Method and system for predicting the transmission path of pests and diseases

Abstract

The present invention provides a method and system for predicting pest and disease transmission paths, which relate to the field of computer technology. In the present invention, first, a global crop image and a local crop image are acquired; second, a first semantic encoding is performed on the global crop image to form a global crop image feature; then, a second semantic encoding is performed on the local crop image to form a local crop image feature; further, based on the global crop image feature, the local crop image features are separately semantically mined to form a crop-associated image feature; further, based on the corresponding crop-associated image features, the pest and disease transmission situation between two adjacent sub-regions is analyzed; finally, based on the pest and disease transmission situation between each two adjacent sub-regions, a predicted pest and disease transmission path is obtained. Based on the above method, the problem of relatively low accuracy in pest and disease transmission path prediction in the prior art can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and system for predicting the transmission path of pests and diseases. Background Art

[0002] The data of each link of the entire planting and breeding industry chain, including crop growth management, breeding, feeding and breeding, slaughtering and processing, sales and trading, etc., will be incorporated into the trusted data space of the planting and breeding industry to form data contribution income, and then the main players in the planting and breeding industry chain will be empowered through data integration and analysis feedback. Promote the innovation and circulation application of multi-source data integration, improve the digitalization level of early warning, supervision, governance and decision-making of the entire planting and breeding industry chain, optimize production processes, improve efficiency, reduce costs, and ensure food safety and quality, forming a positive cycle of the value of the industry's trusted data space. Among them, in agricultural scenarios, the impact of pests and diseases is very large and is an important factor affecting agricultural development. Therefore, pests and diseases need to be focused on control.

[0003] However, in the existing technology, for the control of pests and diseases in agriculture, relevant personnel generally conduct corresponding analysis and prediction based on experience to determine the transmission path of pests and diseases. However, due to the influence of subjective factors of relevant personnel and the limited experience, it is generally difficult to effectively and reliably predict the transmission path of pests and diseases, making it difficult to effectively control the spread of pests and diseases. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method and system for predicting the transmission path of pests and diseases, so as to improve the problem of relatively low accuracy of the prediction of the transmission path of pests and diseases in the prior art.

[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0006] A method for predicting the spread path of pests and diseases, comprising:

[0007] Acquire a global crop image corresponding to the crops in the target area and a local crop image corresponding to the crops in each sub-area included in the target area;

[0008] performing a first semantic encoding on the crop global image to form a crop global image feature;

[0009] performing a second semantic coding on each of the crop local images to form a crop local image feature corresponding to each of the crop local images, wherein the second semantic coding is different from the first semantic coding and includes multi-scale mining and fusion in the frequency domain;

[0010] Based on the crop global image features, performing association semantic mining on each of the crop local image features to form each crop associated image feature;

[0011] For every two adjacent sub-regions, based on the crop-related image features corresponding to the two adjacent sub-regions, the disease and insect pest transmission situation between the two adjacent sub-regions is analyzed;

[0012] Based on the pest and disease transmission conditions between every two adjacent sub-regions in the target region, a predicted pest and disease transmission path in the target region is obtained.

[0013] In some preferred embodiments, in the above-mentioned method for predicting the transmission path of pests and diseases, the step of performing second semantic encoding on each of the crop local images to form crop local image features corresponding to each of the crop local images includes:

[0014] For each of the crop local images, performing Fourier transforms of multiple sizes on the crop local image to form multiple crop local frequency spectra corresponding to the crop local image, wherein the multiple crop local frequency spectra have different sizes;

[0015] Semantically encoding each of the crop local spectrum graphs, and semantically enhancing the high-frequency features during the semantic encoding process to form high-frequency enhanced features of each crop;

[0016] Each of the crop high-frequency enhanced features is fused to form a crop local image feature.

[0017] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, the step of semantically encoding each of the crop local spectrum graphs and semantically enhancing the high-frequency features during the semantic encoding process to form the high-frequency enhanced features of each crop includes:

[0018] Masking low-frequency information on the crop local spectrum to form a crop high-frequency spectrum;

[0019] Performing depth convolution on the crop local spectrum map and the crop high-frequency spectrum map respectively to form crop local depth features corresponding to the crop local spectrum map and crop high-frequency depth features corresponding to the crop high-frequency spectrum map;

[0020] Based on the high-frequency depth features of the crop, semantic enhancement of the high-frequency features is performed on the local depth features of the crop to form a high-frequency enhanced feature of the crop corresponding to the local spectrum graph of the crop.

