A pest information acquisition system for smart agriculture

Abstract

The present application relates to the technical field of edge computing, and discloses a pest information collection system for smart agriculture, which is deployed in open farmland, orchards and facilities greenhouse, and realizes automatic monitoring and intelligent analysis of insect situation. The system is composed of a multi-source sensing terminal, a field edge computing node, a pest information collection server and a pest information management platform. The terminal collects trapping images and environmental data; the edge node completes pest detection and board-level statistics based on an improved YOLOv8n model; the cloud platform generates insect situation trend prediction and risk results through a time-space dual-track verification AI model (TCN-BiGRU and GAT-TCN); and the agricultural management terminal realizes insect situation query and spatio-temporal visualization through a REST interface. The present application realizes intelligent monitoring and prediction of pests in an end-edge-cloud collaborative manner, and improves the insect situation identification accuracy, spatio-temporal trend judgment ability and agricultural management efficiency.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a pest information collection system for smart agriculture. Background Technology

[0002] With the rapid development of smart agriculture and computer vision technology, existing pest monitoring systems are gradually moving from traditional sensor threshold alarms to AI perception and image target recognition. However, existing intelligent pest information collection technologies are mostly based on general visual models for detection or use a single-track solution of cloud-based aggregation and recognition. These technologies still have significant limitations in open weather environments such as real farmland. On the one hand, when some lightweight YOLO models or convolutional visual networks are directly transferred to agricultural small target detection tasks, they lack multi-scale semantic role decomposition and channel stability calibration mechanisms tailored to the characteristics of insect scales. This leads to an imbalance in the coupled expression of crop canopy, small insects, and complex agricultural background features, easily resulting in high class confidence but low accuracy. Significant structural errors due to positional offsets; on the other hand, although current AIoT edge deployment solutions introduce local inference capabilities, they generally lack sufficient NPU computing power or do not enable geometrically stable offset unbiased calibration logic, making it easy for target box overlap redundancy and density statistical bias fluctuations to occur between traps in dense monitoring networks; in addition, existing insect infestation prediction models mostly use single temporal or graph networks for modeling, but lack time-space dual-track collaborative learning and consistency and reliability loss feedback verification mechanisms, which cannot effectively suppress the accumulation of non-stationary structural errors driven by wind direction, distance-related background similarity during insect outbreaks, and cannot form a reliable and traceable multi-period insect infestation spatiotemporal index retrieval capability. Summary of the Invention

[0003] To address the technical bottlenecks of existing smart agriculture pest data collection systems, such as the difficulty in balancing small target recognition and scale semantic coupling under open farmland meteorological backgrounds, insufficient geometric stability of edge-side inference, target box overlap leading to fluctuations in counting and statistical bias under dense deployment across traps, and the lack of temporal-spatial dual-track consistency verification and multi-period index retrieval reliable response mechanisms, this invention provides a pest information collection system for smart agriculture. This system achieves significant improvements in edge-cloud collaborative inference acceleration, unbiased deduplication and aggregation counting of pests across multiple traps, edge representation learning based on wind direction / distance causal correlation of pest spread, self-verification of multi-period pest inference, and distribution of highly consistent and interpretable index retrieval query services. These improvements are achieved through the construction of a multi-scale semantic responsibility-defined convolutional visual coding framework, channel stability calibration and non-negative reliable attention projection mechanism, instance size prior-guided unbiased center offset regression calibration logic, and a temporal-graph network dual-track consistent verification and re-marking mechanism and composite primary key pest spatiotemporal index retrieval architecture. A systematic breakthrough has been achieved, enabling the output of board-level insect infestation prediction vectors, plot-level insect infestation evolution trend detection sequences, insect detection box sets, spatiotemporal insect density distribution, and reliable risk level labels. This significantly improves the reliable retrieval and interaction capabilities of AI-driven automatic insect infestation collection in real agricultural planting environments, consistent response modeling and verification of temporal and spatial trends, unbiased deduplication counting statistics across trap boards, and meteorological background-driven diffusion maps for open-field or facility agriculture. It provides highly reliable, highly consistent response, and interpretable self-verification capabilities for precise monitoring, intelligent early warning, and decision-making linkage of spatiotemporal insect outbreaks in farmland or greenhouse management.

