Agricultural intelligent bird repelling method and system based on AI recognition

Abstract

The invention relates to an agricultural intelligent bird repelling method and system based on AI recognition, and the method achieves the high-precision perception and recognition of farmland bird information through the deep learning and multi-source heterogeneous information fusion technology, remarkably improves the pertinence and effectiveness of a final early warning strategy, and improves the early warning efficiency. And multi-source time sequence data is constructed according to the first time sequence model and the second time sequence model, and a final driving strategy is formulated by taking the activity rhythm of the birds as guidance, so that the interference on the normal physiological rhythm of the birds is minimized. By integrating various driving means such as laser, ultrasonic waves and flash lamps, the driving duration is greatly prolonged, meanwhile, the driving strategy optimization based on the multi-target optimization model enables the system to have autonomous learning and adaptive capacity, the manual intervention requirement is greatly reduced, and the operation and maintenance cost is reduced by 60% or above compared with a traditional method.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart agriculture, and in particular relates to an agricultural intelligent bird-repelling method and system based on AI recognition. Background Art

[0002] With the rapid development of modern agriculture, bird damage in farmland has become one of the important factors affecting crop yield and quality. Traditional bird damage prevention and control methods, such as manual driving away, sound and light bird driving, etc., often have problems such as short-term effect, high labor intensity and high cost. Although new bird driving technologies such as ultrasound and laser have been applied in recent years, facing the complex and changeable farmland environment and bird behavior, the effect of a single bird driving method is often unsatisfactory.

[0003] At the same time, with the advancement of ecological civilization construction, agricultural production has put forward higher requirements for environmental friendliness and ecological sustainability. How to effectively prevent and control bird damage while minimizing interference with farmland ecosystems and achieving harmonious coexistence of people, birds and crops has become a problem that needs to be solved urgently.

[0004] In addition, factors such as the complexity of farmland environment, the randomness of bird damage, and the limitations of farmland power supply have posed severe challenges to the design of intelligent bird repellent systems. Therefore, it is urgent to develop a comprehensive farmland bird repellent system that integrates intelligent perception, multi-modal repellent, and ecological regulation to meet the needs of bird damage prevention and control in modern agricultural production. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to propose an intelligent agricultural bird-repelling method and system based on AI recognition, which solves the problem that bird damage cannot be effectively solved in existing agricultural production.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention provides an intelligent agricultural bird-repelling method based on AI recognition, the method comprising:

[0008] Acquire a first image sequence in the farmland scene, where the first image sequence is a plurality of first images in a first preset sampling period;

[0009] Inputting the first image sequence into a first convolutional neural network model for feature extraction to obtain first image feature information, wherein the first image feature information includes at least one of a plurality of first bird activity trajectories and a plurality of first bird behavior patterns in the current first image sequence;

[0010] and simultaneously acquiring thermal infrared information in the farmland scene, the thermal infrared information being a plurality of thermal infrared image sequences within a first preset sampling period;

[0011] Inputting the thermal infrared information into a second convolutional neural network model for feature extraction to obtain second thermal infrared feature information, wherein the second thermal infrared feature information includes at least one of a plurality of second bird activity trajectories and a plurality of second bird behavior patterns in the current thermal infrared information;

[0012] Acquire bird attribute information associated with the farmland scene, construct a bird attribute semantic dictionary, map the attribute into a semantic vector through an attribute embedding matrix, and obtain a third attribute semantic vector, where the third attribute semantic vector includes at least one of a plurality of third bird activity trajectories and a plurality of third bird behavior patterns in the current bird attribute information;

[0013] Inputting the first image feature information, the second thermal infrared feature information and the third attribute semantic vector into a multimodal gated fusion unit for feature interaction and fusion to obtain a semantically enhanced multimodal fusion feature sequence;

[0014] Inputting the multimodal fusion feature sequence into a long short-term memory network, modeling the temporal dependency of bird activities, and obtaining a first temporal model, wherein the first temporal model includes first bird temporal information, and the first bird temporal information includes activity trajectories and behavior patterns of the birds in the first time series;

[0015] Acquire a second image sequence of the farmland scene within a second preset sampling period, where the second image sequence is a plurality of second images within the second preset sampling period;

[0016] Inputting the second image sequence into a third convolutional neural network model for feature extraction to obtain fourth image feature information, obtaining bird category information based on the fourth image feature information, and obtaining a fourth bird activity trajectory and a fourth bird behavior pattern based on the second image sequence;

[0017] Encoding is performed according to the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information to obtain a second time series model, wherein the second time series model includes the second bird time series information, and the second bird time series information includes the activity trajectory and behavior pattern of the bird in the second time series;

[0018] A final expulsion strategy and / or a final early warning strategy is generated according to the first time series model, the bird category information, and the second time series model.

[0019] In some embodiments, the first image sequence is input into the first convolutional neural network model for feature extraction, and the first image feature information is obtained by formula (1), which is as follows:

[0020] F img =f cnn (I; θ cnn );

[0021] In formula (1), F img is the first image feature information, f cnn is the feature extractor, θ cnn is the learning parameter of the first convolutional neural network model, I is the first image sequence, T is the first preset sampling period;

[0022] The first convolutional neural network model is generated based on the convolutional neural network model. The performance evaluation of the first convolutional neural network model is obtained through the cross entropy loss function. The cross entropy loss function is expressed by formula (2). Formula (2) is as follows:

[0023]

[0024] In formula (2), L is the loss function value, y i The true category of the sample bird data during the training of the first convolutional neural network model, Probability distribution for prediction of sample bird data during training of the first convolutional neural network model;

[0025] The thermal infrared information is input into the second convolutional neural network model for feature extraction, and the second thermal infrared feature information is expressed by formula (3). Formula (3) is as follows:

[0026] F ir =f cnn (H; θ ir );

[0027] In formula (3), F ir is the second thermal infrared characteristic information, f cnn is the feature extractor, θ ir is the learning parameter of the second convolutional neural network model, H is the thermal infrared information, T is the first preset sampling period.

[0028] In some embodiments, the bird attribute semantic dictionary is obtained by the following steps:

[0029] Construct a semantic dictionary. The semantic dictionary is expressed by formula (4), which is as follows:

[0030] D=d 1 ,d 2 ,...,d N ;

[0031] In formula (4), D is the semantic dictionary, d i is an attribute word, which represents an attribute of a bird;

[0032] Match a semantic embedding vector to each attribute word to form an attribute embedding matrix and a semantic representation space;

[0033] The attribute embedding matrix is ​​expressed by formula (5), which is as follows:

[0034] E=[e 1 ,e 2 ,…,e N ];

[0035] In formula (5), E is the attribute embedding matrix, e i is a semantic embedding vector, which represents the position of the attribute word in the semantic representation space.

[0036] In some embodiments, the first image feature information, the second thermal infrared feature information, and the third attribute semantic vector are input into a multimodal gated fusion unit for feature interaction and fusion, and a semantically enhanced multimodal fusion feature sequence is obtained, which is expressed by formulas (6) to (8). Formula (6) is as follows:

[0037] F′ img =F img +f att (F img ,F ir ,E)⊙F ir ;

[0038] Formula (7) is as follows:

[0039] F′ ir =F ir +f att (F ir ,F img ,E)⊙F img ;

[0040] Formula (8) is as follows:

[0041] F fused =W 1 F′img+W 2 F′ ir +W 3 (F′ img ⊙F′ ir );

[0042] In formulas (6) to (8), F' img is the first feature information after fusion, F img is the first image feature information, f att is the first attention mechanism function, E is the attribute embedding matrix, F' ir is the second feature information after fusion, F fused is the first fusion feature information, W 1 , W 2 , W3 is the learning parameter of fusion gating.

[0043] In some embodiments, the multimodal fusion feature sequence is input into a long short-term memory network, and the temporal dependency of the bird activity is modeled to obtain a first temporal model, wherein the first temporal model includes first bird temporal information, and the first bird temporal information includes the activity trajectory and behavior pattern of the bird in the first time series, including:

[0044] The first long short-term memory network is used to encode the multimodal fusion feature sequence and expressed by formula (9):

[0045]

[0046] In formula (9), h t and c t They represent the hidden state and memory cell state of the first long short-term memory network at the tth moment, respectively, and f lstm is the first long short-term memory network function, θ lstm are the parameters of the first long short-term memory network;

[0047] The first time series model is trained using a multi-task learning mode to obtain a trained first time series model. The training process is expressed by formula (10):

[0048]

[0049] In formula (10), is the total loss function value, represents the loss function of the bird detection task, represents the loss function of the bird tracking task, represents the loss function of the bird behavior recognition task, λ 1 , 2 and λ 3 To balance the weight coefficients of different tasks.

[0050] In some embodiments, the second image sequence is input into the third convolutional neural network model for feature extraction to obtain fourth image feature information, the bird category information is obtained based on the fourth image feature information, and the fourth bird activity trajectory and the fourth bird behavior pattern are obtained according to the second image sequence, including:

[0051] The second attention mechanism function is added to the third convolutional neural network model and expressed by formula (11). Formula (11) is as follows:

[0052] F = A⊙X;

[0053] In formula (11), F is the attention enhancement feature in the second attention mechanism function, A is the attention weight matrix, X is the convolution feature map in the third convolutional neural network model, and ⊙ is element-wise multiplication;

[0054] The attention weight matrix is ​​expressed by formula (12), which is as follows:

[0055]

[0056] In formula (12), W a and W x is the learning parameter in the second attention mechanism function.

[0057] In some embodiments, encoding is performed according to the fourth bird activity trajectory, the fourth bird behavior pattern, and the bird category information to obtain a second time series model, the second time series model includes the second bird time series information, and the second bird time series information includes the activity trajectory and behavior pattern of the bird in the second time series.

