CN-121982709-A - Method and system for analyzing bemisia tabaci feeding trend based on artificial intelligence

CN121982709ACN 121982709 ACN121982709 ACN 121982709ACN-121982709-A

Abstract

The invention relates to the technical field of insect feeding trend analysis, in particular to a bemisia tabaci feeding trend analysis method and system based on artificial intelligence. The method comprises the steps of utilizing an attention-enhanced YOLO network to extract a three-dimensional behavior track of insects from a stabilized video sequence, analyzing track segments marked with feeding events through track dynamics and gesture sequences to obtain a feeding event sequence, digging associated weights of the feeding event sequence and volatile dynamic response characteristics through multiple heads of attention to obtain behavior chemical decision characteristics, obtaining track plant interaction characteristics based on the three-dimensional behavior track and plant phenotype characteristics through a space-time diagram neural network taking track points as nodes, and predicting bemisia tabaci population feeding hot spot distribution by utilizing multi-source perception fusion trend decision modeling of deep reinforcement learning in combination with the plant phenotype characteristics, the behavior chemical decision characteristics and the track plant interaction characteristics. The invention can realize the deep fusion of multi-source data and accurately analyze the eating trend.

Inventors

HUANG LINJIE
LI ZHENGGANG
HE ZIFU
SHE XIAOMAN
TANG YAFEI

Assignees

华南农业大学
广东省农业科学院植物保护研究所

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (10)

1. The bemisia tabaci feeding trend analysis method based on artificial intelligence is characterized by comprising the following steps of: separating bemisia tabaci foreground and plant background from the monitoring video sequence to obtain a foreground video sequence and a background model, and guiding plant non-rigid motion compensation constrained by plant structures through optical flow based on the foreground video sequence and the background model to obtain a stabilized video sequence with effective compensation; extracting three-dimensional behavior tracks of insects from the stabilized video sequence by using an attention-enhancing YOLO network, and analyzing track segments marked with feeding events through track dynamics and gesture sequences to obtain a feeding event sequence; digging the associated weights of the feeding event sequence and the volatile dynamic response characteristic through multiple heads of attention to obtain a behavioural-chemical decision characteristic; Based on the three-dimensional behavior track and the plant phenotype characteristics, obtaining track plant interaction characteristics through a space-time diagram neural network taking track points as nodes; And predicting the distribution of the bemisia tabaci population feeding hot spots by combining the plant phenotype characteristics, the behavioral chemistry decision characteristics and the trajectory plant interaction characteristics and utilizing the multisource perception fusion trend decision modeling of deep reinforcement learning.
2. The method for analyzing the feeding trend of bemisia tabaci based on artificial intelligence according to claim 1, wherein the steps of separating bemisia tabaci foreground and plant background from the monitoring video sequence to obtain a foreground video sequence and a background model comprise the following steps: based on the monitoring video sequence, a clear denoising image frame sequence with complete bemisia tabaci detail is obtained by utilizing bilateral filtering; and carrying out dynamic background modeling and foreground separation according to the denoising image frame sequence to obtain a foreground video sequence and a background model.
3. The artificial intelligence based bemisia tabaci feeding trend analysis method according to claim 1, wherein the stabilized video sequence with effective compensation is obtained by non-rigid motion compensation of plants constrained by optical flow guidance and plant structure based on the foreground video sequence and background model, comprising the steps of: capturing displacement characteristics of curling and shaking of plant leaves through optical flow according to the foreground video sequence and the background model; Based on the displacement characteristics, the plant geometric structure prior constraint compensation direction and amplitude are utilized to eliminate the interference of plant motion on the bemisia tabaci track, and a stabilized video sequence with effective compensation is obtained.
4. The artificial intelligence based bemisia tabaci feeding trend analysis method according to claim 1, wherein the extracting the three-dimensional behavior trace of the insect from the stabilized video sequence using the attention-enhancing YOLO network comprises the steps of: obtaining three-dimensional position information of bemisia tabaci from the stabilized video sequence using an attention-enhancing YOLO network; and based on the three-dimensional position information, tracking and extracting the three-dimensional behavior track of the insect through the track constrained by the space-time continuity.
5. The artificial intelligence based bemisia tabaci feeding trend analysis method according to claim 1, wherein the track section marked with feeding events is analyzed by track dynamics and gesture sequence to obtain feeding event sequence, comprising the steps of: analyzing track dynamics characteristics according to the three-dimensional track coordinate sequence to obtain feeding candidate track points; carrying out gesture sequence analysis based on the feeding candidate track points to obtain gesture classification results; And weighting and fusing the track dynamics characteristics and the gesture classification result to obtain a ingestion event sequence.
6. The artificial intelligence based bemisia tabaci feeding trend analysis method according to claim 1, wherein the step of mining the associated weights of the feeding event sequence and the volatile dynamic response feature through multiple head attentions to obtain a behavioural chemical decision feature comprises the following steps: Constructing a multi-head attention fusion model; taking the eating event as a query end, taking the volatile matter characteristic as a key and value end, and mining the associated weight through the multi-head attention fusion model to obtain the behavior chemistry decision characteristic.
7. The artificial intelligence-based bemisia tabaci feeding trend analysis method according to claim 1, wherein the obtaining of the trace plant interaction characteristics through the space-time diagram neural network with the trace points as nodes based on the three-dimensional behavior trace and the plant phenotype characteristics comprises the following steps: Taking the track points as nodes of the graph, and constructing a space-time graph by taking the plant phenotypes as node attributes; And excavating the track plant interaction characteristics of the space-time diagram through a space-time diagram convolution network.
8. The artificial intelligence based bemisia tabaci feeding trend analysis method according to claim 1, wherein the combining the plant phenotype characteristics, the behavioral chemical decision characteristics and the trajectory plant interaction characteristics, utilizing deep reinforcement learning multisource perception fusion trend decision modeling, predicting bemisia tabaci feeding hotspot distribution comprises the following steps: constructing a deep reinforcement learning model; And taking the plant phenotype characteristics, the behavioral chemical decision characteristics and the trajectory plant interaction characteristics as state input of the deep reinforcement learning model, taking feeding hot spot prediction as an action target, and predicting the distribution of the feeding hot spots of the bemisia tabaci population through a reward function.
9. The artificial intelligence based bemisia tabaci feeding trend analysis method according to claim 1, wherein extracting the plant phenotype features comprises the steps of: Fusing the multi-view plant images to obtain a full-view three-dimensional plant phenotype image; and carrying out super-pixel segmentation on the full-view three-dimensional plant phenotype image, and extracting phenotype feature vectors from segmentation results.
10. An artificial intelligence-based bemisia tabaci feeding trend analysis system, which is characterized by comprising an input device, an output device, a processor and a memory, wherein the input device, the output device, the processor and the memory are mutually connected, and the memory comprises program instructions for executing the artificial intelligence-based bemisia tabaci feeding trend analysis method according to any one of claims 1-9.

