Method and system for automatic identification and positioning of small pests based on multi-modal macro imaging

Abstract

The application discloses a small-sized pest automatic identification and positioning method and system based on multi-modal micro-distance imaging, relates to the technical field of pest monitoring, and comprises the following steps: generating a stable image orthographic image and a pixel-millimeter conversion table in a single frame by using an intelligent attractant card, reading a fingerprint of a reader, a near-attractant temperature and a relative value of volatile intensity, and forming an effective sampling range priori; detecting and constructing an instance signature in the stable image coordinate block, matching across time frames, only counting the number of new individuals and saving the new small pictures and matching records; estimating a standardized capture index by combining the number of new individuals with a release rate estimation and an effective sampling range estimation, and rolling quantile normalization; triggering an alarm based on the cross-line determination and the trend determination and superimposing the same group control; and switching the sampling gear and packing the alarm event stream in an uplink according to the risk in a side-cloud cooperative manner.

Description

Technical Field

[0001] This invention relates to the field of pest monitoring technology, specifically to a method and system for automatic identification and location of small pests based on multimodal macro imaging. Background Technology

[0002] Monitoring of bark beetles in woodlands typically relies on traps such as multi-funnel, window-type, or cross-shaped traps containing pheromone attractants. Individuals are guided by the pheromone plume and fall into collection cups, where their numbers are recorded manually or by camera. While remote monitoring using cameras and algorithms has emerged in recent years, it primarily involves collecting single-frame images daily or weekly and reporting total counts. On-site monitoring is affected by factors such as terrain obstruction, weak netting, diurnal temperature variations and condensation, reflection, and dirt, leading to fluctuations in image quality and geometric stability. Furthermore, differences in trap equipment and deployment (attractant type and filling volume, release time, hanging height and orientation, location at the forest edge or within the forest) and meteorological conditions (temperature, wind, radiation, and humidity) systematically alter the entry rate. Specifically, the attractant release rate increases with rising temperature and decreases over time; wind and canopy shape plume morphology and coverage; and forest edge effects alter encounter probabilities. The combined effect of these factors results in raw counts obtained at different sites and at different times not being on the same scale, often exhibiting significant fluctuations and lacking comparability across different locations and time periods.

[0003] Existing camera-based systems often rely on total cup counts or manual verification, making it difficult to obtain time series data for newly entered individuals. On one hand, individuals within the cup can drift and overlap due to solution disturbance, wind, or movement, and single-frame differential analysis can easily misinterpret displacement as additions or losses. On the other hand, dense stacking and posture damage obscure detection and segmentation boundaries, leading to both duplicate and missed counts. Long-term outdoor operation causes optical and geometric drift, and many devices lack stable pixel-to-millimeter conversion and synchronized recording of equipment metadata, making it impossible to perform scale-consistent physical quantity inferences or interpret cross-device differences. Furthermore, some research or product threshold alerts are directly based on raw counts or fixed zoning thresholds, failing to standardize for inducer release intensity, effective sampling space, and microclimate / deployment differences. This results in inconsistent meanings of the same threshold at different sites, easily leading to false alarms and missed alarms.

[0004] Limited power and communication resources in forest areas lead to high costs associated with frequent transmission and manual verification of high-resolution full-map data, while also limiting network density and response speed. If these problems persist, management will struggle to identify peak flight periods and spatial hotspots in a timely manner, potentially missing early intervention windows and causing a decline in timber value. It could also result in unnecessary logging and transportation due to misjudgments, leading to direct economic and ecological consequences. In pheromone-trapping camera monitoring scenarios, the lack of new individual time-series triggers that can both eliminate intra-cup displacement / overlap interference and standardize the release rate of the attractant and differences in deployment / meteorological conditions results in incomparable counts, unstable early warning systems, and high maintenance costs. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for automatic identification and localization of small pests based on multimodal macro imaging. This includes stable image coordinate block detection and instance signature construction; cross-time frame matching that only counts newly added individuals and saves newly added small images and matching records; obtaining a standardized capture index by combining the number of newly added individuals with release rate estimation and effective sampling range estimation, and then rolling quantile normalization; triggering alarms based on boundary crossing and trend determination, and superimposing co-group comparisons; and edge-cloud collaboration that switches sampling levels and packages alarm event streams according to risk. This method achieves comparable counting, reduced false counting, and more stable early warning, thereby solving the technical problems described in the background art.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for automatic identification and localization of small pests based on multimodal macro imaging includes,

[0010] Deploy smart lure cards, read the device fingerprint and near lure temperature, generate stable orthophotos and pixel-to-millimeter conversion tables according to the coded calibration pattern, and combine them with site metadata to form a priori effective sampling range;

[0011] Perform target detection and instance segmentation on the stabilized orthophoto, construct instance signatures, perform time-frame matching at unified coordinates, count unmatched individuals as new individuals, obtain the number of new individuals, and save the new small images and matching records.

[0012] Release rate was estimated based on near-attractant temperature and number of days since release. Effective sampling range was estimated based on prior information on effective sampling range, wind direction and speed and forest edge information. Standardized capture index was obtained by dividing the number of new individuals by the above two quantities.

[0013] Cross-line determination and trend determination are performed under the standardized capture index scale. The index of neighboring stations is weighted by wind direction and distance to form a group comparison and screen alarms. Based on the determination, the sampling level is issued and the uplink is packaged and alarm records are generated.

[0014] Furthermore, the smart attractant card simultaneously carries an coded calibration pattern, equipment fingerprint identification, proximity temperature sensor, color temperature scale, and gas-sensitive element or passive sampling sheet; the camera completes reading and generates a pixel-to-millimeter conversion table, equipment fingerprint field table, and deployment days within a single frame, and forms a stable image homography and timestamp association record locally, while binding the relative values ​​of proximity attractant temperature and volatility intensity with site metadata into the database.

[0015] Furthermore, the cup cavity is equipped with cross-polarized ring lighting, a matte black liner, and a white positioning ring at the cup mouth, along with an anti-fog coating and a drainage structure. After generating a stable orthophoto, an effective sampling range prior is formed based on the forest edge distance, wind direction and speed, and canopy density, and updated with each frame. A key is established with the pixel-to-millimeter conversion table for unified access. At the same time, the lighting and orientation configuration are recorded in the equipment fingerprint field table.

[0016] Furthermore, a fixed-overlapping block inference output candidate volume mask is used on the stabilized orthophoto; when constructing the instance signature, at least a color histogram, texture response, shape description and millimeter volume scale are included and written in a fixed field order.

[0017] Cross-time frame matching generates matching records and newly added individual small images based on weighted costs, and binds them to a pixel-to-millimeter conversion table. The block boundaries and overlap widths are fixed at the site level and recorded with the version.

[0018] Furthermore, quality scores are calculated for three items: defocus, glare, and dirt. Frames with scores below the threshold are not included in the increment. Candidates with adhesion or unclosed boundaries are placed in a rejection queue and local small images are exported.