[0058] The second long short-term memory network is used to encode the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information, and is expressed by formulas (13) to (18). Formula (13) is as follows:

[0059] i t =σ(W ii x t +W hi h t-1 +b i );

[0060] Formula (14) is as follows:

[0061] f t =σ(W if xt+W hf h t-1 +b f );

[0062] Formula (15) is as follows:

[0063] o t =σ(W io x t +W ho h t-1 +b o );

[0064] Formula (16) is as follows:

[0065]

[0066] Formula (17) is as follows:

[0067]

[0068] Formula (18) is as follows:

[0069] h t ′=o t ⊙tanh(C t );

[0070] In formulas (13) to (18), i t is the input gate in the second LSTM network, f t is the forget gate in the second long short-term memory network, o t is the output gate in the second LSTM network, C t is the cell state in the second long short-term memory network, is the candidate cell state in the second LSTM network, h t ' is the hidden state in the second long short-term memory network, W ii , W hi is the weight matrix of the input gate, b i is the bias vector of the input gate, W if , W hf is the weight matrix of the forget gate, b f is the bias vector of the forget gate, W io , W ho is the weight matrix of the output gate, b o is the bias vector of the output gate;

[0071] The second time series model is trained using a multi-task learning mode to obtain a trained first time series model. The training process is expressed by formula (19):

[0072] L'=λ' 1 L recog +λ' 2 L' behav ; In formula (19), L′ is the total loss function value, is the loss function for the bird recognition task, L' behav is the loss function for the bird behavior understanding task, λ' 1 and λ' 2 is the weight coefficient to balance the two tasks.

[0073] In some embodiments, generating a final expulsion strategy according to the first timing model and the second timing model further includes:

[0074] An intelligent adjustment model is constructed, and observation information is input into the intelligent adjustment model to obtain an initial driving strategy, wherein the observation information includes environmental parameters, a first time series model, a second time series model, and bird category information;

[0075] Evaluate the expulsion effect of the initial expulsion strategy, optimize the initial expulsion strategy according to the expulsion effect evaluation, and obtain the final expulsion strategy;

[0076] The final expulsion strategy is expressed by formula (20), which is as follows:

[0077]

[0078] In formula (20), π * (s) is the final expulsion strategy, s is the observed state, is the device parameter set of the driving device used by the intelligent adjustment model, a is the driving action in the final driving strategy, Q(s,a;θ * ) is the state-action value function, θ * is the optimal parameter of the state-action value function.

[0079] In some embodiments, generating a final early warning strategy according to the first time series model and the second time series model further includes:

[0080] Construct multi-source time series data according to the first time series model and the second time series model, wherein the multi-source time series data includes an agricultural ecological calendar, a bird migration calendar, and a meteorological calendar;

[0081] Generate the macroscopic time series law of bird damage based on multi-source time series data, divide the farmland scene into multi-scale grids, and generate the spatial distribution heat map of farmland bird damage;

[0082] A multi-objective optimization model is constructed. The input data of the multi-objective optimization model are multi-source time series data, spatial distribution heat map, bird species diversity index and degree of human intervention. The output result of the multi-objective optimization model is the final early warning strategy.

[0083] In a second aspect, the present invention further provides an agricultural intelligent bird-repelling system based on AI recognition, which is applicable to the agricultural intelligent bird-repelling method of the first aspect, and the system includes a control unit, a communication module, an energy supply module, a bird-repelling device, an image acquisition module and an AI recognition module;

[0084] The control unit is used to execute the method described in the first aspect; the communication module includes at least one of a Bluetooth module, a wifi module, and an antenna, and the communication module is used to establish a communication connection with an external device; the energy supply module includes at least one of a solar panel, a micro wind turbine, and a battery pack, and the energy supply system is used to provide power; the bird-repelling device includes at least one of a laser transmitter, an ultrasonic generator, and an LED flash array, and the bird-repelling device is used to execute the final bird-repelling strategy; the image acquisition module includes a high-definition camera and an infrared imager, and the image acquisition module is used to obtain a first image sequence, a second image sequence, and thermal radiation information; the AI ​​recognition module is used to perform image processing on the first image sequence, the second image sequence, and the thermal radiation information, and the AI ​​recognition module is used to train and correct the first neural network model, the second neural network model, and the third neural network model.

[0085] By adopting the above technical solution, this technical solution realizes high-precision perception and recognition of farmland bird information by using deep learning and multi-source heterogeneous information fusion technology, significantly improving the pertinence and effectiveness of the final early warning strategy, and constructing multi-source time series data based on the first time series model and the second time series model, and formulating the final expulsion strategy based on the rhythm of bird activity, minimizing the interference with the normal physiological rhythm of birds. By integrating multiple expulsion methods such as laser, ultrasound, and flash, the duration of expulsion is greatly extended. At the same time, the expulsion strategy optimization based on the multi-objective optimization model enables the system to have autonomous learning and adaptability, greatly reducing the need for manual intervention, and reducing the operation and maintenance cost by more than 60% compared with traditional methods.

[0086] The above-mentioned records related to the content of the invention are only an overview of the technical solution of the present invention. In order to enable ordinary technicians in the field to more clearly understand the technical solution of the present invention, and then implement it according to the text of the specification and the contents recorded in the drawings, and in order to make the above-mentioned purpose and other purposes, features and advantages of the present invention easier to understand, the following is a description in combination with the specific implementation mode of the present invention and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The drawings are only used to illustrate the principles, implementations, applications, features and effects of the specific embodiments of the present invention and other related contents, and shall not be considered as limiting the present invention.

[0088] In the drawings of the specification:

[0089] Figure 1 A schematic diagram of the generation of the first timing module in a specific implementation manner;

[0090] Figure 2 A schematic diagram of generating the second timing module in a specific implementation manner;

[0091] Figure 3 It is a schematic diagram of the generation principle of the final expulsion strategy described in the specific implementation method;

[0092] Figure 4 This is a schematic diagram of the generation principle of the final early warning strategy described in this specific implementation method;

[0093] Figure 5 It is a schematic diagram of the agricultural intelligent bird-repelling system described in a specific implementation method. DETAILED DESCRIPTION

[0094] In order to explain in detail the possible application scenarios, technical principles, specific schemes that can be implemented, and the purposes and effects of the present invention, the following is a detailed description in combination with the specific embodiments listed and the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of the present invention, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present invention.

[0095] Reference to "embodiment" herein means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The word "embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or association with other embodiments. In principle, in the present invention, as long as there is no technical contradiction or conflict, the various technical features mentioned in each embodiment can be combined in any way to form a corresponding implementable technical solution.

[0096] Unless otherwise defined, the technical terms used herein have the same meanings as those generally understood by those skilled in the art to which the present invention belongs; the use of relevant terms herein is only for describing specific embodiments and is not intended to limit the present invention.

[0097] In the description of the present invention, the term "and / or" is an expression used to describe the logical relationship between objects, indicating that three relationships may exist, for example, A and / or B, which means: A exists, B exists, and A and B exist at the same time. In addition, the character " / " in this article generally indicates that the objects before and after are in a logical relationship of "or".

[0098] In the present invention, terms such as "first" and "second" are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship of quantity, priority or sequence between these entities or operations.

[0099] Without further restrictions, in the present invention, the words "include", "comprises", "has" or other similar open-ended expressions used in sentences are intended to cover non-exclusive inclusion. These expressions do not exclude the presence of additional elements in the process, method or product including the elements, so that the process, method or product including a series of elements may include not only those limited elements, but also other elements not explicitly listed, or also include elements inherent to such process, method or product.

[0100] Similar to the understanding in the Examination Guidelines, in the present invention, expressions such as "greater than", "less than", "exceed" and the like are understood to exclude the number itself; expressions such as "above", "below", "within" and the like are understood to include the number itself. In addition, in the description of the embodiments of the present invention, the meaning of "multiple" is more than two (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups", "multiple times", etc., unless otherwise clearly and specifically limited.

[0101] In the description of the embodiments of the present invention, the space-related expressions used, such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicate the orientation or position relationship based on the orientation or position relationship shown in the specific embodiments or drawings, and are only for the convenience of describing the specific embodiments of the present invention or facilitating the reader's understanding, and do not indicate or imply that the referred device or component must have a specific position, a specific orientation, or be constructed or operated in a specific orientation, and therefore cannot be understood as a limitation on the embodiments of the present invention.

[0102] Unless otherwise expressly specified or limited, in the description of the embodiments of the present invention, the terms such as "install", "connect", "connect", "fix", "set" and the like used should be understood in a broad sense. For example, the "connection" can be a fixed connection, a detachable connection, or an integrated setting; it can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection, or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. For those skilled in the art of the technical field to which the present invention belongs, the specific meanings of the above terms in the embodiments of the present invention can be understood according to the specific circumstances.