[0021] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, the step of semantically enhancing the high-frequency features of the local depth features of the crop based on the high-frequency depth features of the crop to form the high-frequency enhanced features of the crop corresponding to the local spectrum graph of the crop includes:

[0022] performing an upsampling operation on the current crop high-frequency depth feature according to the size of the previous crop high-frequency depth feature to form a crop high-frequency upsampled feature, wherein after sorting the multiple crop local spectrograms corresponding to the crop local image in a direction from large to small in size, the corresponding crop high-frequency depth features are sorted according to the sorting result to form a corresponding order relationship;

[0023] Based on the previous crop high-frequency depth feature, attention processing is performed on the crop high-frequency upsampled feature to form a crop high-frequency attention feature;

[0024] Based on the crop local depth feature, downsampling the crop high-frequency attention feature to form a crop high-frequency downsampling feature;

[0025] An importance parameter distribution is formed based on the high-frequency downsampling feature map of the crop, and based on the importance parameter distribution, importance weighting processing is performed on the local depth features of the crop, and the local depth features of the crop are connected to the output of the importance weighting processing to form the high-frequency enhanced features of the crop corresponding to the local spectrum map of the crop.

[0026] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, the step of fusing each of the crop high-frequency enhanced features to form crop local image features includes:

[0027] Traversing the plurality of crop high-frequency enhancement features in a direction from small to large in size of the plurality of crop local frequency spectra to form a currently traversed crop high-frequency enhancement feature;

[0028] If the currently traversed crop high-frequency enhancement feature belongs to the first crop high-frequency enhancement feature, then the crop high-frequency enhancement feature is used as the first local fusion feature;

[0029] If the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, then based on the previous local fusion feature, the currently traversed crop high-frequency enhancement feature is fused and output to form the current local fusion feature;

[0030] Based on the last formed local fusion feature, the crop local image feature is determined.

[0031] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, if the currently traversed crop high-frequency enhanced feature does not belong to the first crop high-frequency enhanced feature, the step of fusing and outputting the currently traversed crop high-frequency enhanced feature based on the previous local fusion feature to form the current local fusion feature includes:

[0032] If the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, then based on the size of the currently traversed crop high-frequency enhancement feature, the previous local fusion feature is upsampled to form a local upsampled feature, and the currently traversed crop high-frequency enhancement feature and the local upsampled feature are convoluted and fused to form a high-frequency convolution fusion feature;

[0033] forming an importance parameter distribution based on the high-frequency convolution fusion feature map, and performing importance weighting processing on the currently traversed high-frequency enhanced features of the crop based on the importance parameter distribution;

[0034] The currently traversed high-frequency enhanced features of the crop are connected to the output of the importance weighting process to form the current local fusion features.

[0035] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, the step of performing association semantic mining on each of the local crop image features based on the global crop image features to form each crop association image feature includes:

[0036] Performing multiple segmentations on the crop global image features to form multiple segmentation sets, wherein the number of crop segmentation image features included in each segmentation set is equal to the number of the crop local image features;

[0037] Based on the feature correlation between the crop segmentation image features and the crop local image features, determining a segmentation set that best matches the multiple crop local image features from the multiple segmentation sets, and determining the segmentation set as a target segmentation set;

[0038] Based on each crop segmentation image feature in the target segmentation set, attention processing is performed on each corresponding crop local image feature to form each crop-related image feature.

[0039] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, the step of performing a first semantic encoding on the crop global image to form crop global image features includes:

[0040] Performing Fourier transform on the crop global image to form a crop global spectrum map;

[0041] Performing depth convolution on the crop global spectrum map to form a crop global depth feature;

[0042] Self-attention processing is performed on the crop global depth features to form crop global image features.

[0043] In some preferred embodiments, in the above-mentioned pest and disease transmission path prediction method, the step of analyzing the pest and disease transmission situation between each two adjacent sub-regions based on the crop-related image features corresponding to the two adjacent sub-regions includes:

[0044] For every two adjacent sub-regions, performing a difference operation on the crop-related image features corresponding to the two adjacent sub-regions to obtain the crop image difference features;

[0045] Performing convolution fusion on the crop-related image features corresponding to two adjacent sub-regions and the crop image difference features to form crop convolution fusion features corresponding to the two adjacent sub-regions;

[0046] Performing full connection processing on the crop convolution fusion feature to form a crop fully connected feature, wherein the size of the crop fully connected feature is determined based on the number of pest and disease transmission types;

[0047] activating the fully connected features of the crop to form a propagation type probability distribution, wherein the probability values ​​in the propagation type probability distribution are used to represent the likelihood of a corresponding pest and disease propagation type, the pest and disease propagation type including propagation from the first sub-region to the second sub-region, propagation from the second sub-region to the first sub-region, and no propagation between the first sub-region and the second sub-region;

[0048] The pest and disease transmission situation between two adjacent sub-areas is determined based on the pest and disease transmission type corresponding to the probability with the maximum value in the transmission type probability distribution.