[0019] Targets within the occlusion reconnection window are reconnected without being counted as new, and the quality score is bound to the frame timestamp for archiving. The rejection queue is sorted by priority and the instance signature field is updated and reviewed in the cloud.

[0020] Furthermore, the release rate is estimated by combining the near-attractant temperature, the number of days since release, and the relative value of volatility intensity; the effective sampling range is corrected by combining the effective sampling range prior with forest edge distance, wind direction and speed, and canopy density; both are timestamped for recalculation and written into the parameter record table with a unified identifier for cross-station alignment, and are used as the denominator of the same frame for reference in the standardized capture index, and are linked to the release date, trap model, and orientation in the equipment fingerprint field table.

[0021] Furthermore, after constructing the standardized capture index, the median and upper and lower quantile bands are generated by frame-by-frame recursion and archived together with the standardized capture index. The quantile bands serve as the reference baseline for boundary crossing and trend determination, and an index relationship is established with the newly added individual number and parameter record table. When archiving, the site number and timestamp are written to maintain consistent reference with the equipment fingerprint field table and the number of deployment days.

[0022] Furthermore, the threshold for determining the boundary is formed by extrapolating the upper edge of the quantile zone and the bandwidth, and the trend determination is formed by recursively extrapolating the slope of the exponential weighting to form a Boolean indicator. After the two are combined to generate candidate alarms, the index of the neighboring stations is aggregated according to wind direction and distance to form a group comparison and a deviation ratio threshold is set to generate candidate alarm records with station numbers and timestamps. At the same time, newly added individual small map references are retained for auditing.

[0023] Furthermore, a risk score is obtained by weighting three types of Boolean indicators: exceeding the limit, trend, and deviation from the same group comparison. The risk score is then mapped to a deterministic switch between the edge-side sampling level and the uplink packaging method.

[0024] The platform generates alarm event streams, operation and maintenance instructions and work orders in sync, and associates the work orders with alarm records and assigns numbers. At the same time, it records sampling level switching logs and uplink packet lists on the edge side.

[0025] A system for automatic identification and location of small pests based on multimodal macro imaging includes,

[0026] The calibration and acquisition module is equipped with smart attractant cards, which read the fingerprints of the equipment and the near-attractant temperature. It generates a stable orthophoto and a pixel-to-millimeter conversion table according to the coded calibration pattern, and combines it with site metadata to form a priori effective sampling range.

[0027] The deduplication module performs target detection and instance segmentation on the stabilized orthophoto, constructs instance signatures, performs cross-time frame matching at unified coordinates, counts unmatched individuals as new additions, obtains the number of new individuals, and saves the new small images and matching records.

[0028] The index calculation module estimates the release rate based on near-attractant temperature and number of days since release, and estimates the effective sampling range based on prior information on effective sampling range, wind direction and speed and forest edge information. The standardized capture index is obtained by dividing the number of new individuals by the above two quantities.

[0029] The early warning linkage module performs boundary crossing and trend determination under the standardized capture index scale. It uses the index of neighboring stations to form a group comparison by weighting the index by wind direction and distance, and filters alarms. Based on the determination, it sends the sampling level and uplink package and generates alarm records.

[0030] (III) Beneficial Effects

[0031] This invention provides a method and system for automatic identification and localization of small pests based on multimodal macro imaging, which has the following beneficial effects:

[0032] With the support of the smart decoy card, the system acquires stable orthophotos, pixel-to-millimeter conversion tables, equipment fingerprints, relative values ​​of near-decoy temperature and volatilization intensity in the same frame, and generates a priori data of effective sampling range. This enables unified archiving of geometric scale and decoy status, reduces inconsistencies caused by installation drift and on-site differences, and forms a replayable physical benchmark and site-level input.

[0033] Block detection and instance segmentation are performed within the stable image coordinates. Instance signatures are constructed and cross-time frame matching is performed. Combined with occlusion reconnection and rejection queue, only newly entered individuals are counted to alleviate the duplication and undercounting of total difference. At the same time, small images of newly added individuals and matching records are output to facilitate subsequent verification and data backflow training.

[0034] A standardized capture index is constructed by estimating the number of new individuals and the release rate, as well as the effective sampling range. A quantile band rolling normalization is used to unify the counting scale under different sites, seasons, deployments, and wind field conditions, supporting subsequent judgments based on a single trigger quantity and reducing instability caused by human threshold shifting.

[0035] By combining boundary crossing and trend determination, and introducing a group comparison weighted by wind direction and distance, candidate alarms are screened to separate regional synchronous fluctuations from local anomalies, maintaining the certainty of alarm entry points; and by connecting with fixed fields of alarm event streams and work orders, a continuous link from determination to execution is ensured.

[0036] By driving edge-cloud collaboration through risk scoring, sampling levels and uplink packaging are switched deterministically. In high-risk segments, only newly added small images and statistical summaries are uploaded, while in low-risk segments, heartbeats and periodic summaries are maintained. This reduces the transmission and power supply pressure under weak network and energy constraints, and improves the sustainability and manageability of long-term deployments. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the method for automatic identification and location of small pests according to the present invention;

[0038] Figure 2 This is a schematic diagram of the system structure for automatic identification and location of small pests according to the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 This invention provides a method for automatic identification and location of small pests based on multimodal macro imaging, including:

[0041] Step 1: Based on the smart decoy card, generate a stable orthophoto image and a pixel-to-millimeter conversion table. Read the relative values ​​of equipment fingerprint, near-decoy temperature and volatilization intensity, and combine them with site metadata to form a priori valid sampling range, providing traceable input for subsequent instance-level incremental counting and unified scaling calculation. At the same time, establish a correlation between stable homography, pixel-to-millimeter conversion table, equipment fingerprint field table and timestamp to ensure a consistent entry point for subsequent cross-frame comparison, cross-site calls and audit recalculation.

[0042] Insert the smart attractant card into the positioning slot of the attractant holder, ensuring that the card's coding and marking patterns are unobstructed and the white positioning ring at the cup rim is fully visible. Secure the camera to the mounting hole and tighten it. Initiate a calibration acquisition using the buttons on the front panel of the device. The system will acquire one frame of image inside the cup in a predetermined sequence, read the graphic identification code on the card, and record the attractant model, batch, release date, trap model, and orientation. Simultaneously, it will acquire the near-attractant temperature from the temperature sensor close to the attractant sac and read the evaporation signal from the gas-sensitive element in the small diffusion area of ​​the cup cavity.

[0043] After data acquisition, the device locally displays that the pixel-to-millimeter conversion table has been generated, the equipment fingerprint field table has been recorded, and the near-inducer temperature and volatility intensity have been entered into the database. The stabilized orthophoto image and the prior value of the effective sampling range are simultaneously displayed on the maintenance terminal page. The operator signs for the calibration record on the terminal, completing the handover. If calibration fails, the system rolls back to the previously verified pixel-to-millimeter conversion table and equipment fingerprint field table, and prompts the user to clean the card surface again before re-acquiring data.