[0103] See also Figures 1 to 5 In a first aspect, this embodiment provides an agricultural intelligent bird-repelling method based on AI recognition, comprising:

[0104] Acquire a first image sequence in the farmland scene, where the first image sequence is a plurality of first images in a first preset sampling period;

[0105] Inputting the first image sequence into the first convolutional neural network model for feature extraction to obtain first image feature information, wherein the first image feature information includes at least one of a plurality of first bird activity trajectories and a plurality of first bird behavior patterns in the current first image sequence; and simultaneously obtaining thermal infrared information in the farmland scene, wherein the thermal infrared information is a plurality of thermal infrared sequences within a first preset sampling period;

[0106] Inputting the thermal infrared information into a second convolutional neural network model for feature extraction to obtain second thermal infrared feature information, wherein the second thermal infrared feature information includes at least one of a plurality of second bird activity trajectories and a plurality of second bird behavior patterns in the current thermal infrared information;

[0107] Acquire bird attribute information associated with the farmland scene, construct a bird attribute semantic dictionary, map the attribute into a semantic vector through an attribute embedding matrix, and obtain a third attribute semantic vector, where the third attribute semantic vector includes at least one of a plurality of third bird activity trajectories and a plurality of third bird behavior patterns in the current bird attribute information;

[0108] Inputting the first image feature information, the second thermal infrared feature information and the third attribute semantic vector into a multimodal gated fusion unit for feature interaction and fusion to obtain a semantically enhanced multimodal fusion feature sequence;

[0109] Inputting the multimodal fusion feature sequence into a long short-term memory network, modeling the temporal dependency of bird activities, and obtaining a first temporal model, wherein the first temporal model includes first bird temporal information, and the first bird temporal information includes activity trajectories and behavior patterns of the birds in the first time series;

[0110] Acquire a second image sequence of the farmland scene within a second preset sampling period, where the second image sequence is a plurality of second images within the second preset sampling period;

[0111] Inputting the second image sequence into a third convolutional neural network model for feature extraction to obtain fourth image feature information, obtaining bird category information based on the fourth sequence feature information, and obtaining a fourth bird activity trajectory and a fourth bird behavior pattern based on the second image sequence;

[0112] Encoding is performed according to the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information to obtain a second time series model, wherein the second time series model includes the second bird time series information, and the second bird time series information includes the activity trajectory and behavior pattern of the bird in the second time series;

[0113] A final expulsion strategy and / or a final early warning strategy is generated according to the first time series model, the bird category information, and the second time series model.

[0114] For further information, see Figure 1 In some embodiments, the first image sequence is input into the first convolutional neural network model for feature extraction, and the first image feature information is obtained by formula (1), which is as follows:

[0115] F img =f cnn (I; θ cnn );

[0116] In formula (1), F img is the first image feature information, f cnn is the feature extractor, θ cnn is the learning parameter of the first convolutional neural network model, I is the first image sequence, T is the first preset sampling period;

[0117] The first convolutional neural network model is generated based on the convolutional neural network model. The performance evaluation of the first convolutional neural network model is obtained through the cross entropy loss function. The cross entropy loss function is expressed by formula (2). Formula (2) is as follows:

[0118]

[0119] In formula (2), L is the loss function value, y i The true category of the sample bird data during the training of the first convolutional neural network model, Probability distribution for prediction of sample bird data during training of the first convolutional neural network model;

[0120] The thermal infrared information is input into the second convolutional neural network model for feature extraction, and the second thermal infrared feature information is expressed by formula (3). Formula (3) is as follows:

[0121] F ir =f cnn (H; θ ir );

[0122] In formula (3), F ir is the second thermal infrared characteristic information, f cnn is the feature extractor, θ ir is the learning parameter of the second convolutional neural network model, H is the thermal infrared information, T is the first preset sampling period.

[0123] In this embodiment, a feature extractor is used to obtain a first image sequence In order to fully explore the spatial structure information in the image, the system uses convolutional neural network (CNN) to extract features from the first image sequence:

[0124] F img =fcnn(I;θ cnn )

[0125] Among them, f cnn represents the feature extractor, θ cnn is the learning parameter of the first neural network model, F img is the first feature information extracted.

[0126] The first neural network model achieves high-precision recognition of birds such as egrets, sparrows, and crows by extracting spatial features from images. When using the first neural network model for image recognition, the cross-entropy loss function is used to evaluate the performance of the model, which is defined as follows:

[0127]

[0128] where y i is the real category, is the probability predicted by the model.

[0129] In this formula, L represents the cross-entropy loss function. It is a commonly used loss function for evaluating the performance of classification models. In this formula, L calculates the probability of the model prediction. and the true category y i The difference between.

[0130] Specifically, for each sample i, y i is the true category of the sample, and is the probability distribution of the model's prediction for this sample. It means taking the logarithm of the predicted probability to convert it to a positive value. Then, the logarithmic probability of each true category is multiplied by -1. The negative sign is taken to convert the loss value to a positive value. Finally, the loss values ​​of all samples are summed up to get the overall loss function value L.

[0131] The cross entropy loss function can be used for multi-classification problems. By minimizing this loss function, the model can better fit the training data and improve the performance of the classification model.

[0132] The system uses the classic CNN architecture (such as ResNet, DenseNet, etc.) as the backbone network, and performs targeted fine-tuning on farmland scene data to improve the scene adaptability of feature representation.

[0133] In terms of thermal infrared modality, the system uses a thermal imager to obtain thermal infrared information sequences of farmland scenes.

[0134] Different from visible light images, thermal infrared information mainly reflects the temperature distribution information on the surface of an object, and has unique advantages in capturing biological features (such as flying birds). The system uses a similar CNN architecture to extract the feature representation of the second feature information:

[0135] F ir =f cnn (H; θ ir );

[0136] Among them, θir is the learning parameter of the second neural network model. Considering the difference in imaging principle between thermal infrared information and the first image, the system makes adaptive adjustments in the design of feature extractor, such as using smaller convolution kernel size and shallower network layers to adapt to the low resolution and high noise characteristics of thermal infrared images.

[0137] See also Figure 1 In some embodiments, the bird attribute semantic dictionary is obtained by the following steps:

[0138] Construct a semantic dictionary. The semantic dictionary is expressed by formula (4), which is as follows:

[0139] D=d 1 ,d 2 ,…,d N ;

[0140] In formula (4), D is the semantic dictionary, d i is an attribute word, which represents an attribute of a bird;

[0141] Match a semantic embedding vector to each attribute word to form an attribute embedding matrix and a semantic representation space;

[0142] The attribute embedding matrix is ​​expressed by formula (5), which is as follows:

[0143] E=[e 1 ,e 2 ,…,e N ];

[0144] In formula (5), E is the attribute embedding matrix, e i is a semantic embedding vector, which represents the position of the attribute word in the semantic representation space.

[0145] In this embodiment, in order to further enhance the semantic representation ability of the flying bird target, the system also introduces the flying bird attribute information based on prior knowledge. By constructing the semantic dictionary D = d 1 ,d2 ,…,d N , each attribute word d i Corresponding to a semantic embedding vector e i , the system uses the attribute embedding matrix E = [e 1 ,e 2 ,…,e N ]Map discrete attributes into a continuous semantic representation space.

[0146] Specifically, the system enhances the semantic representation ability of the bird target by introducing bird attribute information based on prior knowledge: First, the system constructs a semantic dictionary of bird attributes, which contains N attribute words d i , each attribute word corresponds to a semantic embedding vector e i These attribute words can describe the characteristics or features of the bird, such as color, size, wing shape, etc.

[0147] Next, the system uses the attribute embedding matrix E = [e 1 ,e 2 ,...,e N ], mapping discrete attributes to continuous semantic representation space. The attribute embedding matrix is ​​a matrix composed of semantic embedding vectors corresponding to attribute words, and each column represents a semantic embedding vector corresponding to an attribute word. By mapping discrete attributes to continuous semantic representation space, the system can better express the attribute characteristics of the flying bird target.

[0148] In this way, the system can describe the flying bird target more accurately and comprehensively in the semantic representation space. With the help of the attribute embedding matrix, the system can transform discrete attribute information into continuous semantic representation, thereby better capturing the characteristics and features of the flying bird target. This enhanced semantic representation capability can help the system better understand and handle tasks and problems related to the flying bird target.

[0149] Based on this, the logical steps for constructing a semantic dictionary of bird attributes can refer to the following example:

[0150] 1. Collect information about the attributes of flying birds as prior knowledge;

[0151] 2. Based on the collected attribute information, a semantic dictionary D is constructed, where each attribute word d i Represents a flying bird attribute;

[0152] 3. For each attribute word d i , assigning it a corresponding semantic embedding vector e i , the vector represents the position of the attribute in the semantic space;

[0153] 4. For each attribute word d iIts corresponding semantic embedding vector e i Pair them and form the attribute embedding matrix E = [e 1 ,e 2 ,…,e N ];

[0154] 5. Use the attribute embedding matrix E to map the discrete bird attributes into a continuous semantic representation space, thereby enhancing the semantic representation capability of the bird target.

[0155] Through the above logical steps, a semantic dictionary of bird attributes can be constructed, in which each attribute word forms a corresponding relationship with its corresponding semantic embedding vector, thereby enhancing the semantic representation ability of the bird target.

[0156] See also Figure 1 In some embodiments, the first image feature information, the second thermal infrared feature information, and the third attribute semantic vector are input into the multimodal gated fusion unit for feature interaction and fusion, and the semantically enhanced multimodal fusion feature sequence is expressed by formula (6) to formula (8). Formula (6) is as follows:

[0157] F′ img =F img +f att (F img ,F ir ,E)⊙F ir ;

[0158] Formula (7) is as follows:

[0159] F′ ir =F ir +f att (F ir ,F img ,E)⊙F img ;

[0160] Formula (8) is as follows:

[0161] F fused =W 1 F′img+W 2 F′ ir +W 3 (F′ img ⊙F′ ir );

[0162] In formulas (6) to (8), F' img is the first feature information after fusion, F img is the first image feature information, f att is the first attention mechanism function, E is the attribute embedding matrix, F' ir is the second feature information after fusion, Ffused is the first fusion feature information, W 1 , W 2 , W3 is the learning parameter of fusion gating.

[0163] In this embodiment, after obtaining the first image feature information, the second thermal infrared feature information, and the third semantic attribute vector, the system realizes the interaction and fusion of different modal information through a multimodal feature fusion module.