[0049] An embodiment of the present invention also provides a pest and disease transmission path prediction system, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above-mentioned pest and disease transmission path prediction method.

[0050] The pest and disease transmission path prediction method and system provided by the embodiment of the present invention first obtains a global crop image and a local crop image; secondly, performs a first semantic encoding on the global crop image to form a global crop image feature; then, performs a second semantic encoding on the local crop image to form a local crop image feature; further, based on the global crop image feature, performs association semantic mining on the local crop image features to form a crop association image feature; further, based on the corresponding crop association image feature, analyzes the pest and disease transmission situation between two adjacent sub-regions; finally, based on the pest and disease transmission situation between each two adjacent sub-regions, obtains the pest and disease transmission prediction path. Based on the above method, on the one hand, because the local crop images of the sub-regions are subjected to multi-scale mining and fusion in the frequency domain, the detailed information in the local crop images can be fully mined, and this detailed information is closely related to the growth of the crop (pests and diseases directly affect the growth). Therefore, based on the mined detailed information, the pest and disease situation can be reliably predicted. On the other hand, since the features of the sub-regions will also be associated and mined based on the global image of the crops in the target area, the obtained crop-associated image features can not only characterize the detailed information, but also characterize the overall situation of the crops. In this way, the reliability of the analyzed pest and disease transmission situation can be further improved. Therefore, the reliability of the pest and disease transmission prediction path obtained by further analysis can be guaranteed, thereby improving the relatively low accuracy of the pest and disease transmission path prediction in the existing technology, so that relevant personnel or equipment can focus on the pest and disease transmission sub-region according to the pest and disease transmission prediction path.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a structural block diagram of the pest transmission path prediction system provided by an embodiment of the present invention.

[0053] Figure 2 A schematic diagram of the various modules included in the pest transmission path prediction device provided in an embodiment of the present invention.

[0054] Figure 3 A flowchart of the steps of the method for predicting the transmission path of pests and diseases provided in an embodiment of the present invention.

[0055] Figure 4 A schematic diagram of a first semantic encoding provided by an embodiment of the present invention.

[0056] Figure 5 A schematic diagram of the second semantic coding provided by an embodiment of the present invention.

[0057] Figure 6 A schematic diagram of semantically enhancing high-frequency features during semantic encoding provided by an embodiment of the present invention.

[0058] Figure 7 A schematic diagram of semantic enhancement of high-frequency features provided by an embodiment of the present invention.

[0059] Figure 8 A schematic diagram of a pest and disease transmission network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0061] like Figure 1 As shown, an embodiment of the present invention provides a system for predicting the spread of pests and diseases. The system may include a memory and a processor.

[0062] Specifically, the memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, the electrical connection between them can be achieved through one or more communication buses or signal lines. The memory can store at least one software function module (such as a software module) in the form of software or firmware. Figure 2 The processor may be configured to execute the executable computer program stored in the memory, thereby implementing the pest transmission path prediction method provided by the embodiment of the present invention (as described below).

[0063] Optionally, the modules included in the pest transmission path prediction device may be:

[0064] An image acquisition module is configured to acquire a global crop image corresponding to the crops in the target area and a local crop image corresponding to the crops in each sub-area included in the target area; a first semantic coding module is configured to perform a first semantic coding on the global crop image to form a global crop image feature; a second semantic coding module is configured to perform a second semantic coding on each of the local crop images to form a local crop image feature corresponding to each local crop image, wherein the second semantic coding is different from the first semantic coding and the second semantic coding includes multi-scale mining and fusion in the frequency domain; an association semantic mining module is configured to perform association semantic mining on each local crop image feature based on the global crop image feature to form a related image feature of each crop; a pest and disease transmission analysis module is configured to analyze the pest and disease transmission situation between each two adjacent sub-areas based on the crop related image features corresponding to the two adjacent sub-areas; and a pest and disease transmission prediction module is configured to obtain a predicted pest and disease transmission path in the target area based on the pest and disease transmission situation between each two adjacent sub-areas in the target area.