[0004] This invention proposes a pest information collection system for smart agriculture, which includes a multi-source sensing terminal, a field edge computing node, a pest information collection module, a pest information management platform, and an agricultural management terminal.

[0005] Multi-source sensing terminals are deployed near the crop canopy in open farmland, orchards, or greenhouses to continuously monitor and capture data on pest activity areas, generating raw monitoring data with high-precision time and location markers. The raw monitoring data includes pest trapping images (images of pests on trapping boards or crop surfaces collected by imaging sensors) and multi-dimensional environmental data of the adult pest trapping area (including air temperature, relative humidity, light intensity, wind speed and direction, and crop plot codes) to reflect the light and meteorological background conditions for pest occurrence. The raw monitoring data is then uploaded to field edge computing nodes.

[0006] The field edge computing node, with its embedded CPU+NPU processing architecture and built-in AI inference acceleration capabilities, performs distortion correction, adaptive brightness and contrast compensation, noise filtering, and cropping on pest-trapped images to generate standardized pest images. An improved YOLOv8n model is used to perform inference on these standardized pest images. This improved YOLOv8n model includes an improved backbone network, an improved Neck structure, and a decoupled detection head, outputting granular pest detection results, including the bounding box coordinates, species category label, and confidence value for each pest target. Cross-board level target bounding box deduplication and consistency counting are performed using the trap board ID, yielding pest quantity statistics and category density distribution indexed by the trap board ID. The pest quantity statistics, category density distribution, and corresponding environmental parameters are then structured, uniformly encoded, and hashed into a pest information collection record, which is uploaded to the pest information collection module via a wireless network.

[0007] The pest information collection module is deployed on the physical server of the agricultural cloud data center. It receives pest information collection records and constructs pest and pest archive data at the plot level.

[0008] The pest information management platform is deployed on a cloud server cluster. It receives pest and insect pest archive data, has a built-in time-series-spatial dual-track verification AI model, generates board-level pest prediction results and plot-level pest trend prediction results, and provides REST standardized pest query interface services and risk level output.

[0009] The temporal-spatial dual-track verification AI model includes a board-level temporal verification model and a plot-level spatial verification model: the board-level temporal verification model is based on a temporal convolutional network-bidirectional gated recurrent unit structure for prediction; the plot-level spatial verification model is based on a graph attention network-temporal convolutional structure for modeling.

[0010] The agricultural management terminal connects to the pest information management platform via a wireless network. It is used to call the REST standardized pest query interface service provided by the platform and the results of pest prediction and reliable risk level labeling at the board / plot level. It can also perform multi-key retrieval and visualization of pest counts, species density distribution, spatial heat map, and 1-365 day time series pest evolution data under the monitoring network, to assist agricultural technicians in pest control analysis and scheduling decision support.

[0011] Furthermore, the process of using an improved YOLOv8n model to perform inference on standardized pest images and output pest granularity detection results specifically includes the following steps:

[0012] Step S1: Based on YOLOv8n, a multi-scale semantic decomposition strategy for pest collection scenarios is introduced into the YOLOv8n backbone network, so that features of different scales can respectively undertake the tasks of sensitive texture of small pests, semantic modeling of main body contours, and auxiliary representation of farmland background anti-disturbance; before the output of the backbone network enters the Neck structure, LayerNorm channel stabilization calibration and non-negative confidence weighted projection are performed to construct an improved backbone network; the standardized pest image is input into the improved backbone network to form a geometrically stable intermediate feature map;

[0013] Step S2: Input the geometrically stable intermediate feature map into the improved Neck structure and output the fused feature map;

[0014] Step S3: Decoupled detection heads are deployed on the fused feature map to perform agricultural pest detection inference, outputting bounding box confidence logits, center point offset probability distribution bins, species category classification logits, and category probability for each candidate pest instance; then, instance size priors based on the width and height of the candidate instance bounding boxes are introduced, and unbiased, proportional scaling and probability integration are performed on the center point offset probability distribution bins. Combined with the upper layer logits prior and the current layer offset Δlogits correction, layer-by-layer expected offset iterative calibration is performed to complete the distortion correction calibration of the bounding boxes and the fine regression of the expected offset of the pixel sites, generating a set of highly reliable pest detection boxes that are distortion-corrected and unbiased.