[0164] The specific example can be used to understand: Construct a multimodal fusion algorithm model based on the attention mechanism:

[0165]

[0166] Among them, f att W represents the first attention mechanism function, which is used to dynamically calculate the correlation weights between different modal features. 1 ,W 2 ,W3 is the learning parameter of fusion gating, which is used to control the importance of different modal features. Through the attention mechanism and gating unit, MGFU can adaptively adjust the interaction strength of different modal information and generate semantically enhanced cross-modal fusion feature representation F fused .

[0167] See also Figure 1 In some embodiments, the multimodal fusion feature sequence is input into a long short-term memory network to model the temporal dependency of bird activities to obtain a first temporal model, the first temporal model includes first bird temporal information, the first bird temporal information includes the activity trajectory of the bird in the first time series and the behavior pattern includes:

[0168] The first long short-term memory network is used to encode the multimodal fusion feature sequence and expressed by formula (9):

[0169]

[0170] In formula (9), h t and c t They represent the hidden state and memory cell state of the first long short-term memory network at the tth moment, respectively, and f lstm is the first long short-term memory network function, θ lstm are the parameters of the first long short-term memory network;

[0171] The first time series model is trained using a multi-task learning mode to obtain a trained first time series model. The training process is expressed by formula (10):

[0172]

[0173] In formula (10), is the total loss function value, represents the loss function of the bird detection task, represents the loss function of the bird tracking task, represents the loss function of the bird behavior recognition task, λ 1 , 2 and λ 3 To balance the weight coefficients of different tasks.

[0174] In this embodiment, based on the fusion of multimodal features, the system further introduces the first time series model to capture the continuity and correlation of bird activities in the time dimension. The system uses a long short-term memory network (LSTM) to encode the fused feature sequence:

[0175]

[0176] Among them, h t and c t They represent the hidden state and memory cell state of LSTM at the tth moment respectively, capturing the dependency of bird activities in the time dimension.

[0177] In multimodal feature fusion, the system integrates features from different sensors or modalities to obtain more comprehensive and rich information. These features may include different types of data such as images, sounds, and motion. In order to achieve time series modeling, the system selected LSTM as the encoder. LSTM is a recursive neural network that can process and model time series data. By inputting the fused feature sequence into the LSTM network, the system can learn and capture the time series patterns and correlations in the sequence. Through the encoding process of LSTM, the system can convert the fused feature sequence into a hidden state representation with time series modeling capabilities. The hidden state can be used for subsequent tasks, such as predicting the behavior of flying birds, identifying the state of flying birds, etc.

[0178] In summary, by introducing the time series modeling module and using LSTM to encode the fused feature sequence, this embodiment enables the system to better capture the continuity and correlation of bird activities in the time dimension, thereby improving the understanding and prediction capabilities of bird behavior.

[0179] Specifically, during the training phase, the system adopts a multi-task learning paradigm to simultaneously optimize the objective functions of tasks such as bird detection, tracking, and behavior recognition:

[0180]

[0181] in, and They represent the loss functions of the bird detection task, bird tracking task, and bird behavior recognition task, respectively, and λ1 , 2 and λ 3 To balance the weight coefficients of different tasks. Through end-to-end joint optimization, the system achieves deep fusion of cross-modal information and collaborative enhancement between tasks.

[0182] In addition, to further improve the robustness of the model to different environmental changes, the system also introduces data enhancement strategies, such as random cropping, rotation, scaling and other image transformations, as well as noise simulation of thermal infrared information. At the same time, the system also uses the method of transfer learning, using the CNN model pre-trained on large-scale datasets such as ImageNet to initialize the feature extractor, and fine-tuning it on farmland monitoring data to alleviate the problem of small sample learning.

[0183] This embodiment uses a variety of sensors to obtain information about farmland scenes, including images, thermal infrared, and semantic attributes. The system uses deep learning methods to extract and fuse features to achieve accurate perception of bird activities. Specifically, the system uses a convolutional neural network to extract features from images and thermal infrared data, and uses attention mechanisms and gating units to achieve fusion of multimodal information, and then encodes the fused features through a time series modeling module. During the training phase, the system optimizes tasks such as bird detection, tracking, and behavior recognition through multi-task learning. In addition, the system also uses methods such as data enhancement and transfer learning to improve the robustness and generalization ability of the model.

[0184] See also Figure 2 In some embodiments, the second image sequence is input into the third convolutional neural network model for feature extraction to obtain fourth image feature information, the bird category information is obtained based on the fourth image feature information, and the fourth bird activity trajectory and the fourth bird behavior pattern are obtained according to the second image sequence, including:

[0185] The second attention mechanism function is added to the third convolutional neural network model and expressed by formula (11). Formula (11) is as follows:

[0186] F = A⊙X;

[0187] In formula (11), F is the attention enhancement feature in the second attention mechanism function, A is the attention weight matrix, X is the convolution feature map in the third convolutional neural network model, and ⊙ is element-wise multiplication;

[0188] The attention weight matrix is ​​expressed by formula (12), which is as follows:

[0189]

[0190] In formula (12), W a and W xis the learning parameter in the second attention mechanism function.

[0191] In this embodiment, in terms of bird recognition, the system is based on a convolutional neural network (CNN) and designs a bird recognition model for fine-grained classification. The model draws on the classic CNN structure (such as VGG, ResNet, etc.), and uses multi-layer convolution and pooling operations to extract hierarchical feature representations of the second image sequence. At the same time, in order to further enhance the model's ability to depict subtle recognition features, an attention mechanism is introduced at the end of the network. By learning the distribution of saliency weights of different parts of birds (such as the head, wings, tail, etc.), the recognition model can adaptively allocate the importance of features and focus on the most discriminative local areas, thereby improving the accuracy of fine-grained classification.

[0192] Specifically, the second attention mechanism function adaptively weights the convolution feature map X by introducing the weight matrix A to obtain the attention-enhanced feature representation F:

[0193] F = A⊙X;

[0194] Where ⊙ represents element-wise multiplication. The attention weight matrix A is obtained by learning the similarity between each position of the feature map and the standard saliency template:

[0195]

[0196] W a and W x is the learning parameter in the second attention mechanism function. Through end-to-end joint optimization, the attention mechanism and convolutional feature extraction are synergistically enhanced, which enables the recognition model to achieve significant performance improvement in the fine-grained classification task of birds.

[0197] See also Figure 2 In some embodiments, encoding is performed according to the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information to obtain a second time series model, the second time series model includes the second bird time series information, and the second bird time series information includes the activity trajectory and behavior pattern of the bird in the second time series, including:

[0198] The second long short-term memory network is used to encode the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information, and is expressed by formulas (13) to (18). Formula (13) is as follows:

[0199] i t =σ(W ii x t +W hi h t-1 +b i );

[0200] Formula (14) is as follows:

[0201] f t =σ(W if xt+W hf h t-1 +b f );

[0202] Formula (15) is as follows:

[0203] o t =σ(W io x t +W ho h t-1 +b o );

[0204] Formula (16) is as follows:

[0205]

[0206] Formula (17) is as follows:

[0207]

[0208] Formula (18) is as follows:

[0209] h t ′=o t ⊙tanh(C t );

[0210] In formulas (13) to (18), i t is the input gate in the second LSTM network, f t is the forget gate in the second long short-term memory network, o t is the output gate in the second LSTM network, C t is the cell state in the second long short-term memory network, is the candidate cell state in the second LSTM network, h t ' is the hidden state in the second long short-term memory network, W ii , W hi is the weight matrix of the input gate, b i is the bias vector of the input gate, W if , W hf is the weight matrix of the forget gate, b f is the bias vector of the forget gate, W io , W ho is the weight matrix of the output gate, b o is the bias vector of the output gate;

[0211] The second time series model is trained using a multi-task learning mode to obtain a trained first time series model. The training process is expressed by formula (19):

[0212] L'=λ' 1 L recog +λ' 2 L' behav ;

[0213] In formula (19), L′ is the total loss function value, is the loss function for the bird recognition task, L' behav is the loss function for the bird behavior understanding task, λ' 1 and λ' 2 is the weight coefficient to balance the two tasks.

[0214] In this embodiment, in terms of understanding the behavior of flying birds, the system designs a behavior analysis framework based on spatiotemporal sequence modeling. For the continuous motion sequence of flying birds, the system uses a long short-term memory network (LSTM) to model its temporal dependency. The second long short-term memory network can effectively capture the long-distance dependency of flying birds in the time dimension by introducing a gating mechanism (including input gate, forget gate and output gate), thereby characterizing its complex dynamic behavior patterns such as flight trajectory and feeding rhythm. Given a flying bird motion sequence x 1 ,x 2 ,…,x T ,LSTM gradually updates the hidden state h t To integrate historical information:

[0215] i t =σ(W ii x t +Whhih t-1 +b i );

[0216] f t =σ(W if xt+W hf h t-1 +b f );

[0217] o t =σ(W io x t +W ho h t-1 +b o );

[0218]

[0219] h t =o t ⊙tanh(Ct );

[0220] Among them, i t 、f t , o t They represent the input gate, forget gate and output gate in the second long short-term memory network, respectively. t It is the memory cell in the second long short-term memory network. The second long short-term memory network realizes the adaptive modeling of the long-term and short-term dependencies in the bird behavior sequence by dynamically adjusting the states of each gate. In the specific application scenario of expulsion, the system also introduces the attention mechanism in a targeted manner. By learning the importance weights of the behavior states at different times to the expulsion decision, it further enhances the pertinence and explanatory nature of the behavior analysis. In terms of model training, the system adopts an end-to-end joint optimization strategy. The bird recognition and behavior understanding models share the underlying convolutional feature extraction network to achieve full information interaction in the feature space. At the same time, the system introduces the paradigm of multi-task learning, which connects the recognition and behavior understanding tasks in series through a shared feature representation layer, and constructs a joint loss function:

[0221] L'=λ' 1 L recog +λ' 2 L' behav

[0222] in, and L' behav Respectively represent the loss functions of bird recognition and behavior understanding tasks, λ' 1 and λ' 2 To balance the weight coefficients of the two tasks. Through end-to-end back-propagation optimization, the recognition model and the behavior analysis model achieve synergistic enhancement, and the intrinsic relationship between appearance features and behavior patterns is mined from the data level.