[0065] Optionally, the memory may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), a system on a chip (SoC), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0066] and, Figure 1 The structure shown is for illustration only. The pest transmission path prediction system may also include Figure 1 More or fewer components than shown, or with Figure 1The different configurations shown, for example, may include a communication unit for exchanging information with other devices. In an alternative example, the pest transmission path prediction system may be a server.

[0067] Combine Figure 3 The present invention also provides a method for predicting pest transmission paths, which can be applied to the above-mentioned pest transmission path prediction system. The method steps defined in the process related to the pest transmission path prediction method can be implemented by the pest transmission path prediction system.

[0068] The following will Figure 3 The specific process shown is explained in detail.

[0069] Step S110 , obtaining a global crop image corresponding to the crops in the target area and a local crop image corresponding to the crops in each sub-area included in the target area.

[0070] In an embodiment of the present invention, the pest and disease transmission path prediction system can obtain a global crop image corresponding to the crops in the target area and a local crop image corresponding to the crops in each sub-area included in the target area. Exemplarily, the global crop image can be formed by long-distance shooting of the target area by airborne equipment, and is used to reflect the overall growth of the crops in the target area. The local crop image can be formed by close-up shooting of the sub-area by airborne equipment, and is used to reflect the growth of the crops in the sub-area. That is to say, compared with the global crop image, the local crop image can have more detailed information and can better characterize the pest and disease situation. In addition, the target area may include multiple sub-areas.

[0071] Step S120 : performing a first semantic coding on the crop global image to form a crop global image feature.

[0072] In an embodiment of the present invention, after obtaining the global crop image, the pest and disease transmission path prediction system can perform a first semantic encoding on the global crop image to form a global crop image feature, that is, to mine the semantic information in the global crop image to form a corresponding global crop image feature. It should be noted that, in an embodiment of the present invention, the specific form of expression of each feature formed can be a vector (or matrix).

[0073] Step S130 : performing second semantic coding on each of the crop partial images to form crop partial image features corresponding to each of the crop partial images.

[0074] In an embodiment of the present invention, after acquiring the crop local images, the pest and disease transmission path prediction system may perform a second semantic encoding on each of the crop local images to form a crop local image feature corresponding to each crop local image. The second semantic encoding is different from the first semantic encoding and includes multi-scale mining and fusion in the frequency domain to fully capture detailed information in the crop local images, thereby improving the ability of the generated crop local image features to represent the pest and disease situation.

[0075] Step S140 : Based on the crop global image feature, each of the crop local image features is subjected to associated semantic mining to form each crop associated image feature.

[0076] In an embodiment of the present invention, after forming the global crop image features and the local crop image features, the pest and disease transmission path prediction system can perform associated semantic mining on each of the local crop image features based on the global crop image features to form each crop associated image feature. It should be noted that although the local crop image features have a stronger representational capability in terms of detailed information, the global crop image features also carry semantic information related to the pest and disease situation that is not present in the local crop image features. Therefore, through associated semantic mining, the semantic enhancement of the local crop image features can be achieved, thereby improving the semantic representation accuracy of the formed crop associated image features.

[0077] Step S150 : for every two adjacent sub-regions, based on the crop-related image features corresponding to the two adjacent sub-regions, analyzing the disease and insect pest transmission situation between the two adjacent sub-regions.

[0078] In an embodiment of the present invention, after forming the crop-associated image features, the pest and disease transmission path prediction system can analyze the pest and disease transmission situation between each two adjacent sub-regions based on the crop-associated image features corresponding to the two adjacent sub-regions. For example, sub-region A and sub-region B have an adjacent relationship, and the pest and disease transmission situation between sub-region A and sub-region B can be analyzed based on the crop-associated image features corresponding to sub-region A and the crop-associated image features corresponding to sub-region B.

[0079] Step S160: obtaining a predicted path for the spread of pests and diseases in the target area based on the pest and disease spread between every two adjacent sub-areas in the target area.