[0044] After prolonged outdoor operation, the lens focal length, mounting posture, and position of the cup-based camera will experience slight drift. If the geometric relationships are not reconstructed in each acquisition, pixel-level measurements cannot be mapped to the millimeter scale, resulting in unstable data for inter-period comparisons. Therefore, it is necessary to solve the homography relationship with a visible, fixed graphic on each frame and generate a pixel-to-millimeter conversion table to give the target's boundaries, length, and position physical meaning.

[0045] Within the same frame, the homography relationship between the cup rim plane and the camera imaging plane is first solved using the coded calibration pattern to obtain the transformation from pixel to physical plane; then, the pixel-to-millimeter scale coefficient is calculated using the known card surface geometry and corner coordinates, and a conversion table that can be directly used for subsequent body size and trajectory measurement is output.

[0046] A corner point chessboard is arranged on the card surface, and the world coordinates of the corner points remain fixed within the card surface plane. During acquisition, the pixel coordinates of the corner points are identified, and the homography matrix is ​​solved jointly using least squares and anti-outspot strategies to form a bidirectional mapping from the physical plane to the pixel plane. This mapping is then used for intra-frame image stabilization and orthorectification, where:

[0047]

[0048] Where: pixel x-coordinate Pixel ordinate Non-negative integers; scale factor A value greater than zero is used for homogeneous coordinate normalization.

[0049] Pixel-to-millimeter homography matrix ,for A matrix, with real numbers as elements, representing the projective transformation from the physical plane to the pixel plane; the physical x-axis... Physical vertical axis The unit is millimeters, and the range is limited by the size of the outer rectangle of the coding and calibration pattern.

[0050] The pixel coordinates are obtained by corner detection. Combined with the known physical coordinates of the corners on the card, the initial value of the homography matrix is ​​obtained by singular value decomposition. A robust estimator is used to suppress the influence of dirty corners. Then, the entire frame image is orthorectified according to this transformation to generate a stable orthorectified image and a pixel-to-millimeter conversion table.

[0051] In application, homography pulls the cup rim image of each frame back to the same geometric reference plane, eliminating drift and slight tilt; the stabilized edges and contours have a consistent direction, which can be directly used for inter-period matching; the pixel-to-millimeter conversion table is updated with each frame, making millimeter-level measurements repeatable.

[0052] Even with only homography, dimensional calibration is still required. The actual measured side length of the card surface must be used as the scale anchor point to directly obtain the pixel-to-millimeter ratio and generate a conversion table across the entire frame, as follows:

[0053]

[0054] Where: pixel-to-millimeter scale factor , is a positive number; physical coordinates of the two corner points Given by the card's geometry design; pixel coordinates of the two corner points. It is obtained by corner detection.

[0055] On the stabilized orthophoto image, adjacent corner point pairs are selected, and the scale coefficient is calculated using the above formula. A pixel-to-millimeter conversion table is generated for the entire frame, containing the millimeter length and area conversion factor corresponding to each pixel position. The conversion table and the homography matrix are archived together with timestamps. In application, the scale coefficient unifies image measurements to the millimeter dimension, and the calculation of volume length, spacing, and area has direct physical meaning. The conversion table can be directly called in step two, avoiding repeated derivation. The timestamped conversion table facilitates auditing and traceability.

[0056] The intensity of pheromone release is clearly related to its near-end temperature and the number of days since release; the evaporation path and plume coverage are also affected by wind and boundaries. If the attractant state is not directly read from the acquisition frame and a priori valid sampling range is not generated, any subsequent normalization and cross-site comparisons will lack executable input.

[0057] Therefore, by fusing near-attractant temperature and color temperature scale readings within the same frame, a definite temperature value is obtained; then, the volatile signal is read to generate a relative value of volatile intensity, and the effective sampling range prior is calculated using forest edge distance, wind speed, and wind direction, so that the attractant state and spatial coverage are quantified on the same time scale.

[0058] Temperature probes provide instantaneous temperature, while color temperature scales provide redundant information. Both may be affected by sludge or condensation, necessitating the establishment of a confident fusion temperature at the frame level, along with the number of days since application, to create a complete inducer status, as detailed below:

[0059]

[0060] Where: near-attractant temperature The output is used for subsequent release rate estimation; temperature probe temperature Temperature is read by a proximity sensor; color temperature scale The unit is Celsius, obtained from color mapping;

[0061] Color temperature scale A quadratic polynomial is used to establish the mapping relationship between RGB channel values ​​and temperature, as shown in the following formula:

[0062]

[0063] Color-coded temperature scale; Normalized channel values ​​( Factory calibration constants; fusion weights Dimensionless, taking values ​​in The result is calculated based on color saturation, contrast, and sensor self-test results.

[0064] First, the color temperature scale is obtained by mapping the tristimulus values ​​of the color scale on the card face using a calibrated polynomial. Then take the temperature probe reading. The near-inducing temperature is obtained according to the above formula. Simultaneously, the number of days for deployment is calculated using the deployment date from the equipment fingerprint field table and the timestamp of the current frame. And put into storage, number of days of release This serves as the input for calculating the release rate in the subsequent step three.

[0065] When applied, temperature fusion generates usable near-attractor temperatures within a single frame, resisting temporary distortions from a single channel; the storage of release days gives subsequent release rate calculations a time dimension, and the attractant status becomes a traceable sequence.

[0066] The volatile signal reflects the instantaneous release background of the attractant capsule, but there are differences in zero point and span between different devices; at the same time, spatial cover is related to forest edge distance, wind speed and wind direction, and a usable prior needs to be obtained at the frame level, where:

[0067]

[0068] Where: the effective sampling range prior Used for subsequent normalization; reference area This provides the baseline coverage for the installation point under windless, open conditions; attenuation coefficient. A value greater than 0 represents the intensity of cover attenuation due to forest edge distance; forest edge distance Wind gain coefficient, measured by on-site distance measurement or electronic map. The unit is per meter per second, and the value is greater than 0, representing the linear gain of wind speed on coverage; wind speed Read by anemometer; relative wind direction angle The unit is radians, representing the angle between the wind direction and the front of the trap, with values ​​ranging from... to .

[0069]

[0070] Relative value of volatility intensity After normalization, the value range is [0, 1].

[0071] raw readings The original detection values ​​(such as voltage and grayscale value) of the gas-sensitive element or passive sampling plate;

[0072] midnight and span : Real number, a fixed constant specified at the factory ( This is the original value when there is no volatilization. (The difference between the original value at full-scale evaporation); Clipping function Defined as This ensures that the output value falls within the range [0, 1].