[0223] In this embodiment, in order to further improve the robustness and generalization performance of the model, the system also introduces strategies such as data augmentation and transfer learning during the training process. Data augmentation expands the diversity of training samples and enhances the model's adaptability to changes in lighting, viewing angles, etc. by randomly rotating, scaling, and cropping the original bird images. Transfer learning uses CNN models (such as ResNet, etc.) pre-trained on large general data sets such as ImageNet to initialize parameters and fine-tune them in the bird-repelling scenario, which significantly speeds up the convergence of the model and improves performance.

[0224] This embodiment extracts the feature representation of the bird image through multi-layer convolution and pooling operations, and introduces the attention mechanism to enhance the ability to depict subtle features. In terms of bird recognition, the model focuses on the most discriminative local areas through the attention mechanism, which improves the accuracy of fine-grained classification. In terms of bird behavior understanding, the model uses LSTM to model the temporal dependency of the bird motion sequence to capture the dynamic behavior pattern of the bird. At the same time, the attention mechanism is introduced to learn the importance weight of the behavioral state to the expulsion decision, which enhances the pertinence and interpretability of the behavior analysis. The model training adopts an end-to-end joint optimization strategy, shares the convolutional feature extraction network, and introduces the paradigm of multi-task learning to balance the recognition and behavior understanding tasks. In order to improve the robustness and generalization performance of the model, the system also adopts strategies such as data enhancement and transfer learning. Experimental results show that the system has achieved good performance in the tasks of fine-grained bird recognition and bird-repelling behavior warning.

[0225] In practical applications, through the collaborative design and optimization of the intelligent identification of flying birds and the behavior understanding model, this system has built a set of intelligent behavior analysis engines for farmland bird repellent. Experimental results show that the analysis engine has achieved a Top-1 accuracy of 92.5% in the fine-grained bird identification task, and also showed good performance in the foraging behavior warning of typical harmful birds such as egrets and sparrows, and achieved the capture of high-risk dynamics 2-3 observation cycles in advance on average.

[0226] See also Figure 3 In some embodiments, generating a final expulsion strategy according to the first timing model and the second timing model further includes:

[0227] An intelligent adjustment model is constructed, and observation information is input into the intelligent adjustment model to obtain an initial driving strategy, wherein the observation information includes environmental parameters, a first time series model, a second time series model, and bird category information;

[0228] Evaluate the expulsion effect of the initial expulsion strategy, optimize the initial expulsion strategy according to the expulsion effect evaluation, and obtain the final expulsion strategy;

[0229] The final expulsion strategy is expressed by formula (20), which is as follows:

[0230]

[0231] In formula (20), π * (s) is the final expulsion strategy, s is the observed state, is the device parameter set of the driving device used by the intelligent adjustment model, a is the driving action in the final driving strategy, Q(s,a;θ * ) is the state-action value function, θ * is the optimal parameter of the state-action value function.

[0232] In this embodiment, the system uses environmental parameters (such as temperature, humidity, wind speed, etc.), the first time series model, the second time series model, bird category information (species, quantity, distribution, activity pattern, etc.) and expulsion effect evaluation (such as expulsion rate, warning distance, duration, etc.) as input, and gradually optimizes the expulsion strategy through reinforcement learning agent (Agent). The observation state s of the agent is provided by the multi-source heterogeneous information fusion module, including multi-dimensional continuous state (environmental parameters) and discrete state (such as bird species). The action space of the agent It is composed of the parameters of each driving subsystem, such as laser power, ultrasonic frequency, flash brightness, etc., and executes actions Adjust various expulsion methods. The reward r fed back to the agent by the environment is given by the expulsion effect evaluation model, which comprehensively considers the effectiveness and energy consumption of the expulsion. The system uses the Deep Q-Learning algorithm based on value function approximation to train the agent. By continuously interacting with the environment, the state-action value function Q(s, a; θ) is updated, and finally the optimal expulsion strategy is learned.

[0233] See also Figure 3 In some embodiments, generating a final early warning strategy according to the first time series model and the second time series model further includes:

[0234] Construct multi-source time series data according to the first time series model and the second time series model, wherein the multi-source time series data includes an agricultural ecological calendar, a bird migration calendar, and a meteorological calendar;

[0235] Generate the macroscopic time series law of bird damage based on multi-source time series data, divide the farmland scene into multi-scale grids, and generate the spatial distribution heat map of farmland bird damage;

[0236] A multi-objective optimization model is constructed. The input data of the multi-objective optimization model are multi-source time series data, spatial distribution heat map, bird species diversity index and degree of human intervention. The output result of the multi-objective optimization model is the final early warning strategy.

[0237] In this embodiment, in order to accurately grasp the spatiotemporal dynamics of farmland bird damage, the system constructs a multi-scale and multi-granularity farmland bird sentiment perception model. In the time dimension, the system integrates multi-source time series data such as agricultural ecological calendar, bird migration calendar and meteorological calendar to depict the macroscopic temporal law of bird damage. Studies have shown that more than 90% of farmland bird damage in China is concentrated in spring (March-May) and autumn (September-November), especially in spring. By performing spatiotemporal clustering of bird monitoring data from 600 observation points across the country for 5 consecutive years, the system constructs a digital "bird calendar" for farmland bird damage, divides my country into 10 bird damage zones, including the Northeast, North China, Northwest, and the middle and lower reaches of the Yangtze River, and finely depicts the key phenological periods of high incidence of bird damage in each zone, with an average positioning accuracy of better than 3 days.

[0238] For example, in the spatial dimension, the system uses a multi-scale gridding method to divide the farmland area into nested grids (such as 100 meters, 500 meters, and 1 kilometer). In each grid, the system deploys an acoustic, optical, and multi-modal bird-repellent device, and is equipped with a bird monitoring unit (such as an infrared camera, sound sensor, etc.) to achieve real-time monitoring of the bird situation in the grid. Through the spatial interpolation and rule tree mapping of multi-grid bird data, the system can generate a spatial distribution heat map of farmland bird damage, which intuitively reflects the spatial differentiation characteristics of bird damage at different scales such as local, regional, and global. Measured data show that the bird damage heat map generated by this system has a spatial resolution of better than 80% at a 100-meter grid scale, which can effectively guide farmland zoning policies and the precise deployment of bird-repellent devices.

[0239] Based on the accurate perception of bird conditions in farmland, the system further constructed a farmland bird-repelling ecological control model based on intelligent optimization. According to different farming stages and regional characteristics, the model can adaptively adjust the working parameters of the sound, light and electricity multimodal bird-repelling device (such as frequency, intensity and duration of repelling), effectively controlling bird damage while minimizing the cost of bird repelling and ecological disturbance.

[0240] Specifically, the regulation model takes farmland bird monitoring data as input, and takes crop bird damage control rate, bird species diversity index, and human intervention rate as constraints to construct a multi-objective optimization problem:

[0241]

[0242] g i (x)≤0,i=1,2,…,ph j (x)=0,j=1,2,…,qx min ≤x≤x max Indicates that the vector function F(x) is to be minimized, where F(x) contains m objective functions f i(x). x is the parameter vector of the bird repellent device, F(x) is a vector function including multiple optimization objectives such as bird repellent cost and ecological impact, g i (x) and h j (x) are inequality and equality constraints, respectively. The model uses a multi-objective evolutionary algorithm (such as NSGA-III) to optimize the bird-repelling parameters, selects the optimal parameter combination that meets the constraints on the Pareto front, and sends it to each bird-repelling node for execution.

[0243] In areas with high incidence of bird damage in rice fields, the system conducted a test demonstration of the control model for a growing season. Through online parameter optimization once a week, the bird damage rate was effectively controlled below 5%, which was 3 percentage points lower than the conventional bird repellent method; while the population of more than 10 common farmland birds such as sparrows and egrets remained basically the same as before the experiment, and the Shannon diversity index remained at a high level of more than 2.5. Compared with manual bird repellent, intelligent control reduces labor input by more than 90% and reduces comprehensive costs by more than 60%.

[0244] Farmland bird-repelling strategy guided by bird activity rhythm: In order to further reduce the impact of bird-repelling behavior on bird physiological rhythms and population behavior and improve the ecological friendliness of bird-repelling, a new farmland bird-repelling strategy guided by bird activity rhythm is systematically and innovatively proposed. This strategy uses the key behavioral rhythms of birds such as circadian rhythms and seasonal migration as constraint boundaries, and focuses on repelling birds in the "ecological interval period" when bird activities are relatively infrequent, while isolation and induction are mainly used in the "ecological sensitive period" when bird activities are frequent, so as to minimize the interference of repelling and maintain the normal physiological rhythm of birds.

[0245] Studies have shown that the foraging activities of most farmland birds show a clear circadian rhythm, concentrated in 2 hours after sunrise and 1 hour before sunset. Based on this, the system designed a daytime bird-repelling strategy of "avoiding peaks and driving away", that is, increasing the frequency of driving away during the non-peak activity period of birds, and appropriately reducing the intensity of driving away during the peak period, regulating bird activities through biological clock guidance, and reducing bird droppings pollution. Measured data show that "avoiding peaks and driving away" reduces the amount of bird droppings in farmland during the day by 20-30%, and reduces the potential threat of bird enemies by more than 15%.