Description

Method and system for analyzing bemisia tabaci feeding trend based on artificial intelligence Technical Field The invention relates to the technical field of insect feeding trend analysis, in particular to a bemisia tabaci feeding trend analysis method and system based on artificial intelligence. Background Bemisia tabaci is a worldwide agricultural pest, and adults and nymphs of the bemisia tabaci damage crops by feeding on host plant juice, and can transmit various plant viruses at the same time, so that the yield and quality of the crops are seriously affected. The food tendency of the bemisia tabaci is accurately analyzed, the preferred host plant area and relevant influencing factors are defined, and the method has important significance in formulating an accurate prevention and control strategy and reducing pesticide abuse. At present, the related technology of bemisia tabaci feeding trend analysis has the following defects: The detection accuracy of the tiny target is insufficient, namely the length of the bemisia tabaci is smaller than 1mm, the traditional target detection algorithm such as a general YOLO model is not adapted to the tiny target, and the tiny target is easy to be interfered by plant textures and environmental noise, so that the detection omission rate and the false detection rate are high; the background interference treatment is poor, namely, in a natural monitoring scene, host plants can generate slight shaking, curling and other non-rigid motions due to environmental factors, the traditional video image stabilizing technology is designed aiming at a rigid background, the interference of plant motions on bemisia tabaci track extraction cannot be effectively eliminated, and the plant motions are easily misjudged as bemisia tabaci behaviors; The multisource data fusion is inaccurate, namely the food intake trend of the bemisia tabaci is affected by multiple factors such as plant phenotype, volatile matters and the like, but the prior art multi-single-dimension analysis data does not realize the deep fusion of behavior data, plant phenotype data and volatile matter chemical data, and the problem that the volatile matter collection and the behavior monitoring are not synchronous exists; The generalization capability of the chemotaxis prediction model is poor, the traditional prediction method mostly adopts regression, general CNN and other models, the characteristic customization design of the bemisia tabaci feeding behavior is not combined, the comprehensive influence of multiple factors on the feeding trend cannot be accurately quantized, and the hot spot prediction accuracy is low. Therefore, a bemisia tabaci feeding trend analysis technology capable of accurately processing tiny target detection, effectively eliminating plant dynamic interference, realizing multi-source data deep fusion and accurately predicting is needed, so as to solve the defects in the prior art. . Disclosure of Invention Aiming at the inadequacy of the existing method and the requirement of practical application, the method aims at solving the problems. In one aspect, the invention provides an artificial intelligence-based bemisia tabaci feeding trend analysis method, which comprises the following steps: The method comprises the steps of separating bemisia tabaci foreground and plant background from a monitoring video sequence, obtaining a foreground video sequence and a background model, guiding plant non-rigid motion compensation constrained by a plant structure through optical flow based on the foreground video sequence and the background model, obtaining a stabilized video sequence with effective compensation, extracting three-dimensional behavior tracks of insects from the stabilized video sequence through a concentration enhancement YOLO network, analyzing track segments marked with feeding events through track dynamics and gesture sequences, obtaining a feeding event sequence, excavating associated weights of the feeding event sequence and volatile dynamic response characteristics through multi-head attention, obtaining behavior chemical decision characteristics, obtaining track plant interaction characteristics through a space-time graph neural network taking track points as nodes based on the three-dimensional behavior tracks and the plant phenotype characteristics, combining the plant phenotype characteristics, the behavior chemical decision characteristics and the track plant interaction characteristics, and predicting the feeding distribution of bemisia tabaci population through multi-source perception fusion trend decision modeling of depth reinforcement learning. Optionally, the separation of bemisia tabaci foreground and plant background from the monitoring video sequence to obtain a foreground video sequence and a background model comprises the following steps: And carrying out dynamic background modeling and foreground separation according to the denoising image frame sequence to obtain a fore