[0073] The relative value of volatile intensity is obtained by performing a linear normalization on the original signal of the gas-sensitive element with the factory zero point and span constant. The data is archived frame by frame; the effective sampling range is calculated using the above formula, taking into account the distance to the forest edge, wind speed, and wind direction. If the wind direction is away from the decoy, because A value of zero indicates that the a priori value immediately reflects the weakened coverage.

[0074] In application, the relative value of volatile intensity compresses the differences between devices to a uniform scale, serving as an auxiliary record of the release state; the effective sampling range is a priori. Spatial coverage is expressed as a closed expression, which is available with each frame and provides an entry point for spatial scale for subsequent normalization and threshold determination.

[0075] Step 2: Construct instance signatures through block detection and instance segmentation, perform cross-temporal frame matching, and combine occlusion reconnection windows and quality gating. Unmatched instances are counted as new additions, and new individual small images and matching records are output simultaneously as reliable molecules for subsequent standardized capture indices. At the same time, the new addition list is indexed with stabilized orthophotos and pixel-to-millimeter conversion tables to ensure a unified entry point and consistent format for cross-batch retrieval, verification reflow, and training sample acquisition.

[0076] For scenarios involving densely packed, easily drifting small pests within cups, the system selects the current site and selects "This Acquisition" on the maintenance terminal. The device retrieves the previously archived stabilized orthophoto image, pixel-to-millimeter conversion table, and equipment fingerprint field table, and then acquires the current frame. The system first displays a preview of the stabilized orthophoto image on the terminal page, then automatically performs single-frame target detection and instance segmentation, backfilling the bounding box, mask, and body size of each candidate. Subsequently, it calls cross-time frame matching within the same geometric coordinates to generate the assignment results for the previous and current frames. Unassigned candidates are displayed as thumbnails in the newly added individuals list. The terminal simultaneously generates the number of newly added individuals in this period, a list of newly added individual local thumbnails, cross-frame matching records (including lists of matched and unmatched pairs), and a rejection queue (requiring manual confirmation).

[0077] Click "Confirm Archive" on the terminal, and the system will store the above deliverables in the warehouse and send them back to the central platform. If the matching or quality gating fails, the system will switch to the keyframe sampling + manual review priority queue mode, retaining the original image of the current frame and the candidate object thumbnail, waiting for subsequent review.

[0078] The background inside the cup is composed of a matte black liner and reflections from the cup wall. The insects are small in scale, high in density, and exhibit varied postures. Direct detection and measurement at the original pixel coordinates would be highly susceptible to minute variations in installation posture and lighting differences. Therefore, it is necessary to first pull the current frame back to a stabilized orthophoto image, then perform block detection and mask generation within this geometric reference frame, using a pixel-to-millimeter conversion table as the scale anchor point to unify the bounding box and body size to millimeters. Subsequently, an instance signature containing both appearance and geometry is constructed for each candidate, providing a discriminative identity for cross-time frame matching.

[0079] First, orthophotos the current frame is unfolded according to the image stabilization homography. Then, block inference is performed at a uniform scale to ensure that small targets can be separated. Next, the pixel length and area are converted to the millimeter scale using a pixel-to-millimeter conversion table. Then, color and texture statistics, contour shape and volume scale elements are extracted on the basis of the mask, assembled into an instance signature, and output together with the mask and bounding box.

[0080] In the current frame, the archived image stabilization homography matrix and pixel-to-millimeter conversion table are invoked to expand the cup-shaped region to the same geometric reference frame as the previous frame. To ensure the separability of dense targets, a fixed-overlapping block partitioning strategy is adopted to avoid boundary effects caused by targets being segmented across blocks along their long sides. For each candidate object, its size is converted according to the pixel-to-millimeter conversion table to obtain its length and area description in millimeters. To ensure a unified size variable for subsequent link calls, the physical length is defined as the product of the pixel length and the pixel-to-millimeter scale coefficient.

[0081]

[0082] Where: physical length , indicating time , No. Length of each candidate object along the principal axis; pixel-to-millimeter scale factor The unit is millimeters per pixel, derived from the pixel-to-millimeter conversion table in step one; pixel length. The unit pixel is calculated from the principal axis of the candidate volume mask.

[0083] In the stabilized orthophoto image, blocks are formed in row priority order, with fixed block sizes and overlapping widths between blocks within a preset range. For each candidate volume mask, the principal axis direction and pixel length are calculated, and the physical length is obtained using a formula. And store the instance signature.

[0084] In application, through image stabilization and scale anchoring, the geometry and size of the candidate object fall within a unified physical dimension, avoiding measurement fluctuations caused by slight changes in installation posture and focal length; block processing ensures that small targets at the edges are completely included, and the geometric consistency of subsequent matching is guaranteed.

[0085] Within each candidate body mask, color histograms, texture filter responses, and shape moment descriptions are extracted, and length, area, aspect ratio, and skeleton segment lengths in millimeter dimensions are incorporated into the instance signature. Instance signatures are concatenated in a fixed order to maintain comparability across frames. To ensure a stable numerical range for appearance similarity in subsequent matching, high-dimensional appearance vectors are normalized to units of a sphere, and geometric quantities are linearly normalized at the millimeter scale, allowing different dimensions to contribute controllably to the weighted matching cost.

[0086] Therefore, local small images are cropped from the stabilized orthophoto image according to the mask, the color histogram and texture response are calculated, and the instance signature vector is generated by combining the geometric quantities at the millimeter scale; the mask and bounding box are saved as materials for subsequent matching and local small image export.

[0087] In application, instance signatures couple the three elements of appearance, geometry, and volume scale within the same coordinate system, providing a distinguishable identity for cross-time frame assignment; unit sphere and linear normalization suppress numerical dominance between different dimensions, improving the stability of subsequent cost functions.

[0088] Insects inside the cup may shift slightly due to solution disturbance and wind-induced movement, and simply calculating the total number of insects in the cup may misjudge the displacement as new additions. Therefore, it is necessary to establish an assignment cost within the stable image coordinates, constrained by appearance similarity, spatial overlap, and motion consistency, and solve the matching between the previous frame and the current frame based on this. At the same time, a quality gating function should be established at the frame level to shield out-of-focus and highlight frames, avoiding the identification of stains and reflections as new additions. Adhesive bodies that cannot be distinguished should be placed in a rejection queue, reviewed by the cloud, and then fed back.

[0089] First, a matching cost matrix consisting of overlap, appearance cosine distance, and center displacement penalty is constructed, and then assignment is performed. Candidates that are not matched and are not within the occlusion reconnection window are judged as new individuals, and candidates with uncertain adhesion are marked as rejected. Finally, at the frame level, sharpness, glare ratio, and stain coverage are scored, and frames below the threshold are not included in the increment.