[0246] On a seasonal scale, the system formulates a "southward retreat, northward avoidance" bird-repelling strategy based on the migration patterns of birds and local conditions. Specifically, it means increasing the intensity of bird repelling in the 1-2 weeks before the southward migration period to induce them to migrate south in advance; and during the northward migration period, the bird flocks are avoided to the greatest extent through habitat isolation and alternative food sources, providing them with a buffer adaptation period. Monitoring data show that "southward retreat, northward avoidance" has caused farmland birds and migratory populations to form a staggered distribution in time and space, and the migratory bird feeding rate in farmland has decreased by more than 25%.

[0247] Farmland bird damage risk warning based on niche simulation: In order to further expand the application value of this system in regional farmland bird situation warning, the research team constructed a bird damage risk warning model based on niche simulation. The model is supported by bird occurrence survey data and remote sensing habitat element data, and constructs a niche model of bird species distribution through the maximum entropy principle, and combines bird damage sensitivity assessment to generate a bird damage risk distribution map at a regional scale.

[0248] The niche model estimates the species' preference for the environment through data on species occurrence sites and environmental variables. The model assumes that the environmental conditions of the known distribution sites of the species represent its niche requirements, so the environmental combination closest to the known distribution points is most likely to provide a suitable habitat for the species. This study selected 300 bird survey points and extracted more than 40 habitat elements such as elevation, vegetation index, water distance, and farmland area ratio within a 5-kilometer radius as model inputs. The distribution probability of each species was obtained through iterative optimization:

[0249]

[0250] Among them, y represents the presence or absence of species, x is the environmental variable vector, and f(x) is the niche function. Based on the species-habitat preference relationship, the model is spatially extrapolated to the entire study area to obtain a suitability probability map of species distribution. By weighted fusion of the distribution maps of each species, the spatial distribution of regional bird richness can be generated. Based on the bird richness distribution map, the system further introduces a bird damage sensitivity assessment model to map the bird damage risk of farmland in the study area. The sensitivity assessment model comprehensively scores farmland from three dimensions: exposure, hazard, and vulnerability, and obtains a bird damage risk index:

[0252] R = E × H" × V;

[0253] Among them, E represents the degree of farmland exposure to harmful birds, H" represents the ability of harmful birds to harm crops, and V represents the vulnerability of crops after being harmed. Through AHP hierarchical analysis and expert scoring, the system established a weight matrix of evaluation indicators in each dimension, and based on this, spatial mapping of bird damage risks in farmland in the study area was carried out.

[0254] IoT farmland bird monitoring system: In order to achieve real-time monitoring of farmland bird conditions and bird repelling process, a farmland bird monitoring system based on low-power wide-area Internet of Things (LPWAN) was designed. The system consists of a terminal perception layer, a near-field communication layer, and a cloud platform layer. It can monitor the dynamics of birds in farmland and the working conditions of bird repellent devices at any time, and transmit the monitoring big data to the cloud management platform in real time.

[0255] At the terminal perception layer, the system deploys a variety of bird monitoring units, such as infrared cameras, sound sensors, and feces collectors, on different carriers such as fields, agricultural machinery, and bird repellent devices to obtain multi-source data such as images, sounds, and feces of bird activities in farmland. Each monitoring unit is powered by solar energy and can work continuously for more than 2 years.

[0256] At the near-field communication layer, the system uses LPWAN communication protocols such as LoRa and NB-IoT to transmit the data collected by the monitoring unit to the regional aggregation node with low power consumption. A single aggregation node can cover a radius of 5-10 kilometers, and the transmission rate can reach 10-50Kbps. The aggregation node intelligently screens and compresses the received multi-source heterogeneous data, and transmits it back to the cloud platform through the 4G / 5G network.

[0257] At the cloud platform level, the system has built an intelligent cloud platform for monitoring farmland bird conditions, which stores, normalizes, and integrates the massive monitoring data uploaded from farmland, and provides intelligent application services such as data analysis and visualization. Users can view the farmland bird situation in real time through the Web or mobile App, and obtain bird damage warning information for the next 1-7 days. The cloud platform can also remotely diagnose faulty devices and push maintenance plans based on the working condition data of the bird repellent device.

[0258] See also Figure 5 In the second aspect, the present embodiment further provides an agricultural intelligent bird-repelling system based on AI recognition, which is applicable to the agricultural intelligent bird-repelling method of the first aspect, and the system includes a control unit, a communication module, an energy supply module, a bird-repelling device, an image acquisition module and an AI recognition module;

[0259] The control unit is used to execute the method described in the first aspect; the communication module includes at least one of a Bluetooth module, a wifi module, and an antenna, and the communication module is used to establish a communication connection with an external device; the energy supply module includes at least one of a solar panel, a micro wind turbine, and a battery pack, and the energy supply system is used to provide power; the bird-repelling device includes at least one of a laser transmitter, an ultrasonic generator, and an LED flash array, and the bird-repelling device is used to execute the final bird-repelling strategy; the image acquisition module includes a high-definition camera and an infrared imager, and the image acquisition module is used to obtain a first image sequence, a second image sequence, and thermal infrared information; the AI ​​recognition module is used to perform image processing on the first image sequence, the second image sequence, and the thermal infrared information, and the AI ​​recognition module is used to train and correct the first neural network model, the second neural network model, and the third neural network model.

[0260] The following is a specific application example based on the above system:

[0261] Laser and sound, light and electricity multi-modal collaborative repelling forms a comprehensive and three-dimensional farmland bird repelling protection system through the intelligent combination of laser, ultrasound, flash light and other repelling means. According to the physiological characteristics and behavioral habits of different birds, the system designs a multi-modal repelling strategy optimization framework based on reinforcement learning, which realizes the dynamic scheduling and synergistic efficiency of repelling means.

[0262] Laser repelling subsystem: Laser repelling uses a directional high-energy laser beam to repel flying birds through thermal effects and visual interference. The system uses a laser parameter optimization model based on the physiological characteristics of birds. By adjusting parameters such as laser wavelength, power density and pulse frequency, the repelling effect is maximized without causing permanent damage to the birds. Specifically, the system uses a green laser with a wavelength of 532nm. Its spectral distribution is highly matched with the visual sensitive area of ​​birds and can form significant visual stimulation. At the same time, the system uses pulse modulation technology to convert continuous laser into a frequency-variable pulse sequence, forming thermal effects and frightening effects within a safe threshold. Experimental data show that 532nm laser has a power density of 10-50mW / cm 2 At a power density of 1000 Hz, it can effectively drive away most farmland birds. Continuous irradiation for 5-10 seconds can drive away more than 95% of the flying birds.

[0263] Ultrasonic bird repellent subsystem: Ultrasonic bird repellent uses the sensitivity of birds to sound waves in a specific frequency band to form a highly directional sound wave beam through an ultrasonic speaker array to cause auditory interference to flying birds. The system uses an ultrasonic parameter optimization model based on the auditory characteristics of birds, and achieves accurate repelling of different birds by adjusting parameters such as ultrasonic frequency, sound pressure level and modulation method. Studies have shown that ultrasonic waves with a frequency between 20-30kHz have a significant repelling effect on most farmland pest birds (such as sparrows, egrets, etc.). The system uses a phased array-based beamforming technology, through which more than 40 ultrasonic transducers work together to form a directional controllable ultrasonic beam, which effectively covers a farmland area of ​​0.5-5 hectares. Field test data shows that at a sound pressure level of more than 120dB, the system has an effective repelling rate of 92% for sparrows and other birds, and the average repelling time is 2-3 seconds.

[0264] Flash light repelling subsystem: Flash light repelling bird uses the sensitivity of the bird's visual system to changes in light to generate high-frequency flashes through a flash array to create visual interference for flying birds. The system uses a flash parameter optimization model based on the visual characteristics of birds to achieve precise repelling of different birds by adjusting parameters such as flash frequency, brightness and wavelength. Studies have found that flash stimulation with a frequency of 5-10Hz has a significant repelling effect on most birds. The system uses a high-brightness LED array to achieve precise modulation of flash frequency and brightness through PWM control. Field deployment data shows that at a flash brightness of 500-1000lux, the system's bird repelling rate reaches more than 90%, and the average repelling distance is 50-100 meters.

[0265] To verify the effectiveness of multi-modal collaborative repelling, large-scale field tests were carried out in different seasons and different crop planting areas. The experimental plots cover typical farmland and orchards such as rice, grapes, and corn, with a cumulative observation sample size of more than 10,000 times. Test data show that in rice fields where egrets are rampant, the comprehensive efficiency of laser-ultrasound-flash ternary collaborative repelling reaches 98.5%, which is 15.2% higher than that of single-mode repelling; in wheat fields where sparrows frequently appear, the average duration of laser-ultrasound-flash collaborative repelling is more than 2 hours, which is more than 50 minutes longer than that of single mode. Compared with traditional manual repelling, the system saves more than 95% of labor time and reduces the crop loss rate by more than 5 percentage points. Through intelligent multi-modal collaborative repelling, the system can autonomously adjust the parameter combination of each repelling method according to the real-time status of bird damage, ensuring long-term repelling while minimizing energy consumption. The measured data showed that compared with the fixed parameter combination, the adaptive multi-modal driving strategy increased the driving efficiency by 10% and reduced the system energy consumption by 20-30%. From an eco-friendly perspective, strategic driving also significantly reduced the potential risk of harm to birds. The optimized regulation of laser and ultrasound reduced the incidence of stress behavior in birds to less than 5%.

[0266] Modular design is a method of dividing a complex system into several independent and reusable functional units, aiming to reduce system coupling and improve maintainability and scalability. This system adopts a layered modular architecture, dividing the underlying hardware, middle-level communication, and upper-level applications into relatively independent subsystems, which are further subdivided into modules such as perception, expulsion, edge computing, data transmission, and energy management.