[0015] Step S4: Perform confidence screening on the high-confidence pest detection box set to obtain the pest particle size detection results.

[0016] Furthermore, step S1 specifically includes the following steps:

[0017] Step S11: The improved backbone network performs cascaded feature processing and spatial progressive compression on the standardized pest image through Conv convolutional chains, C2f lightweight geometric sensitive feature extraction module, and multi-level downsampling convolutions with increasing stride (stride4, 16, 32), outputting intermediate feature maps; among them, the downsampling layer with stride=4 is used to form a high-resolution detail basis, the downsampling layer with stride=16 is used to carry the semantics of the main structure, and the downsampling layer with stride=32 is used to construct a global background auxiliary semantic representation;

[0018] Step S12: Based on the intermediate feature map, perform explicit semantic role decoupling and scale responsibility division in the agricultural scenario to form decoupled multi-scale feature data;

[0019] Step S13: Perform LayerNorm normalization on the channel dimension of the decoupled multi-scale feature data to suppress feature distribution bias drift during agricultural trapping imaging; and combine the scale prior of candidate instances to implement non-negative confidence weighted compression and scale-aligned projection of the embedding space to ensure that the offset feature amplitude is proportional to the instance size and the error is not amplified, thus forming a geometrically stable intermediate feature map.

[0020] Furthermore, step S2 specifically includes the following steps:

[0021] Step S21: Use a top-down Feature Pyramid Network (FPN) and a bottom-up Path Aggregation Network (PAN) to perform multi-scale feature upsampling, downsampling and lateral connection fusion on the geometrically stable intermediate feature map to generate a preliminary fused feature map;

[0022] Step S22: From the initially fused feature map, extract Nc scale template query vectors along the scale dimension, perform channel-level LayerNorm normalization calibration on all scale template query vectors, and obtain a scale-stable calibration template vector set;

[0023] Step S23: From the initially fused feature map, Nq pest instance query vectors are extracted along the instance dimension using a learnable linear projection layer to form an instance query vector set;

[0024] Step S24: Based on the scale-stable calibration template vector set and the instance query vector set, perform decoupled two-stage Self-Attention to obtain scale semantic consistency encoding and instance subject semantic encoding;

[0025] Decoupled two-stage Self-Attention includes a scale consistency modeling stage and an instance semantic modeling stage;

[0026] Scale consistency modeling stage: Self-Attention modeling is performed on Nc scale template query vectors to learn the consistency of cross-scale geometric semantics and the credibility of contour association, forming scale semantic consistency encoding;

[0027] Instance semantic modeling stage: Self-Attention modeling is performed on Nq pest instance query vectors to learn the main semantic attributes of pests, contour associations and local semantic expression capabilities, and obtain the main semantic encoding of the instance.

[0028] Step S25: Based on scale semantic consistency encoding and instance subject semantic encoding, perform instance-scale composite query vector binding and fuse them to generate composite query vector;

[0029] Step S26: In the local Cross-Attention update stage of the improved Neck structure, the composite Query vector is subjected to local geometric enhancement feature sampling and multi-scale semantic attention interaction update to suppress occlusion or weakly associated Query responses, and view-scale consistency constraints and outlier attention interaction screening are completed to output the fused feature map.

[0030] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0031] This invention utilizes an improved YOLOv8n model deployed in field edge computing nodes. It constructs an innovative detection link with stable channel-scale-instance encoding and fine-grained center offset regression correction, enhancing the localization accuracy of tiny insect target boxes in open farmland and greenhouse canopy backgrounds, as well as the geometric stability of NPU end-side inference. This solves key problems commonly found in existing agricultural vision detection models, such as high confidence but large localization offsets, channel activation distribution drift, coupling of responsibilities at different scales, and overlapping target boxes leading to repetitive responses and difficulty in converging statistical bias fluctuations in dense trap scenarios. Through LayerNorm multi-scale channel calibration, Soft-Bins unbiased offset layer-by-layer back-index correction, and a hash-based deduplication aggregation mechanism for trap IDs, it provides the system with a high-quality, non-repetitive, statistically stable, and reliable pest detection and counting basis.