[0090] In the image-stabilized coordinate system, the candidate object sets from the previous and current frames already possess unified masks, boundaries, and instance signatures. To balance geometric overlap, appearance similarity, and motion consistency, a weighted matching cost is defined and assignment is performed accordingly:

[0091]

[0092] Where: matching cost Dimensionless, smaller values ​​are better; weighting Dimensionless, taking values ​​in Up to 1, and the sum of the three is Overlap function The value ranges from 0 to , which represents the intersection-union ratio of the two bounding boxes;

[0093]

[0094] in, and These are two target regions located in the same stable image coordinate system; Let the area be the region, and the union area be zero. .

[0095] Cosine distance of appearance The value ranges from 0 to , representing the angular distance between the normalized appearance vectors of a unit sphere; center displacement penalty This represents the ratio of the center distance after image stabilization to the allowable distance;

[0096]

[0097] The center pixel coordinates of the candidate volume. The upper limit of allowable displacement (unit: mm) Euclidean norm; previous frame bounding box , current frame bounding box It has been defined within the stable image coordinates; the appearance vector of the previous frame. , appearance vector of this frame The unit sphere has been normalized.

[0098] Processing method: Construct a cost matrix and obtain one-to-one matches using a Hungarian assignment solver; for each pair of matches, if Less than the threshold or If the value exceeds a threshold, the match is cancelled and the target is moved to the unmatched set. Within the occlusion reconnection window, targets that are temporarily invisible are allowed to reconnect with subsequent frames without being counted repeatedly. In application, this cost combines three complementary pieces of information: overlap limits spatial consistency, appearance cosine distance characterizes identity similarity, and center displacement penalty suppresses large displacement mismatches. Solving within stable image coordinates ensures intertemporal geometric consistency, and the unmatched set is the direct source of new candidate objects. Measurable effects and measurement conditions: Under fixed lighting and uniform lens conditions, drawing... The distribution and matching pass rate at the threshold; the algorithm configuration includes a Hungarian solver, bilinear interpolation for center coordinate interpolation, and does not introduce statistical variance indicators.

[0099] To ensure that new judgments have clear thresholds, the acceptance criteria are defined as follows:

[0100]

[0101] Where: Crossover Union (CUC) threshold The value is in Up to 1, used for minimum overlap requirement; cost threshold Positive numbers control the maximum acceptable cost;

[0102] Only when the above equations are satisfied will As a valid match; if not satisfied, it will be... Add to the unmatched set.

[0103] In application, the dual-threshold structure ensures that both spatial and appearance-based evidence meet the requirements simultaneously, reducing mismatches entering the deduplication process; the unmatched set is filtered through an occlusion window to form a list of new individuals. Measurable effects and measurement conditions: Changing the cross-union ratio (CUI) threshold under a unified scenario. With cost threshold Observe the stability curve of the newly added list; the solver and interpolation are the same as described above.

[0104] At the frame level, image sharpness, glare ratio, and dirt coverage are synthesized into a score; frames with scores below a threshold are not included in incremental calculations. At the candidate volume level, individuals with adhesion or unclosed boundaries are marked as rejected. To improve the interpretability of the gating, the quality score is defined as a linear combination of monotonic functions:

[0105]

[0106] Where: quality score Dimensionless, the larger the value, the higher the quality; weight Dimensionless, taking values ​​in to And the sum of the three is 1;

[0107] Monotonic mapping is a publicly known piecewise linear function, defined by non-negative real numbers, with a range of . .

[0108] Sharpness mapping: , Maximum reference sharpness value; Glare ratio mapping: , Maximum acceptable glare ratio; Stain coverage mapping: , To the maximum acceptable stain coverage ratio;

[0109] Sharpness index The unit is grayscale energy, derived from Laplacian energy or gradient energy accumulation; sharpness index :

[0110]

[0111] For the working area of ​​the cup rim, It is a grayscale image. for Directional gradient (calculated using Sobel / Laplacian operator), The reference maximum resolution energy is the one specified at the factory.

[0112] Glare ratio The value is in to This indicates the percentage of saturated pixels; glare ratio. :

[0113]

[0114] Pixel saturation threshold This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise.

[0115] Stain coverage The value ranges from 0 to 1, representing the percentage of the area suspected to be a stain or water droplet; stain coverage. :

[0116]

[0117] in, The dark spot threshold, radius Morphological closing operation, This represents the number of pixels in the stained area.

[0118] In calculation Substitute into the above formula to obtain the quality score Compare with the quality threshold; if the quality score If the value is below the threshold, the candidate body of the current frame will not participate in the cross-time frame assignment. At the candidate body level, mask topology inspection and skeleton continuity criteria are used to put non-closed and inseparable sticky samples into the rejection queue, export local small images and mark them to be reviewed.

[0119] In application, quality gating eliminates out-of-focus and strong reflection frames at the source, reducing false positives; the rejection queue outputs individuals with high uncertainty in a centralized manner, which facilitates unified review in the cloud and subsequent retraining, maintaining the stability of the list of newly added individuals.

[0120] Step 3: Based on the relative values ​​of near-attractant temperature, number of days since release, and volatility intensity, release rate is estimated. Combined with prior data on effective sampling range and corrections for wind direction and speed, forest edge, and canopy, cover estimate is obtained. A standardized capture index is constructed and quantiles are generated for subsequent threshold and trend determination. Release rate estimate, effective sampling range estimate, and standardized capture index are archived by timestamp and station number to maintain cross-station alignment and recalculation consistency, providing a unified data source for boundary crossing determination and group comparison in Step 4.

[0121] When selecting the current site - calculation for this period in the operation and maintenance terminal, the system automatically loads the number of newly added individuals, near-attractant temperature, number of days of release, relative value of volatility intensity and prior of effective sampling range, and displays the parameter list simultaneously;

[0122] Subsequently, the release rate estimation, effective sampling range correction, index construction and quantile normalization are performed in the order of the procedure. The standardized capture index curve (coverage of the points on the day), index quantile band and parameter record table (including timestamp) appear on the terminal page. The on-duty personnel click to confirm and archive, and the system generates index calculation records and puts them into the database. If any input is missing or the verification fails, the system rolls back to the last qualified parameter status and prompts for supplementation and recalculation.

[0123] The release intensity of pheromone attractants is dominated by near-attractant temperature and number of days since release, and individual differences can be corrected by the relative value of volatility intensity; the range of insect encounter is affected by forest edge distance, wind speed and direction, and canopy density.

[0124] If these two types of physical quantities are not explicitly quantified at the same time, the number of new individuals will lose its unified scale. Therefore, using the physical quantities at the time of a single collection as input, the release rate is first estimated, and then the effective sampling range is corrected based on prior knowledge, thereby converging time-temperature-release-wind field-forest edge-canopy into two calculable quantities.

[0125] First, the release rate estimate is derived using the temperature-time relationship, and then a first-order correction is performed by combining the relative value of the volatile intensity. Next, the effective sampling range is a priori corrected according to wind speed, wind direction and canopy density to obtain a spatial coverage estimate consistent with the sampling moment. Both are used as denominators in the exponent construction.