[0267] Each functional module is connected by a standardized interface, adopting the plug-and-play design concept. This loosely coupled architecture allows each module to be independently designed, developed, tested and upgraded, greatly reducing the complexity of system integration. For example, the bird repellent device can be flexibly matched with laser, ultrasonic, flash and other repellent modules according to the differences in the protected objects; the edge computing module can be equipped with high-performance GPU, FPGA and other accelerators to meet the computing power requirements of the AI ​​algorithm; the data transmission module can flexibly select Zigbee, LoRa, NB-IoT and other communication protocols according to the communication distance and bandwidth requirements.

[0268] Furthermore, this system adopts the microservices architecture to split the business logic into fine-grained, loosely coupled service units. For example, bird sentiment awareness, bird damage warning, driving away strategy, equipment management and other functions are encapsulated as independent microservices, and communication and collaboration between services are achieved through lightweight RESTful APIs. The microservices architecture enables the system to flexibly respond to changes in demand, and the development team can choose the most suitable technology stack for a single service, improving development efficiency and system performance.

[0269] To ensure interoperability and interchangeability between modules, this system strictly follows the industrial general interface standards. At the hardware level, the industrial-grade M12 interface is used as the physical connection between modules. The interface has reliable waterproof and dustproof performance (IP67) and can adapt to harsh farmland environments. In terms of data transmission, the system uses the RS-485 bus protocol to achieve serial communication and multi-point interconnection between modules. RS-485 has good anti-interference and long-distance transmission capabilities, with a maximum transmission distance of up to 1,200 meters.

[0270] At the software level, the system has designed a unified API specification based on the RESTful architecture. Each microservice provides a standardized HTTP interface, which implements data serialization and deserialization in JSON format. This interface design based on Web standards enables services developed in different languages ​​and platforms to be seamlessly connected, greatly improving the convenience of system integration. At the same time, the system also introduces the API Gateway mode to provide a unified entry and access control for external calls, simplifying the complexity of interaction between clients and microservices.

[0271] Farmland bird repellent devices are usually deployed in areas far from the power grid, and stable energy supply is the prerequisite for ensuring the long-term operation of the system. Aiming at the application scenario of farmland, this system has built a set of diversified energy supply solutions, which comprehensively utilizes new energy supply technologies such as solar energy, wind energy, and environmental energy collection to minimize dependence on external power sources.

[0272] The system uses high-efficiency monocrystalline silicon solar panels as the main power generation unit. The maximum output power of a single panel can reach 50W, and the photoelectric conversion efficiency reaches more than 22%. Considering the wind conditions of the farmland environment, the system also integrates a micro wind turbine. Under the condition of an average wind speed of 3m / s, the output power of a single generator can reach 20W. The system uses the maximum power point tracking (MPPT) technology to adjust the working status of photovoltaics and wind turbines in real time, so that they always work at the optimal power output point, improving the energy conversion efficiency.

[0273] Under low illumination and breeze conditions, the output energy of solar and wind energy is low and it is difficult to support the continuous operation of the equipment. To this end, the system also integrates environmental energy harvesting technology to collect tiny energies such as bird activities, environmental vibrations, and electromagnetic infrared through piezoelectric, thermoelectric, and RF methods to provide supplementary power for low-power sensors and controllers. For example, the system has laid piezoelectric floor tiles around the bird repellent device to generate electrical energy using the mechanical vibration of birds landing. A single floor tile can generate an output power of 1.5mW under a load of 100g.

[0274] In order to balance the volatility of energy output and load demand, the system is equipped with an intelligent microgrid control unit. The unit uses the Model Predictive Control (MPC) algorithm, combined with historical power generation data, load forecasting, SOC estimation and other information, to optimize the power flow of solar energy, wind energy and energy storage, and minimize the probability of power shortage. At the same time, the control unit also supports multiple safety protection measures such as overcharge and over-discharge protection and short-circuit protection, which prolongs the service life of the energy storage battery.

[0275] Through modular design and multiple energy supply guarantees, the system achieves high flexibility, scalability and continuous operation capabilities. Field test data show that the system can continue to operate stably for more than 18 months with an equipment failure rate of less than 1% and an energy shortage probability of less than 0.1%.

[0276] The above example has the following benefits:

[0277] The system uses deep learning and multi-source heterogeneous information fusion technology to achieve high-precision perception and identification of farmland bird conditions. Experimental data show that the system's bird identification accuracy in complex farmland environments reaches 92.5%, significantly surpassing traditional methods. This precise identification capability provides a reliable decision-making basis for subsequent expulsion strategies, significantly improving the pertinence and effectiveness of prevention and control.

[0278] By integrating multiple repelling methods such as laser, ultrasound, and flashlight, the system overcomes the defect that a single repelling method is prone to avoidance. Field test data show that in the rice field environment, the comprehensive efficiency of multi-modal synergistic repelling is as high as 98.5%, which is 15.2% higher than that of a single mode. This synergistic effect greatly prolongs the duration of repelling, which can reach more than 2 hours on average.

[0279] The system innovatively proposed a driving strategy based on the rhythm of bird activity, and minimized the interference with the normal physiological rhythm of birds through measures such as "avoiding peak driving" and "evacuating from the south and avoiding the north". Monitoring data showed that this strategy reduced the amount of bird droppings in farmland by 20-30% during the day, while the number of bird populations and diversity index remained basically stable, reflecting a good ecological balance.

[0280] The optimization of the driving strategy based on reinforcement learning enables the system to have the ability to learn and adapt autonomously. Long-term operation data shows that compared with the fixed parameter combination, the adaptive multi-modal driving strategy reduces the system energy consumption by 20-30% while maintaining the driving efficiency. At the same time, the remote monitoring of the Internet of Things greatly reduces the need for manual intervention, and the operation and maintenance costs are reduced by more than 60% compared with traditional methods.

[0281] The system-integrated ecological niche model provides a scientific basis for early warning of regional farmland bird damage risks. In the 600 square kilometer test area, the model has a spatial positioning accuracy of more than 80% for high-risk areas, and can issue early warning information one month in advance, providing strong support for the macro-decision-making of agricultural management departments.

[0282] The layered modular architecture and standardized interface design make the system highly scalable and adaptable. Field tests show that the system can be quickly deployed and parameter tuned in different farmland environments, with an equipment failure rate of less than 1% and can run stably for more than 18 months.

[0283] Comprehensive evaluations show that the system can effectively control bird damage while saving more than 95% of labor time compared to traditional manual driving methods, and reduce crop loss by an average of 5 percentage points. Considering equipment investment and operation and maintenance costs, the system's return on investment in large-scale application scenarios is about 2-3 years.

[0284] As a typical application of smart agriculture, the system not only solves the specific problem of bird damage, but also provides a model for the refined management of farmland ecosystems. The large amount of farmland ecological data generated by the system provides valuable data support for agricultural scientific research and policy formulation, and effectively promotes the digital and intelligent transformation of agriculture.

[0285] In short, this intelligent agricultural bird-repellent system based on AI recognition has demonstrated significant advantages in technological innovation, eco-friendliness, economic benefits and industry driving force, and represents the future development direction of the field of agricultural pest control.

[0286] In summary, this technical solution achieves high-precision perception and recognition of farmland bird information by utilizing deep learning and multi-source heterogeneous information fusion technology, significantly improving the pertinence and effectiveness of the final early warning strategy. By constructing multi-source time series data based on the first time series model and the second time series model, the final expulsion strategy is formulated based on the bird activity rhythm, minimizing the interference with the normal physiological rhythm of birds. By integrating multiple expulsion methods such as laser, ultrasound, and flash, the duration of expulsion is greatly extended. At the same time, the expulsion strategy optimization based on the multi-objective optimization model enables the system to have autonomous learning and adaptability, greatly reducing the need for manual intervention, and reducing the operation and maintenance cost by more than 60% compared with traditional methods.

[0287] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of the present invention, the patent protection scope of the present invention cannot be limited thereby. All technical solutions generated by replacing or modifying equivalent structures or equivalent processes based on the essential concept of the present invention and using the contents recorded in the specification and drawings of the present invention, as well as directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are included in the patent protection scope of the present invention.

Claims

1. An agricultural intelligent bird-repelling method based on AI recognition, characterized in that: include: Acquire a first image sequence in a farmland scene, where the first image sequence is a plurality of first images in a first preset sampling period; Inputting the first image sequence into a first convolutional neural network model for feature extraction to obtain first image feature information, wherein the first image feature information includes at least one of a plurality of first bird activity trajectories and a plurality of first bird behavior patterns in the current first image sequence; and simultaneously acquiring thermal infrared information in the farmland scene, the thermal infrared information being a plurality of thermal infrared image sequences within a first preset sampling period; Inputting the thermal infrared information into a second convolutional neural network model for feature extraction to obtain second thermal infrared feature information, wherein the second thermal infrared feature information includes at least one of a plurality of second bird activity trajectories and a plurality of second bird behavior patterns in the current thermal infrared information; Acquire the bird attribute information associated with the farmland scene, construct a bird attribute semantic dictionary, map the attribute into a semantic vector through an attribute embedding matrix to obtain a third attribute semantic vector, wherein the third attribute semantic vector includes at least one of a plurality of third bird activity trajectories and a plurality of third bird behavior patterns in the current bird attribute information; Inputting the first image feature information, the second thermal infrared feature information and the third attribute semantic vector into a multimodal gated fusion unit for feature interaction and fusion to obtain a semantically enhanced multimodal fusion feature sequence; Inputting the multimodal fusion feature sequence into a long short-term memory network, modeling the temporal dependency of bird activities, and obtaining a first temporal model, wherein the first temporal model includes first bird temporal information, and the first bird temporal information includes activity trajectories and behavior patterns of the birds in the first time series; Acquire a second image sequence of the farmland scene within a second preset sampling period, where the second image sequence is a plurality of second images within the second preset sampling period; Inputting the second image sequence into a third convolutional neural network model for feature extraction to obtain fourth image feature information, obtaining bird category information based on the fourth image feature information, and obtaining a fourth bird activity trajectory and a fourth bird behavior pattern according to the second image sequence; Encoding is performed according to the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information to obtain a second time series model, wherein the second time series model includes second bird time series information, and the second bird time series information includes the activity trajectory and behavior pattern of the bird in the second time series; A final expulsion strategy and / or a final warning strategy is generated according to the first time series model, the bird category information, and the second time series model.