[0032] In terms of pest prediction and verification, this invention is based on dual-track consistent modeling of TCN-BiGRU (time track) and GAT-TCN (spatial track) and consistent cross-validation closed loop, which improves the numerical stability and consistency of the prediction of multi-period pest temporal fluctuations and spatial diffusion trends under open farmland meteorological background. It solves the problems of inconsistency between temporal and spatial correlation learning, amplification of cross-period error accumulation, fluctuation of pest trend query response and insufficient credibility under dense trapping network in existing smart agriculture pest prediction models. It enables the system to achieve stable pest trend prediction, density distribution response and reliable regression output of risk labels at both the board level and the plot level.

[0033] In summary, this invention forms a complete and innovative closed loop in terms of stable generation of detection boxes for small pests, unbiased deduplication counting across dense trapping boards, consistent prediction of pest infestations in multiple periods of time and space, and stable retrieval feedback verification of composite primary key pest indexes. It improves the system's engineering implementation capabilities in real farmland and greenhouse deployments, ensuring stable target box positioning, no duplicate pest counts, no amplification of spatiotemporal diffusion learning errors, consistent and reliable index retrieval response, and interpretable and credible regression of result risk level output. It overcomes the fundamental defects of existing smart agriculture visual pest systems, such as unstable three-track coordination of detection-prediction-retrieval process and the cumulative impact of duplicate box response and center offset errors on statistical consistency. It provides a system-level outstanding contribution capability that aligns with the objectives of this invention and is feasible for the accurate collection, modeling, retrieval, and intelligent early warning of pests in smart agriculture. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall structure of a pest information collection system for smart agriculture proposed in this invention.

[0035] Figure 2 This is a schematic diagram of the bidirectional path aggregation Neck structure of the FPN and PAN in the improved YOLOv8n model proposed in this invention.

[0036] Figure 2 In this context, Backbone represents the improved backbone network, FPN represents the feature pyramid network, and PAN represents the path aggregation network; Downsampling represents downsampling, and Upsampling represents upsampling. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] Example 1, according to Figure 1 , Figure 2 This invention proposes a pest information collection system for smart agriculture, which can be applied to open farmland, orchards and greenhouse planting environments. It is used to automatically collect and structure the number, species and spatiotemporal distribution information of pests in the target crop area. The system includes a multi-source sensing terminal, a field edge computing node, a pest information collection module, a pest information management platform and an agricultural management terminal.

[0039] Multi-source sensing terminals are deployed near the crop canopy in open farmland, orchards, or greenhouses to continuously monitor and capture data on pest activity areas, generating raw monitoring data with high-precision time and location markers. The raw monitoring data includes pest trapping images (images of pests on trapping boards or crop surfaces acquired by imaging sensors) and multi-dimensional environmental data of the adult pest trapping area (including air temperature, relative humidity, light intensity, wind speed and direction, and crop plot codes) to reflect the light and meteorological background conditions for pest occurrence. The raw monitoring data is then uploaded to field edge computing nodes. The multi-source sensing terminal comprises an adult pest trapping unit, an image acquisition unit, and an environmental sensing unit, specifically:

[0040] The adult insect trapping unit includes an insect-attracting lamp, a rain cover, and a replaceable trapping board, which is used to attract flying pests at a preset height and fix them in the collection area.

[0041] The image acquisition unit is mounted on the fixed bracket of the adult insect trapping unit. The image acquisition unit includes a high-definition visible light camera with automatic exposure and autofocus functions and a ring LED fill light, which is used to image the surface of the trapping board from multiple angles within a preset time interval to generate insect trapping images.

[0042] The environmental sensing unit includes temperature and humidity sensors, light intensity sensors, and wind speed and direction sensors, which are used to simultaneously collect multi-dimensional environmental data of the adult insect trapping area.

[0043] The field edge computing node, with its embedded CPU+NPU processing architecture and built-in AI inference acceleration capabilities, is deployed in a field data acquisition box (with an IP65 protection rating). It performs distortion correction, adaptive brightness and contrast compensation, noise filtering, and cropping on pest-trapped images to generate standardized pest images. An improved YOLOv8n model is used to perform inference on these standardized images. This improved YOLOv8n model includes an improved backbone network, an improved Neck structure, and a decoupled detection head, outputting granular pest detection results. These results include the bounding box coordinates, species category label, and confidence value for each pest target. Cross-board level target bounding box deduplication and consistency counting are performed using the trap ID, yielding pest quantity statistics and category density distribution indexed by the trap ID. The pest quantity statistics, category density distribution, and corresponding environmental parameters are then structured, uniformly encoded, and hashed to form a pest information collection record, which is uploaded to the pest information collection module via a wireless network.