[0126] At the same acquisition time, both the near-attractant temperature and the number of days since release are available, with the relative value of volatility intensity serving as an observation of individual differences. To avoid relying solely on empirical tables, a continuous temperature-time function is established, and a linear gain term for the relative value of volatility intensity is introduced, allowing the release rate to be updated frame by frame.

[0127]

[0128] Where: Release rate estimate The unit is mass / time, and it is one of the components in the denominator of the exponent; benchmark release rate. The unit is mass / time, representing the factory parameters of the attractant formulation at the reference temperature; degree coefficient. Dimensionless, greater than 1, describes the doubling relationship of release for every 10 degrees Celsius increase in temperature; near the inducer temperature. The temperature at which the fusion is achieved;

[0129] Reference temperature Unit: degrees Celsius; factory rated temperature; attenuation coefficient. The unit is Depending on the formulation settings, this characterizes the exponential decay of release with respect to the number of days of application; number of days of application The unit is days, derived from the equipment fingerprint field table; gain coefficient. Dimensionless The linear effect of the relative value of volatility intensity on the release rate was controlled; the relative value of volatility intensity Dimensionless, representing the relative reading obtained in step one; reference volatile intensity. , dimensionless, is the relative reading of the formula under reference conditions.

[0130] In application, this formula binds temperature, time, and individual evaporation differences to a single moment in a closed-form formula, eliminating the reliance on offline tables for release rate estimation. Since all inputs are timestamped, release rate estimation... It can be audited and recalculated, ensuring the physical traceability of the exponent denominator.

[0131] The effective sampling range prior is derived from step one, reflecting the basic state of station geometry and meteorology. To incorporate instantaneous wind field and canopy density into the spatial characterization of the same frame, a linear subtraction term for wind direction cosine and canopy density is introduced into the prior, resulting in the coverage correction:

[0132]

[0133] Where: Effective sampling range estimation As one of the components in the denominator of the exponent; the effective sampling range prior. , is the prior output of step one; attenuation coefficient The unit is The exponential decay intensity characterizing forest edge distance; forest edge distance Unit: meters; values ​​recorded at the station; wind gain coefficient. The unit is , representing the linear gain of wind speed and coverage;

[0134] wind speed The unit is meters per second, representing the anemometer reading for the current frame; relative angle. Units in radians represent the angle between the wind direction and the orientation of the trap; canopy deduction factor. Dimensionless Canopy density Dimensionless Estimated from forest stand data or remote sensing.

[0135] When applied, the coverage modification explicitly incorporates wind directionality and canopy shading effect simultaneously, with directional terms... By incorporating plume directionality into coverage calculations, the canopy term transforms the decrease in encounter probability into a calculable proportional deduction. Consequently, the spatial denominator can be stored with each frame and recalculated.

[0136] The original number of newly added individuals output from different sites is affected by systemic differences in release rate and coverage, and cannot be directly compared. Even if the physical quantity in the denominator is substituted, a robust normalization method that does not require sample variance is still needed to ensure that the trigger quantity has an interpretable center and quantile band within a short window. Therefore, an index is constructed after physical normalization, and quantile bands are generated by updating the quantile bands frame by frame for subsequent threshold-trend joint triggering.

[0137] First, a standardized capture index is constructed using physical quantities as the denominator; then, a quantile band is formed in the sense of a rolling window using a frame-by-frame quantile update algorithm, avoiding the use of standard deviation or variance and maintaining robustness; finally, the index, quantile pairs, and parameter records are output and archived together.

[0138] The number of newly added individuals is combined with the release rate estimate and coverage estimate in the same frame to form a dimensionless standardized capture index, which serves as the sole trigger value:

[0139]

[0140] Where: Standardized capture index Dimensionless, trigger quantity; number of newly added individuals The unit is units per cycle, and this is the output of step two; release rate estimation. Units are mass / time; effective sampling range estimation Unit: square meters;

[0141] When applied, this formula unifies the number of individuals reaching a comparable number under what release intensity and spatial coverage, and the differences across sites are absorbed through the denominator; since the denominator is obtained from the same frame, the exponent has strict temporal consistency.

[0142] To avoid relying on standard deviation or variance, a frame-by-frame quantile update function is used to provide the center and confidence band of the exponent in a rolling sense. For any quantile... The following recursion can be used:

[0143]

[0144] Where: quantile estimation ,and Dimensionless, indicating time... quantiles; step size Dimensionless , can be Set to a fixed constant or a slowly decreasing constant; indicator function The value is 1 if the proposition in parentheses is true, otherwise it is 1. ; quantile level Dimensionless ,For example Indicates the median.

[0145] right , , Synchronous recursion yields the median and upper and lower quantile bands; and , Simultaneously output as a reference band for subsequent threshold-trend joint triggering.

[0146] When applied, quantile recursion does not require variance estimation and has low sensitivity to extreme points. Since the recursion only depends on the quantile value of the previous time step and the current exponent, the computational cost is stable, it can be implemented at the edge, and it can be audited in conjunction with timestamps.

[0147] Step 4: Candidate alarms are generated by combining the quantile band crossing line with the exponentially weighted slope. These are then screened using a group comparison weighted by wind direction and distance. A risk score is calculated, and the sampling level and uplink packaging are deterministically switched to generate alarm event streams, maintenance instructions, and work orders. Alarm records are indexed with newly added individual small images, stabilized orthophotos, and parameter record tables to form a replayable and auditable evidence chain. Simultaneously, a rejection list is sent back for cloud review and model updates, closing the complete link from judgment to handling to retrospective analysis.

[0148] In the operations and maintenance terminal, selecting the current site and today's alerts automatically loads the standardized capture index, quantile estimates, and a list of neighboring sites. The page displays the trend judgment for exceeding the threshold, the baseline comparison for the same group, and the suggested sampling level. The on-duty personnel click to generate an alarm, and the system generates an alarm event stream and pushes it simultaneously to the mobile device. The mobile device displays entries for level, location, time, suggested action, and confidence level. Then, clicking to issue the sampling level and work order immediately switches the sampling frequency and lighting level on the equipment side. The platform generates a work order number, task content, site coordinates, and contact person record and sends it to the work team. Upon completion of the task, a confirmation receipt is sent on the terminal. If any step in the judgment process fails, the system rolls back to the previous version of the threshold and quantile band, suspends high-level alarms, and prompts for supplementary data entry or review before recalculation.

[0149] A single threshold is susceptible to short-duration pulses, and a single trend may lead to frequent false triggers in high-noise segments. Therefore, it is necessary to combine the standardized capture index with its quantile bands. First, cross-limit prediction is performed at the frame level, and then trend evidence is formed through robust slope recursion. Alarm generation only proceeds when both conditions are met. Since the quantile bands are derived from the frame-by-frame quantile updates in step three, the criteria do not need to rely on standard deviation or variance; the slope recursion uses an exponentially weighted form, which can suppress high-frequency oscillations without sacrificing temporal consistency.