[0080] In an embodiment of the present invention, after analyzing the pest and disease transmission situation, the pest and disease transmission path prediction system can obtain a predicted pest and disease transmission path in the target area based on the pest and disease transmission situation between each two adjacent sub-areas in the target area, that is, how pests and diseases spread between the sub-areas in the target area, for example, from sub-area A to sub-area B, and from sub-area B to sub-area C and sub-area D, respectively. In this way, after determining the pest and disease transmission path, focused pest and disease control is performed for crops that have not yet encountered pests and diseases, but have adjacent sub-areas that have encountered pests and diseases, and are located on the pest and disease transmission path. General pest and disease control is performed for crops that have not yet encountered pests and diseases, but have adjacent sub-areas that have encountered pests and diseases, and are not located on the pest and disease transmission path. Lighter pest and disease control is performed for crops that have not yet encountered pests and diseases, but have adjacent sub-areas that have encountered pests and diseases, and are not located on the pest and disease transmission path. In other words, based on the predicted path of pest and disease transmission, the intensity of pest and disease control in each sub-region can be determined, such as controlling the amount of pesticides applied, thus avoiding the problem of waste of resources.

[0081] Specifically, rice is one of the world's most important food crops, and rice planthoppers are a common rice field pest that usually spreads through air currents (making two adjacent sub-regions on the propagation path also physically adjacent sub-regions), causing significant damage to rice. Based on the above method, predicting the propagation path of rice planthoppers is of great significance for effectively controlling pests and increasing yields.

[0082] Based on the above method, on the one hand, multi-scale frequency domain mining and fusion of local crop images in sub-regions enable the full mining of detailed information within these images. This detailed information is closely related to crop growth (pests and diseases directly affect growth), and therefore, based on this mined detailed information, reliable predictions of pest and disease conditions can be made. On the other hand, since association mining of sub-region features is also performed based on the global crop image of the target region, the resulting crop-associated image features not only represent detailed information but also the overall crop condition. This further improves the reliability of the analyzed pest and disease transmission situation, thereby ensuring the reliability of the predicted pest and disease transmission paths obtained through further analysis. This improves the relatively low accuracy of pest and disease transmission path predictions in existing technologies, allowing relevant personnel or equipment to focus on controlling pest and disease transmission sub-regions based on the predicted pest and disease transmission paths.

[0083] In the first part, what needs to be explained about step S110 is that the specific process of obtaining the crop global image and the crop local image is not limited and can be selected according to actual conditions.

[0084] For example, in a feasible implementation, a communication connection may be directly established with an airborne device to obtain the currently collected global image and local image of the crop in real time.

[0085] For example, in another possible implementation, after the airborne device collects and forms a global image of the crop and a local image of the crop, it can first store them in a corresponding storage server, and then, in response to an instruction to perform path prediction, the global image of the crop and the local image of the crop can be obtained from the storage server.

[0086] In the second part, it should be explained with respect to step S120 that the specific process of performing the first semantic encoding on the crop global image is not limited and can be selected according to actual conditions.

[0087] For example, in one feasible implementation, the crop global image may be subjected to convolution processing to implement a first semantic encoding, thereby forming a crop global image feature.

[0088] For example, in another feasible embodiment, in order to improve the reliability of the representation of crop growth conditions by the formed global crop image features, the above-mentioned step S120 may further include step S121, step S122 and step S123, and the specific contents of each step are described as follows.

[0089] Step S121 , performing Fourier transform on the crop global image to form a crop global spectrum graph.

[0090] In an embodiment of the present invention, the crop global image may be subjected to Fourier transform to form a crop global spectrum map. It should be noted that, in an alternative embodiment, the result of the Fourier transform may be directly used as the crop global spectrum map. In another alternative embodiment, the result of the Fourier transform may be further processed, such as centralization, to form a crop global spectrum map. Specifically, in combination with Figure 4 In the result of Fourier transform, the edge area belongs to low-frequency information, and the middle area belongs to high-frequency information. Through centralization processing, the low-frequency information can be concentrated in the middle area, and the high-frequency information can be concentrated in the edge area.

[0091] Step S122: performing depth convolution on the crop global spectrum map to form a global depth feature of the crop.

[0092] In the embodiment of the present application, after forming the crop global spectrum map, the crop global spectrum map can be subjected to deep convolution to form a global crop depth feature. The deep convolution can be implemented by a convolutional neural network model in the training process. The specific network architecture is as follows:

[0093] Convolution layer 1: The first convolution layer uses multiple 3x3 convolution kernels (for example, 32 convolution kernels, and the activation function is ReLU) to perform convolution operations on the input spectrum map; Pooling layer 1: After convolution, perform maximum pooling (MaxPooling), the pooling size is 2x2, and the step size is 2; Convolution layer 2: The second convolution layer uses a slightly larger convolution kernel (for example, the size is 5x5, the number is 64, and the activation function is ReLU) to further convolve the feature map of the previous layer; Pooling layer 2: Perform the second maximum pooling (MaxPooling), the pooling size is 2x2, the step size is 2, and the feature representation ability is further reduced; Global average pooling layer: In order to extract global features, perform global average pooling operation; this step reduces the feature map to a vector; Fully connected layer: The feature vector after global pooling is input into the fully connected layer for extracting the final global deep features. The number of neurons can be 128 or 256, and the activation function is ReLU.