[0044] The pest information collection module is deployed on the physical server of the agricultural cloud data center (the server is equipped with a StackedSSD solid-state storage array, which belongs to the data processing and archiving layer). It receives pest information collection records and constructs pest and insect profile data at the plot level. The pest and insect profile data includes "plot coordinates, trap ID, identification timestamp, number of pests, category density distribution, temperature, humidity, light and wind environmental factors and high-frequency texture statistical indicators".

[0045] The pest information management platform is deployed on a cloud server cluster (belonging to the external service and decision distribution layer). It receives pest and insect pest archive data, has a built-in time-series-spatial dual-track verification AI model, generates board-level pest prediction results and plot-level pest trend prediction results, and provides REST standardized pest query interface service and risk level output.

[0046] The temporal-spatial dual-track verification AI model includes a board-level temporal verification model (time track) and a plot-level spatial verification model (spatial track): The board-level temporal verification model is based on a temporal convolutional network-bidirectional gated recurrent unit (TCN-BiGRU) structure. It takes the historical pest count sequence, species density sequence, and corresponding environmental parameter sequence of the trapping board as input, extracts multi-period fluctuation patterns, performs trend prediction, and outputs a single-board pest prediction vector for the next H periods. The plot-level spatial verification model is based on a graph attention network-temporal convolution (GAT-TCN) structure. It takes crop plots or board groups within plots as graph nodes, and uses spatial distance, wind direction correlation, and background environmental similarity measurement factors as edge encoding basis. It learns the spatial occurrence association and diffusion trend of pests, and outputs a future pest evolution trend prediction sequence and a spatial retrieval map response vector.

[0047] Both prediction paths employ cross-track consistency loss terms and credibility threshold verification rules for dual-track cross-verification during the training and inference phases. When the consistency index falls below the credibility threshold of the lower track, prediction verification and effective relabeling of the current sample are triggered to suppress weak response or cross-scale heterogeneous error amplification and complete dual-track cross-verification of the numerical consistency and spatial trend credibility of the prediction results.

[0048] The agricultural management terminal connects to the pest information management platform via a wireless network. It is used to call the REST standardized pest query interface service provided by the platform and the results of pest prediction and reliable risk level labeling at the board / plot level. It can also perform multi-key retrieval and visualization of pest counts, species density distribution, spatial heat map, and 1-365 day time series pest evolution data under the monitoring network, to assist agricultural technicians in pest control analysis and scheduling decision support.

[0049] Example 2, based on Example 1, describes the process of using an improved YOLOv8n model to perform inference on standardized pest images and output pest granularity detection results. The specific steps include:

[0050] Step S1: Based on YOLOv8n, a multi-scale semantic decomposition strategy for pest collection scenarios is introduced into the YOLOv8n backbone network, so that features of different scales can respectively undertake the tasks of sensitive texture of small pests, semantic modeling of main body contours, and auxiliary representation of farmland background anti-disturbance; before the output of the backbone network enters the Neck structure, LayerNorm channel stabilization calibration and non-negative confidence weighted projection are performed to construct an improved backbone network; the standardized pest image is input into the improved backbone network to form a geometrically stable intermediate feature map;

[0051] Step S2: Input the geometrically stable intermediate feature map into the improved Neck structure and output the fused feature map;

[0052] Step S3: Decoupled detection heads are deployed on the fused feature map to perform agricultural pest detection inference, outputting bounding box confidence logits, center point offset probability distribution bins, species category classification logits, and category probability for each candidate pest instance; then, instance size priors based on the width and height of the candidate instance bounding boxes are introduced, and unbiased, proportional scaling and probability integration are performed on the center point offset probability distribution bins. Combined with the upper layer logits prior and the current layer offset Δlogits correction, layer-by-layer expected offset iterative calibration is performed to complete the distortion correction calibration of the bounding boxes and the fine regression of the expected offset of the pixel sites, generating a set of highly reliable pest detection boxes that are distortion-corrected and unbiased.