[0150] On the same time scale, the over-limit threshold is first constructed using the upper edge of the quantile band and the bandwidth to obtain a Boolean indication; then, the normalized capture index of two adjacent frames is subjected to exponential weighted difference to obtain the trend strength; after combining the two with the deterministic threshold, the frame-level trigger result of the over-limit and trend being established simultaneously is output, and a timestamp is added to enter the alarm pipeline.

[0151] To enable judgments significantly higher than the recent background to be executed, a threshold is constructed using the upper edge of the quantile band and the bandwidth to form a Boolean indicator:

[0152]

[0153] In the formula: Crossing line indicator Dimensionless, taking values ​​of Or 1, a Boolean value indicating whether the boundary has been crossed; Standardized capture index. Dimensionless; quantile estimation and Dimensionless, representing the upper and lower quantiles at the same moment, respectively, derived from the quantile recursion in step three; upper quantile level With lower quantile level Dimensionless, taking values ​​between 0 and 1, and can be used in engineering. , Threshold coefficient , dimensionless, non-negative, used to extrapolate the upper threshold based on bandwidth.

[0154] Read the normalized capture index into the current frame. Quantile estimation and Calculate according to the above formula The Boolean value and the timestamp are then pushed onto the stack; the threshold coefficient value is fixed according to the site-level procedure.

[0155] In application, the over-line discrimination anchors the judgment of high-level anomalies to the quantile band near the window. The bandwidth term provides a margin for local fluctuations, thereby reducing the sensitivity to short pulses. The Boolean result is auditable and can be directly used for subsequent concatenation.

[0156] Crossing the line alone can still lead to false alarms, therefore, an exponentially weighted slope of the difference between adjacent frames is introduced as evidence of a trend:

[0157]

[0158] Where: trend strength ,and Dimensionless, representing the difference between adjacent frames smoothed with exponential weights; smoothing coefficient. Dimensionless A higher value indicates a longer memory duration; Standardized capture index Compared to the previous frame Dimensionless. Trend strength is calculated recursively. After that, with engineering threshold Comparison yields Boolean indications ;Will and Conjunction occurs only if both are true. When the alarm enters the alarm pipeline, candidate alarm records are generated.

[0159] When applied, the exponentially weighted slope provides evidence of an upward trend without introducing variance classifiers; the conjunctive strategy binds high positions to positive growth rates, thereby suppressing false triggers caused by isolated spikes.

[0160] Significant spatial correlation exists in forest areas, and single-point anomalies may be caused by micrometeorological conditions or differences in deployment. It is necessary to construct a co-group control baseline weighted by wind direction and distance, and further screen candidate alarm records. At the same time, in order to maintain stable backhaul under the constraints of weak network and limited power, it is necessary to deterministically switch the edge-side sampling interval and uplink packetization method based on risk scores, thereby increasing the frequency of event-driven events and throttling non-event periods.

[0161] First, the standardized capture index of neighboring sites is normalized and weighted according to wind direction and distance to obtain the same group control baseline; the candidate alarms and the baseline deviation criteria are merged to form the final alarm Boolean value and risk score; based on the score, the edge side sampling and uplink packaging are switched using a deterministic sampling interval function, and work orders and operation and maintenance instructions are generated on the platform side to complete the closed loop.

[0162] To make the control group more closely reflect the direction of plume propagation, the indices of neighboring sites were aggregated using wind direction cosine and distance weighting to form a co-group control:

[0163]

[0164] Where: baseline of the same group control Dimensionless, representing the time of neighboring stations. Weighted index; number of neighboring sites , a positive integer, selected by the platform based on geographical adjacency and communication reachability; weight Dimensionless, non-negative, and normalized to 1 at the same time; static distance weight constant. Dimensionless, commonly used as a discrete value of the exponential decay function of distance; relative angle The unit is radians, representing the angle between the current station and a neighboring station relative to the prevailing wind direction in the current frame; the neighboring station index... , dimensionless, from the output of step three at a neighboring site.

[0165] Among them, the weights are calculated. Subsequent aggregation yielded baseline data for cohort controls. Set the deviation from the threshold ratio ,when If the condition is met, the alarm is considered to be abnormal. This Boolean value is then merged with the aforementioned candidate alarm records to obtain the final alarm record.

[0166] When applied, the baseline weights neighboring sites under the combined influence of wind direction and distance, which can distinguish between regional rises and local anomalies; the sampling frequency is increased only when the baseline deviation is valid, thereby concentrating resources on more suspicious sites.

[0167] To ensure a one-to-one correspondence between edge-side sampling and feedback and risk levels, the risk score is defined as a weighted sum of three Boolean values ​​(crossing the line, trend, and control deviation) followed by linear normalization. Interval.

[0168] The sampling interval employs a deterministic monotonically decreasing function for risk scoring, ensuring that higher risks result in denser sampling, while clearly defining minimum and maximum intervals to balance energy and bandwidth boundaries.

[0169]

[0170] Where: sampling interval , represents the time interval between the next frame and the current frame; minimum interval With the maximum interval In engineering, it is a definite constant that satisfies Slope control coefficient Dimensionless, positive, controls the steepness of the function; risk score Dimensionless, interval This is a weighted and normalized summary of the overshoot indicator, trend indicator, and control deviation indicator;

[0171]

[0172] Risk Score The value range is [0, 1], comprehensively reflecting the risk level of exceeding the line, trend, and deviation; weight. , , : Positive real number, used to adjust the contribution weights of the three types of Boolean indicators; sampling interval The unit is minutes, representing the time interval between the next frame and the current frame; minimum interval. Maximum interval : A positive real number, in minutes, and , which are boundary values ​​set for the project; slope coefficient : Positive real number, controlling the steepness of the sampling interval function; center point The value range is (0, 1), which is the inflection point for switching sampling levels (when the risk score reaches a certain threshold). When, the interval is ).

[0173] Calculate risk score Substituting into the above equation, we obtain the sampling interval. Based on this, the sampling frequency, exposure mode, and lighting level are switched on the device side; if the risk score exceeds the high-level threshold, the uplink packaging is switched to a simplified package that only uploads small images of newly added individuals and statistical summaries, and a work order and a drug replacement reminder are generated simultaneously; if the risk score falls back below the low-level threshold, the regular interval and periodic heartbeat package are restored.

[0174] In application, the deterministic interval function makes the risk-frequency-bandwidth relationship transparent, maintaining continuous backhaul under weak network and limited power conditions; uplink is packaged in the high-risk segment to carry key evidence with minimal load, and the low-risk segment maintains healthy signals, enabling stable aggregation and backhaul on the platform side.

[0175] Please see Figure 2 This invention provides a system for automatic identification and location of small pests based on multimodal macro imaging, comprising:

[0176] The calibration and acquisition module is equipped with smart attractant cards, which read the fingerprints of the equipment and the near-attractant temperature. It generates a stable orthophoto and a pixel-to-millimeter conversion table according to the coded calibration pattern, and combines it with site metadata to form a priori effective sampling range.

