on data analysis according to claim 4, characterized in that: In step 2, the sequences of the patch layer, plot layer, and trap layer are advanced synchronously, and the generation and merging events of connected components are recorded to obtain connected persistent pairs; the appearance and disappearance events of holes are recorded simultaneously on the patch layer and plot layer to obtain persistent pairs of holes; the continuous expansion events of cluster units are recorded on the trap layer to obtain persistent pairs of expansions; and a list of persistent events is established for each plot unit.

6. The method for identifying forest pests and diseases based on data analysis according to claim 5, characterized in that: In step 2, when the same plot unit meets the following three conditions, a backfill bridging edge is introduced in the plot layer to stabilize the topological evidence chain: the first condition is that there are persistent pairs of holes in the patch layer and the persistence level is not less than 30; the second condition is that the trapping layer shows a monotonically increasing trend in two consecutive days and the quantile level increases by not less than 2 levels; the third condition is that a connectivity weakening event occurs in the plot layer within the current guard window; when all three conditions are met, a backfill bridging edge is established between the node with connectivity weakening and the nearest node with stable connectivity in the same guard window in the plot layer, and the backfill bridging edge is retained until the next guard window.

7. The method for identifying forest pests and diseases based on data analysis according to claim 6, characterized in that: In step 2, at the end of each guard window, abnormal vortex core determination is performed on all sample units, and those that meet the following three abnormality determination criteria are marked as abnormal vortex cores: the first abnormality determination criterion is that the number of holes corresponding to the persistent pair of holes in the patch layer is dominant in the neighborhood of the sample unit; The second anomaly discrimination criterion is that the connectivity of the sample area layer decreases by at least one level compared with the previous guard window; the third anomaly discrimination criterion is that the quantile level of the trapping layer is in the upper half of the highest 20 percentile in the region; the abnormal vortex core record includes the core sample area unit identifier, the corresponding hole index and the trapping quantile level; trajectory growth is performed in both space and scale directions starting from the abnormal vortex core, and the trajectory growth strategy is: first expand along the 8-neighborhood in the direction of continued decrease in connectivity, and then expand along the scale sequence to higher levels; the single expansion step is one neighborhood unit or one scale level; when two trajectories meet within 3 steps, they are merged into one trajectory chain. If the difference in the total amount of persistence of the two trajectories exceeds 10 levels, the one with the larger total amount of persistence is retained and the smaller one is marked as a subordinate trajectory.

8. The method for identifying forest pests and diseases based on data analysis according to claim 7, characterized in that: In step 2, evidence for each anomalous vortex trajectory chain is summarized at three levels and synthesized using asymmetric weights, with a weight of 50 for connectivity weakening evidence, 30 for hole expansion evidence, and 20 for entrapment rise evidence. A topological evidence package is generated when the combined score is no less than 70 and contains at least two types of evidence. The topological evidence package includes a list of evidence types, support layer distribution, spatial coverage, guard window span, combined score, and the number of backfill bridge edge usages. A two-level verifier is executed on the topological evidence package of each sample plot unit; the first-level robust consistency verifier requires that connectivity weakening and hole expansion in the same direction appear simultaneously within two consecutive guard windows; the second-level neighborhood counterexample exclusion verifier requires that the number of dominant features opposite to the topological evidence package within the 8-neighborhood is no more than 1; the sample plot unit that passes the two-level verifier is output as a suspected pest and disease trigger unit and is output together with its corresponding topological evidence package.

9. The method for identifying forest pests and diseases based on data analysis according to claim 8, characterized in that: In step 3, a pest and disease atlas is established. The pest and disease atlas describes typical patterns using disease type topological fingerprints, which include at least three types of disease type topological fingerprints: leaf spot type, branch borer type, and canopy wilting type. The disease type topological fingerprint consists of required items and optional items. The required items include at least one of connectivity weakening dominance or hole expansion dominance, and the optional items include at least one of trapping rising support or backfill bridging edge multiple triggering. The topological evidence package output in step 2 is subjected to similarity judgment with the pest and disease atlas. If all required items and at least one optional item of the corresponding disease type topological fingerprint are met, it is determined to be a matching disease type. The matching disease types are divided into three levels of credibility: high, medium and low according to the composite score, and the pest and disease identification results are generated and displayed as a layer on the forest map grid; when adjacent guard windows give different disease types for the same sample unit, the safety retreat rule is executed, the one with the higher credibility level is used, and the other result is recorded as pending for review.

Images