[0053] Step S4: Perform confidence screening on the set of high-confidence pest detection boxes, remove candidate boxes whose classification probability and offset calibration response are lower than the preset confidence threshold, and retain high-confidence pest target boxes with clear responsibilities, complete boundaries and significant responses to obtain pest granularity detection results.

[0054] In conventional technical fields, the process of using YOLOv8n to perform inference on standardized pest images and output pest granularity detection results specifically includes the following steps:

[0055] Step E1: Input the standardized pest image into the YOLOv8n backbone network to obtain the intermediate feature map;

[0056] Step E2: Input the intermediate feature map into the Neck structure of YOLOv8n. The Neck can be composed of FPN top-down upsampling and PAN bottom-up feature aggregation path. Through multi-layer up / downsampling and lateral connection fusion, the fused multi-scale feature map is output.

[0057] Step E3: Call the YOLOv8n decoupled detection head to perform pest target detection inference on the feature map, and output the bounding box coordinates, class logits, and confidence logits for each candidate instance; the center offset probability bins output by the detection head can be scaled and reorganized proportionally according to the width and height of the candidate box to make the box center fit the position of the detected target pixel more closely, thus obtaining the detection box after preliminary correction.

[0058] Step E4: The preliminarily corrected detection boxes are screened based on category probability and confidence level. Low-response or low-confidence candidate boxes are eliminated, and pest target detection boxes with complete boundaries and clear categories are retained to form the final pest instance-level detection results.

[0059] Example 3, this example is based on Example 2. In this example, step S1 specifically includes the following steps:

[0060] Step S11: The improved backbone network performs cascaded feature processing and spatial progressive compression on the standardized pest image through Conv convolutional chains, C2f lightweight geometrically sensitive feature extraction modules, and multi-level downsampling convolutions with increasing stride (stride4, 16, 32), outputting intermediate feature maps. Among them, the downsampling layer with stride=4 is used to form a high-resolution detail basis, the downsampling layer with stride=16 is used to carry the semantics of the main structure, and the downsampling layer with stride=32 is used to construct a global background auxiliary semantic representation. Thus, a three-stage feature extraction link is formed, which is hierarchically progressive and has a stable geometric representation transfer from local details to global semantics.

[0061] Step S12: Based on the intermediate feature map, perform explicit semantic role decoupling and scale responsibility division in the agricultural scenario, extracting and generating three types of decoupled feature maps: 1. High-resolution, small receptive field pest texture-sensitive feature map, used to represent the high-frequency texture, boundary inflection points, and fine-grained geometric structure of small pests near the crop canopy; 2. Medium-resolution, medium receptive field main body contour dynamic semantic map, used to model the main body shape of pests, migration direction, and local contextual features of instances; 3. Low-resolution, large receptive field farmland background auxiliary semantic map, used to represent complex field background texture, global illumination perturbation, and spatial density trend perception signals; forming decoupled multi-scale feature data;

[0062] Step S13: Perform LayerNorm normalization on the channel dimension of the decoupled multi-scale feature data to suppress feature distribution bias drift during agricultural trapping imaging; and combine the scale prior of candidate instances to implement non-negative confidence weighted compression and scale-aligned projection of the embedding space to ensure that the offset feature amplitude is proportional to the instance size and the error is not amplified, thus forming a geometrically stable intermediate feature map.

[0063] Example 4, this example is based on Example 3. In this example, step S2 specifically includes the following steps:

[0064] Step S21: Use a top-down Feature Pyramid Network (FPN) and a bottom-up Path Aggregation Network (PAN) to perform multi-scale feature upsampling, downsampling and lateral connection fusion on the geometrically stable intermediate feature map to generate a preliminary fused feature map;

[0065] Step S22: From the initially fused feature map, extract Nc scale template query vectors along the scale dimension, and perform channel-level LayerNorm normalization calibration on all scale template query vectors to stabilize the numerical distribution of feature channels at each scale, avoid channel activation bias or scale information mixing problems in the early stage of training, and obtain a scale-stable calibration template vector set.

[0066] Step S23: From the initially fused feature map, Nq pest instance query vectors are extracted along the instance dimension using a learnable linear projection layer. Each query is generated by jointly initializing the corresponding pest target's center coordinates prior, regional semantic embedding features, and instance scale feature encoding to form an instance query vector set.