[0094] Step S123 , performing self-attention processing on the crop global depth feature to form a crop global image feature.

[0095] In an embodiment of the present invention, after forming the global depth feature of the crop, the global depth feature of the crop can be self-attention processed to form a global image feature of the crop, that is, the associated semantic information within the global depth feature of the crop is mined, so that important semantic information is represented in a focused manner, thereby obtaining a global image feature of the crop with better representation capability.

[0096] In the third part, it is necessary to explain step S130 that the specific process of performing association semantic mining on each crop local image feature is not limited and can be selected according to actual conditions.

[0097] For example, in a feasible implementation, attention processing (cross attention) may be performed on each of the crop local image features based on the crop global image features, thereby forming each crop-related image feature.

[0098] For example, in another feasible implementation, in order to take into account that the semantic information corresponding to the sub-regions may have certain similarities, and therefore to avoid a certain degree of interference caused by the semantic information corresponding to other sub-regions, the above-mentioned step S130 may further include step S131, step S132 and step S133, and the specific contents of each step are as follows.

[0099] Step S131 : performing multiple segmentations on the crop global image features to form multiple segmentation sets.

[0100] In an embodiment of the present invention, the global crop image features can be subjected to multiple segmentation processes to form multiple segmentation sets. Each segmentation set includes a number of crop segmentation image features equal to the number of local crop image features. It should be noted that the multiple segmentation processes may all be sliding window segmentation (i.e., the resulting crop segmentation image features have intersections), or none of the multiple segmentation processes may be sliding window segmentation (i.e., the resulting crop segmentation image features do not have intersections), or some may be sliding window segmentation while others may not. The specific selection can be based on actual needs.

[0101] Step S132 : Based on the feature correlation between the crop segmentation image features and the crop local image features, a segmentation set that best matches the multiple crop local image features is determined from the multiple segmentation sets, and is determined as a target segmentation set.

[0102] In an embodiment of the present invention, after forming the multiple segmentation sets, a segmentation set that best matches the multiple crop local image features is determined from the multiple segmentation sets based on the feature correlations between the crop segmentation image features and the crop local image features, and is determined as the target segmentation set. Specifically, for each segmentation set, the crop segmentation image features in the segmentation set are matched one-to-one with each of the crop local image features (e.g., the mean of the cosine similarities between each crop segmentation image feature and the corresponding crop local image feature has the maximum value, i.e., among various pairing combinations, the pairing combination with the maximum mean value is selected). In this way, the crop local image feature that best matches each crop segmentation image feature is determined. The mean of the corresponding feature similarities (e.g., cosine similarities) can then be used as the representative similarity for the segmentation set. Furthermore, discrete values ​​of the feature similarities for the segmentation set can be calculated, the reciprocal of the discrete values ​​(or other negatively correlated values) used as adjustment coefficients, and the product of the representative similarity and the adjustment coefficients is calculated to obtain matching parameters for the segmentation set. Finally, the segmentation set with the maximum matching parameter can be selected as the target segmentation set.

[0103] Step S133 : Based on each crop segmentation image feature in the target segmentation set, attention processing is performed on each corresponding crop local image feature to form each crop-related image feature.

[0104] In this embodiment of the present invention, after determining the most reliable target segmentation set, attention processing can be performed on each corresponding crop local image feature based on each crop segmentation image feature in the target segmentation set to form each crop-associated image feature. For example, if crop segmentation image feature 1 corresponds to (i.e., matches) crop local image feature 1, cross-attention processing can be performed on crop local image feature 1 based on crop segmentation image feature 1 to form a crop-associated image feature.

[0105] The fourth part, regarding step S140 , needs to be explained that the specific process of performing the second semantic coding on each crop partial image is not limited and can be selected according to actual conditions.

[0106] For example, in one feasible implementation, the local crop image can be Fourier transformed to form a corresponding spectrum map, and then the spectrum map can be convolved at multiple depths. Finally, the features of the convolution processing at multiple depths can be fused to form the corresponding local crop image features.