2. The agricultural intelligent bird-repelling method based on AI recognition according to claim 1 is characterized in that: The first image sequence is input into the first convolutional neural network model for feature extraction, and the first image feature information is obtained by formula (1), and the formula (1) is as follows: F img =f cnn (I;θ cnn ); In formula (1), F img is the first image feature information, f cnn is the feature extractor, θ cnn is the learning parameter of the first neural network model, I is the first image sequence, T is the first preset sampling period; The first convolutional neural network model is generated based on a convolutional neural network model, and the performance evaluation of the first convolutional neural network model is obtained by a cross entropy loss function, and the cross entropy loss function is expressed by formula (2), and the formula (2) is as follows: In formula (2), L is the loss function value, y i The true category of the sample bird data during the training of the first convolutional neural network model, Probability distribution for prediction of sample bird data during training of the first convolutional neural network model; The thermal infrared information is input into the second convolutional neural network model for feature extraction, and the second thermal infrared feature information is obtained by formula (3), which is as follows: F ir =f cnn (H;θ ir ); In formula (3), F ir is the second thermal infrared characteristic information, f cnn is the feature extractor, θ ir is the learning parameter of the second convolutional neural network model, H is the thermal infrared information, T is the first preset sampling period.

3. The agricultural intelligent bird-repelling method based on AI recognition according to claim 1 is characterized in that: The bird attribute semantic dictionary is obtained by the following steps: Construct a semantic dictionary, and the semantic dictionary is expressed by formula (4), and the formula (4) is as follows: D=d1,d2,...,d N ; In formula (4), D is the semantic dictionary, d i is an attribute word, and the attribute word represents an attribute of a bird; Matching a semantic embedding vector for each of the attribute words to form an attribute embedding matrix to form a semantic representation space; The attribute embedding matrix is ​​expressed by formula (5), which is as follows: E=[e1,e2,…,e N ]: In formula (5), E is the attribute embedding matrix, e i is a semantic embedding vector, which represents the position of the attribute word in the semantic representation space.

4. The agricultural intelligent bird-repelling method based on AI recognition according to claim 1 is characterized in that: The first image feature information, the second thermal infrared feature information and the third attribute semantic vector are input into the multimodal gated fusion unit, and the multimodal fusion feature sequence is obtained by formula (6) to formula (8), and the formula (6) is as follows: F img =F img +f att (F img ,F ir ,E)⊙F ir ; The formula (7) is as follows: F ir =F ir +f att (F ir ,F img ,E)⊙F img ; The formula (8) is as follows: F fused =W1F′img+W2F′ ir +W3(F′ img ⊙F′ ir ); In formulas (6) to (8), F' img is the first feature information after fusion, F img is the first image feature information, f att is the first attention mechanism function, E is the attribute embedding matrix, F' ir is the second feature information after fusion, F fused is the first fusion feature information, W1, W2, W3 are the learning parameters of fusion gating.

5. The agricultural intelligent bird-repelling method based on AI recognition according to claim 4 is characterized in that: The multimodal fusion feature sequence is input into the long short-term memory network, and the temporal dependency of the bird activities is modeled to obtain a first temporal model, wherein the first temporal model includes first bird temporal information, and the first bird temporal information includes the activity trajectory and behavior pattern of the bird in the first time series, including: The multimodal fusion feature sequence is encoded using the first long short-term memory network and expressed by formula (9): In formula (9), h t and c t They represent the hidden state and memory cell state of the first long short-term memory network at the tth moment, respectively, and f lstm is the first long short-term memory network function, θ lstm are the parameters of the first long short-term memory network; The first time series model is trained using a multi-task learning mode to obtain a trained first time series model. The training process is expressed by formula (10): In formula (10), is the total loss function value, represents the loss function of the bird detection task, represents the loss function of the bird tracking task, represents the loss function of the bird behavior recognition task, and λ1, λ2, and λ3 are weight coefficients for balancing different tasks.

6. The agricultural intelligent bird-repelling method based on AI recognition according to claim 5 is characterized in that: Inputting the second image sequence into a third convolutional neural network model for feature extraction to obtain fourth image feature information, obtaining bird category information based on the fourth image feature information, and obtaining a fourth bird activity trajectory and a fourth bird behavior pattern according to the second image sequence include: A second attention mechanism function is added to the third convolutional neural network model and expressed by formula (11), which is as follows: F = A⊙X; In formula (11), F is the attention enhancement feature in the second attention mechanism function, A is the attention weight matrix, X is the convolution feature map in the third convolutional neural network model, and ⊙ is element-wise multiplication; The attention weight matrix is ​​expressed by formula (12), which is as follows: In formula (12), W a and W x is the learning parameter in the second attention mechanism function.

7. The agricultural intelligent bird-repelling method based on AI recognition according to claim 6 is characterized in that: According to the fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information, encoding is performed to obtain a second time series model, wherein the second time series model includes second bird time series information, and the second bird time series information includes the activity trajectory and behavior pattern of the bird in the second time series, including: The fourth bird activity trajectory, the fourth bird behavior pattern and the bird category information are encoded using a second long short-term memory network and expressed by formulas (13) to (18). Formula (13) is as follows: i t =σ(W ii x t +W hi h t-1 +b i ); The formula (14) is as follows: f t =σ(W if xt+W hf h t-1 +b f ); The formula (15) is as follows: o t =σ(W io x t +W ho h t-1 +b o ); The formula (16) is as follows: The formula (17) is as follows: The formula (18) is as follows: h t ’=o t ⊙tanh(C t ); In formulas (13) to (18), i t is the input gate in the second LSTM network, f t is the forget gate in the second long short-term memory network, o t is the output gate in the second LSTM network, C t is the cell state in the second long short-term memory network, is the candidate cell state in the second LSTM network, h t ' is the hidden state in the second long short-term memory network, W ii , W hi is the weight matrix of the input gate, b i is the bias vector of the input gate, W if , W hf is the weight matrix of the forget gate, b f is the bias vector of the forget gate, W io , W ho is the weight matrix of the output gate, b o is the bias vector of the output gate; The second time series model is trained using a multi-task learning mode to obtain a trained first time series model. The training process is expressed by formula (19): L'=λ'1L recog +λ'2L' behav ; In formula (19), L' is the total loss function value, is the loss function for the bird recognition task, L' behav is the loss function of the bird behavior understanding task, and λ'1 and λ'2 are the weight coefficients for balancing the two tasks.

8. The agricultural intelligent bird-repelling method based on AI recognition according to claim 1, characterized in that: Generating a final expulsion strategy according to the first timing model and the second timing model also includes: Constructing an intelligent adjustment model, inputting observation information into the intelligent adjustment model to obtain an initial driving strategy, wherein the observation information includes environmental parameters, a first time series model, a second time series model, and the bird category information; Performing a driving effect evaluation on the initial driving strategy, optimizing the initial driving strategy according to the driving effect evaluation, and obtaining the final driving strategy; The final expulsion strategy is expressed by formula (20), which is as follows: In formula (20), π * (s) is the final expulsion strategy, s is the observed state, is the device parameter set of the driving device used by the intelligent adjustment model, a is the driving action in the final driving strategy, Q(s,a;θ * ) is the state-action value function, θ * is the optimal parameter of the state-action value function.

9. The agricultural intelligent bird-repelling method based on AI recognition according to claim 1, characterized in that: Generating a final early warning strategy according to the first time series model and the second time series model further includes: Constructing multi-source time series data according to the first time series model and the second time series model, wherein the multi-source time series data includes an agricultural ecological calendar, a bird migration calendar, and a meteorological calendar; Generate a macroscopic time series law of bird damage according to the multi-source time series data, divide the farmland scene into multi-scale grids, and generate a spatial distribution heat map of farmland bird damage; A multi-objective optimization model is constructed, wherein the input data of the multi-objective optimization model are the multi-source time series data, the spatial distribution heat map, the bird species diversity index and the degree of human intervention, and the output result of the multi-objective optimization model is the final early warning strategy.

10. An agricultural intelligent bird-repelling system based on AI recognition, characterized in that: The agricultural intelligent bird-repelling method according to any one of claims 1 to 9, wherein the system comprises a control unit, a communication module, an energy supply module, a bird-repelling device, an image acquisition module and an AI recognition module; The control unit is used to execute the method according to any one of claims 1 to 9; The communication module includes at least one of a Bluetooth module, a Wi-Fi module, and an antenna, and the communication module is used to establish a communication connection with an external device; The energy supply module includes at least one of a solar panel, a micro wind turbine and a battery pack, and the energy supply system is used to provide power; The bird-repelling device comprises at least one of a laser transmitter, an ultrasonic generator and an LED flash array, and the bird-repelling device is used to execute a final bird-repelling strategy; The image acquisition module includes a high-definition camera and an infrared imager, and the image acquisition module is used to obtain the first image sequence, the second image sequence and thermal radiation information; The AI ​​recognition module is used to perform image processing on the first image sequence, the second image sequence and the thermal radiation information, and the AI ​​recognition module is used to train and correct the first neural network model, the second neural network model and the third neural network model.

Images