[0067] Step S24: Based on the scale-stable calibration template vector set and the instance query vector set, perform decoupled two-stage Self-Attention to obtain scale semantic consistency encoding and instance subject semantic encoding;

[0068] Decoupled two-stage Self-Attention includes a scale consistency modeling stage and an instance semantic modeling stage;

[0069] Scale consistency modeling stage: Self-Attention modeling is performed on Nc scale template query vectors to learn the consistency of cross-scale geometric semantics and the credibility of contour association, forming scale semantic consistency encoding;

[0070] Instance semantic modeling stage: Self-Attention modeling is performed on Nq pest instance query vectors to learn the main semantic attributes of pests, contour associations and local semantic expression capabilities, and obtain the main semantic encoding of the instance.

[0071] By using the above decoupling process, the scope of attention computation is limited to (within the scale) or (within the instance), thereby avoiding the high complexity and irrational cross-object interference brought about by the Nc×Nq fully connected attention topology and improving training stability.

[0072] Step S25: Based on scale semantic consistency encoding and instance subject semantic encoding, perform instance-scale composite query vector binding, and fuse to generate Nc×Nq composite query vectors with clear responsibilities, which serve as the input basis for the deformation sensitivity representation of agricultural sub-region boundaries and downstream Neck local attention sampling;

[0073] ;

[0074] in, Indicates the first The first pest example in the... Composite Query Vectors under Each Scale Template Indicates the first Each instance semantic encoding vector Indicates the first Scale-consistent encoding vectors corresponding to each scale; Represents the vector fusion operator;

[0075] Step S26: In the local Cross-Attention update stage of the improved Neck structure, the composite Query vector is subjected to local geometric enhancement feature sampling and multi-scale semantic attention interaction update to suppress occlusion or weakly associated Query responses, and view-scale consistency constraints and outlier attention interaction screening are completed to output the fused feature map.

[0076] Example 5, this example is based on Example 4, in this example,

[0077] The pest information management platform is deployed on a cloud server cluster (belonging to the external service and decision distribution layer). It receives pest and insect pest archive data, has a built-in time-series-spatial dual-track verification AI model, generates board-level pest prediction results and plot-level pest trend prediction results, and provides REST standardized pest query interface service and risk level output.

[0078] This embodiment was implemented in a citrus orchard (open-air deployment, covering an area of ​​12 mu).

[0079] Pest trend prediction results for the plot:

[0080] [2025-03-12] 94 animals;

[0081] [2025-03-13] 102 items;

[0082] [2025-03-14] 101 items;

[0083] [2025-03-15] 117 animals;

[0084] [2025-03-16] 129 units;

[0085] [2025-03-17] 128 items;

[0086] [2025-03-18] 142 animals;

[0087] [2025-03-19] 155 pieces;

[0088] [2025-03-20] 153 units;

[0089] [2025-03-21] 172.

[0090] Output:

[0091] In this example, the pest growth rate of the plot from March 12, 2025 to March 21, 2025 was approximately 83%. Output:

[0092] The pest infestation status of plot Plot-A is assessed as: High risk (RED);

[0093] A large-scale insect infestation may spread within the next 10 days and 30 collection cycles.

[0094] The processor recommends the following actions: Immediately issue a control alert for insecticidal lamps and enhance the sampling sparse inflection point.

[0095] The agricultural management terminal connects to the pest information management platform via a wireless network. It is used to call the REST standardized pest query interface service provided by the platform and the results of pest prediction and reliable risk level labeling at the board / plot level. It can also perform multi-key retrieval and visualization of pest counts, species density distribution, spatial heat map, and 1-365 day time series pest evolution data under the monitoring network, to assist agricultural technicians in pest control analysis and scheduling decision support.

[0096] The results of calling the REST insect monitoring API and displaying them on the terminal side are shown in Table 1 (displaying a 5-day range, primary key search mode):

[0097] Table 1

[0098] .