[0107] For example, in another achievable embodiment, in order to fully mine the semantic information that can effectively represent the pest and disease situation during the semantic coding process, the above step S140 can further include step S141, step S142 and step S143. The specific content of each step is as follows, which can be combined with Figure 5 The content shown.

[0108] Step S141 : For each of the crop local images, perform Fourier transforms of multiple sizes on the crop local image to form multiple crop local frequency spectra corresponding to the crop local image.

[0109] In an embodiment of the present invention, a Fourier transform of various sizes (e.g., 256, 512, 1024, etc., which determines the resolution of the Fourier transform and can be selected as a power of 2) can be performed on each crop local image to generate multiple crop local spectrograms corresponding to the crop local image. The multiple crop local spectrograms have different sizes. It should be noted that the generated crop local spectrograms can be either uncentered or centered.

[0110] In step S142 , semantic coding is performed on each of the crop local spectrograms, and semantic enhancement is performed on the high-frequency features during the semantic coding process to form high-frequency enhanced features of each crop.

[0111] In an embodiment of the present invention, after forming the multiple crop local spectrograms, semantic encoding is performed on each of the crop local spectrograms, and during the semantic encoding process, high-frequency features are semantically enhanced to form each crop high-frequency enhanced feature. That is, each crop local spectrogram corresponds to one crop high-frequency enhanced feature. It should be noted that the high-frequency information corresponding to the high-frequency features can effectively represent some edge information in the crop local image. The edge information of crops is closely related to the growth conditions of pests and diseases, etc. Therefore, by semantically enhancing the high-frequency features, the formed crop high-frequency enhanced features can have better representation capabilities.

[0112] Step S143: fusing each of the crop high-frequency enhancement features to form crop local image features.

[0113] In this embodiment of the present invention, after forming each of the crop high-frequency enhancement features, each of the crop high-frequency enhancement features can be fused to form a crop local image feature. That is, for crop local image 1, each of the crop high-frequency enhancement features corresponding to crop local image 1 can be fused to form a crop local image feature corresponding to crop local image 1. For crop local image 2, each of the crop high-frequency enhancement features corresponding to crop local image 2 can be fused to form a crop local image feature corresponding to crop local image 2.

[0114] It can be selected that, in the above step S142, the specific process of semantically enhancing the high-frequency features in the process of semantic encoding is not limited. For example, in an achievable embodiment, in order to reliably capture and mine high-frequency information, the above step S142 can further include step S142a, step S142b and step S142c. The specific content of each step is as follows, and can be combined with Figure 6 The content shown.

[0115] Step S142a: masking low-frequency information of the crop local spectrum to form a crop high-frequency spectrum.

[0116] In an embodiment of the present invention, to improve the reliability of mining high-frequency semantic information, the local crop spectrogram can be masked with low-frequency information to form a high-frequency crop spectrogram. For example, this can be achieved using a masking matrix, in which the parameters of the regions corresponding to low-frequency information are equal to 0, and the parameters of the regions corresponding to high-frequency information are equal to 1. This masking matrix can then be bitwise multiplied with the local crop spectrogram to mask the low-frequency information, thereby forming a corresponding high-frequency crop spectrogram that retains only the high-frequency information.

[0117] Step S142b: performing depth convolution on the crop local spectrum map and the crop high-frequency spectrum map respectively to form crop local depth features corresponding to the crop local spectrum map and crop high-frequency depth features corresponding to the crop high-frequency spectrum map.

[0118] In an embodiment of the present invention, after forming the local crop spectrum map and the high-frequency crop spectrum map, the local crop spectrum map and the high-frequency crop spectrum map can be subjected to deep convolution respectively (which can be achieved through a convolutional neural network model in the training process, for details, please refer to the relevant description above) to form local crop depth features corresponding to the local crop spectrum map and high-frequency crop depth features corresponding to the high-frequency crop spectrum map.

[0119] Step S142c: Based on the high-frequency depth features of the crop, semantic enhancement of the high-frequency features of the local depth features of the crop is performed to form a high-frequency enhanced feature of the crop corresponding to the local spectrum of the crop.