[0177] The deduplication module performs target detection and instance segmentation on the stabilized orthophoto, constructs instance signatures, performs cross-time frame matching at unified coordinates, counts unmatched individuals as new additions, obtains the number of new individuals, and saves the new small images and matching records.

[0178] The index calculation module estimates the release rate based on near-attractant temperature and number of days since release, and estimates the effective sampling range based on prior information on effective sampling range, wind direction and speed and forest edge information. The standardized capture index is obtained by dividing the number of new individuals by the above two quantities.

[0179] The early warning linkage module performs boundary crossing and trend determination under the standardized capture index scale. It uses the index of neighboring stations to form a group comparison by weighting the index by wind direction and distance, and filters alarms. Based on the determination, it sends the sampling level and uplink package and generates alarm records.

[0180] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0181] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0183] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0184] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automatic identification and localization method for small pests based on multimodal macro imaging, characterized in that: include, Deploy smart lure cards, read the device fingerprint and near lure temperature, generate stable orthophotos and pixel-to-millimeter conversion tables according to the coded calibration pattern, and combine them with site metadata to form a priori effective sampling range; Perform target detection and instance segmentation on the stabilized orthophoto, construct instance signatures, perform time-frame matching at unified coordinates, count unmatched individuals as new individuals, obtain the number of new individuals, and save the new small images and matching records. The release rate was estimated based on near-attractant temperature and number of days after release. The effective sampling range was estimated based on prior information on effective sampling range, wind direction and speed and forest edge information. The standardized capture index was obtained by dividing the number of new individuals by the release rate and the effective sampling range. Cross-line judgment and trend judgment are performed under the standardized capture index scale. The index of neighboring stations is weighted by wind direction and distance to form a group comparison and screen alarms. Based on the judgment, the sampling level is issued and the uplink is packaged and alarm records are generated. Release rate is estimated by combining near-attractant temperature, number of days since release, and relative value of volatility intensity; effective sampling range is corrected by combining the effective sampling range prior with forest edge distance, wind direction and speed, and canopy density; Both are timestamped for recalculation and written into the parameter record table with a unified identifier for cross-site alignment. They are also used as the denominator of the same frame for reference in the standardized capture index and are linked to the deployment date, trap model and orientation in the equipment fingerprint field table. Crossing the line is determined by extrapolating the upper edge of the quantile zone and the bandwidth to form a threshold, and the trend is determined by recursively extrapolating the slope of the exponential weight to form a Boolean indicator. After the two are combined to generate candidate alarms, the index of the neighboring stations is aggregated according to wind direction and distance to form a group comparison and a deviation ratio threshold is set to generate candidate alarm records with station numbers and timestamps. At the same time, newly added individual small map references are retained for auditing. The risk score is obtained by weighting three types of Boolean indicators: crossing the line, trend, and deviation from the same group comparison. The risk score is then mapped to a deterministic switch between the edge-side sampling level and the uplink packaging method. The platform generates alarm event streams, operation and maintenance instructions and work orders in sync, and associates the work orders with alarm records and assigns numbers. At the same time, it records sampling level switching logs and uplink packet lists on the edge side.

2. The automatic identification and location method for small pests according to claim 1, characterized in that: The smart attractant card simultaneously carries an coded calibration pattern, equipment fingerprint identification, proximity temperature sensor, color temperature scale, and gas-sensitive element or passive sampling sheet; Within a single frame, the camera completes the reading and generation of a pixel-to-millimeter conversion table, an equipment fingerprint field table, and the number of days of deployment. It also forms a local record of image stabilization homography and timestamp association, and binds the relative values ​​of near-inducer temperature and volatilization intensity with site metadata into the database.

3. The automatic identification and positioning method for small pests according to claim 2, characterized in that: The cup cavity is equipped with cross-polarized ring lighting, matte black lining and white positioning ring at the cup mouth, along with anti-fog coating and drainage structure; after generating stable orthophotos, an effective sampling range prior is formed based on forest edge distance, wind direction and speed and canopy density and updated with each frame, and a correlation key is established with the pixel-to-millimeter conversion table for unified access, while the lighting and orientation configuration is recorded in the equipment fingerprint field table.

4. The automatic identification and positioning method for small pests according to claim 3, characterized in that: On the stabilized orthophoto, a block-based inference output candidate volume mask is used with fixed overlap; when constructing the instance signature, at least a color histogram, texture response, shape description and millimeter volume scale are included and written in a fixed field order; Cross-time frame matching generates matching records and newly added individual small images based on weighted costs, and binds them to a pixel-to-millimeter conversion table. The block boundaries and overlap widths are fixed at the site level and recorded with the version.

5. The automatic identification and location method for small pests according to claim 4, characterized in that: A composite score is generated based on the sharpness index, glare ratio, and stain coverage. Frames with scores below the threshold are not included in the increment. Candidates with adhesion or unclosed boundaries are placed in the rejection queue and local small images are exported. Targets within the occlusion reconnection window are reconnected without being counted as new, and the quality score is bound to the frame timestamp for archiving. The rejection queue is sorted by priority and the instance signature field is updated and reviewed in the cloud.

6. The automatic identification and location method for small pests according to claim 5, characterized in that: After constructing the standardized capture index, the median and upper and lower quantile bands are generated by frame-by-frame recursion and archived together with the standardized capture index. The quantile bands serve as the reference baseline for boundary crossing and trend determination, and an index relationship is established with the newly added individual number and parameter record table. When archiving, the site number and timestamp are written to maintain consistent reference with the equipment fingerprint field table and the number of deployment days.

7. A system for automatic identification and localization of small pests based on multimodal macro imaging, using the method described in any one of claims 1 to 6, characterized in that: include, The calibration and acquisition module is equipped with smart attractant cards, which read the fingerprints of the equipment and the near-attractant temperature. It generates a stable orthophoto and a pixel-to-millimeter conversion table according to the coded calibration pattern, and combines it with site metadata to form a priori effective sampling range. The deduplication module performs target detection and instance segmentation on the stabilized orthophoto, constructs instance signatures, performs cross-time frame matching at unified coordinates, counts unmatched individuals as new additions, obtains the number of new individuals, and saves the new small images and matching records. The index calculation module estimates the release rate based on near-attractant temperature and number of days since release, and estimates the effective sampling range based on prior information on effective sampling range, wind direction and speed and forest edge information. The standardized capture index is obtained by dividing the number of new individuals by the release rate and the effective sampling range. The early warning linkage module performs boundary crossing and trend determination under the standardized capture index scale. It uses the index of neighboring stations to form a group comparison by weighting the index by wind direction and distance, and filters alarms. Based on the determination, it sends the sampling level and uplink package and generates alarm records.

Images