[0099] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A pest information collection system for smart agriculture, characterized in that, The system includes: Multi-source sensing terminals generate raw monitoring data; the raw monitoring data includes images of pest traps and environmental data. The field edge computing node, with its embedded CPU+NPU processing architecture and built-in AI inference acceleration capabilities, performs distortion correction, adaptive brightness and contrast compensation, noise filtering, and cropping on pest-trapped images to generate standardized pest images. An improved YOLOv8n model is used to perform inference on these standardized pest images, outputting granular pest detection results. The trap ID is then used for cross-trap level bounding box deduplication and consistency counting verification to obtain pest quantity statistics and category density distribution. Finally, the pest quantity statistics, category density distribution, and corresponding environmental data are structured, uniformly encoded, and hashed into a pest information collection record. The pest information collection module receives pest information collection records and forms pest information archive data. The pest information management platform receives pest information archive data, has a built-in time-series-spatial dual-track verification AI model, generates board-level pest prediction results and plot-level pest trend prediction results, and provides source interface services. The improved YOLOv8n model includes an improved backbone network, an improved Neck structure, and a decoupled detection head; The process of using an improved YOLOv8n model to perform inference on standardized pest images and output pest granularity detection results includes the following steps: Step S1: Input the standardized pest images into the improved backbone network to form geometrically stable intermediate feature maps; Step S2: Input the geometrically stable intermediate feature map into the improved Neck structure and output the fused feature map; Step S3: Decouple the detection head on the fused feature map to perform agricultural pest detection inference, and output the bounding box confidence logits, center point offset probability distribution bins, species category classification logits and category probability for each candidate pest instance; introduce instance size prior based on the width and height of the candidate instance bounding box, and perform unbiased and proportional scale adaptation scaling and probability integration on the center point offset probability distribution bins to generate a set of highly reliable pest detection boxes; Step S4: Perform confidence screening on the high-confidence pest detection box set to obtain the pest particle size detection results; Step S2 specifically includes the following steps: Step S21: Use a top-down feature pyramid network and a bottom-up path aggregation network to perform multi-scale feature upsampling, downsampling and lateral connection fusion on the geometrically stable intermediate feature map to generate a preliminary fused feature map. Step S22: From the initially fused feature map, extract Nc scale template query vectors along the scale dimension, perform channel-level LayerNorm normalization calibration on all scale template query vectors, and obtain a scale-stable calibration template vector set; Step S23: From the initially fused feature map, Nq pest instance query vectors are extracted along the instance dimension using a learnable linear projection layer to form an instance query vector set; Step S24: Based on the scale-stable calibration template vector set and the instance query vector set, perform decoupled two-stage Self-Attention to obtain scale semantic consistency encoding and instance subject semantic encoding; Step S25: Based on scale semantic consistency encoding and instance subject semantic encoding, perform instance-scale composite query vector binding and fuse them to generate composite query vector; Step S26: In the local Cross-Attention update stage of the improved Neck structure, the composite Query vector is updated by local geometric enhancement feature sampling and multi-scale semantic attention interaction, and the fused feature map is output.

2. The pest information collection system for smart agriculture according to claim 1, characterized in that: The improved backbone network is constructed as follows: based on YOLOv8n, a multi-scale semantic responsibility decomposition strategy is introduced into the YOLOv8n backbone network, so that features of different scales respectively undertake the tasks of sensitive texture of small pests, semantic modeling of main body contours and auxiliary representation of farmland background anti-disturbance; and LayerNorm channel stabilization calibration and non-negative reliable weighted projection are performed before the backbone network output enters the Neck structure to construct the improved backbone network.

3. The pest information collection system for smart agriculture according to claim 1, characterized in that: Decoupled two-stage Self-Attention includes a scale consistency modeling stage and an instance semantic modeling stage; Scale consistency modeling stage: Self-Attention modeling is performed on Nc scale template query vectors to learn the consistency of cross-scale geometric semantics and the credibility of contour association, forming scale semantic consistency encoding; Example Semantic modeling stage: Self-Attention modeling is performed on the query vectors of Nq pest instances to learn the main semantic attributes, contour relationships and local semantic expression capabilities of the pests, and obtain the semantic encoding of the main instance.

4. A pest information collection system for smart agriculture according to claim 1, characterized in that: The temporal-spatial dual-track verification AI model includes a board-level temporal verification model and a plot-level spatial verification model: the board-level temporal verification model is based on a temporal convolutional network-bidirectional gated recurrent unit structure for prediction; The land parcel spatial verification model is modeled based on a graph attention network-temporal convolutional structure.

Images