[0120] In an embodiment of the present invention, after forming the local depth feature of the crop and the high-frequency depth feature of the crop, the semantic enhancement of the high-frequency feature of the local depth feature of the crop can be performed based on the high-frequency depth feature of the crop to form a high-frequency enhanced feature of the crop corresponding to the local spectrum diagram of the crop, that is, based on the high-frequency depth feature of the crop which only has high-frequency semantic information, the semantic enhancement of the high-frequency feature of the local depth feature of the crop is performed, so that the formed high-frequency enhanced feature of the crop can not only focus on the high-frequency semantic information, but also represent the low-frequency information to a certain extent, that is, avoid the problem of losing some low-frequency effective semantic information due to only using high-frequency semantic information.

[0121] It can be selected that, in the above-mentioned step S142c, the specific process of semantically enhancing the high-frequency features of the local depth features of the crop is not limited. For example, in a feasible embodiment, considering that a variety of sizes of Fourier transforms will be performed on a local crop image, thereby forming multiple local crop spectrograms, and the semantic information represented by the multiple local crop spectrograms is basically the same, only the resolution is different, so that the richness of the semantic representation is different. Based on this, in order to further improve the full mining of high-frequency semantic information, the above-mentioned step S142c can further include step c1, step c2, step c3 and step c4. The specific content of each step is as follows, which can be specifically combined with Figure 7 The content shown.

[0122] Step c1 : performing an upsampling operation on the current crop high-frequency depth feature according to the size of the previous crop high-frequency depth feature to form a crop high-frequency upsampled feature.

[0123] In an embodiment of the present invention, an upsampling operation (e.g., interpolation, deconvolution, etc.) can be performed on the current crop high-frequency depth feature based on the size of the previous crop high-frequency depth feature to form an upsampled crop high-frequency depth feature (of the same size as the previous crop high-frequency depth feature). After sorting the multiple crop local spectrograms corresponding to the crop local image from largest to smallest size, the corresponding crop high-frequency depth features are sorted based on the sorting results to establish a corresponding order, i.e., the ordering relationship of the multiple crop local spectrograms is the same. This allows the size of the previous crop high-frequency depth feature to be larger, preserving more high-frequency semantic information.

[0124] Step c2: Based on the previous crop high-frequency depth feature, attention processing is performed on the crop high-frequency upsampled feature to form a crop high-frequency attention feature.

[0125] In an embodiment of the present invention, after forming the high-frequency upsampling feature of the crop, attention processing (cross attention) can be performed on the high-frequency upsampling feature of the crop based on the previous high-frequency depth feature of the crop to form a high-frequency attention feature of the crop, that is, attention processing is performed by using features with more high-frequency semantic information, so that the high-frequency semantic information can be further represented.

[0126] Step c3: based on the crop local depth feature, down-sampling the crop high-frequency attention feature to form a crop high-frequency down-sampling feature.

[0127] In an embodiment of the present invention, after forming the (size of) the crop high-frequency attention feature, a downsampling operation (such as convolution, pooling, etc.) can be performed on the crop high-frequency attention feature based on the crop local depth feature to form a crop high-frequency downsampled feature (the same size as the crop local depth feature).

[0128] Step c4, forming an importance parameter distribution based on the high-frequency downsampling feature map of the crop, and, based on the importance parameter distribution, performing importance weighting processing on the local depth features of the crop, and connecting the local depth features of the crop to the output of the importance weighting processing to form the high-frequency enhanced features of the crop corresponding to the local spectrum map of the crop.

[0129] In an embodiment of the present invention, after forming the crop high-frequency downsampling feature, an importance parameter distribution can be formed based on the crop high-frequency downsampling feature mapping (for example, after linear mapping the crop high-frequency downsampling feature, the output feature of the linear mapping can be further processed by the sigmoid function to form an importance parameter distribution), and, based on the importance parameter distribution, importance weighting processing is performed on the crop local depth feature (such as bitwise multiplication of the importance parameter distribution and the crop local depth feature, so that gated screening of the crop local depth feature can be achieved, so that high-frequency semantic information is fully mined. Since the importance parameter distribution is derived from the cross-attention results of the two-level crop high-frequency depth features, the representation accuracy of the high-frequency semantic information is high, so that the quality of the importance parameter distribution is better, and therefore, it can be used as a basis for gated screening, which can effectively include the reliability of gated screening), and the crop local depth feature is connected to the output of the importance weighting processing (such as adding the crop local depth feature and the output) to form the crop high-frequency enhanced feature corresponding to the crop local spectrum map.

[0130] In addition, it should be noted that for the first crop high-frequency depth feature, self-attention processing can be performed on the crop high-frequency depth feature, and then the result of the self-attention processing is mapped to form an importance parameter distribution, and based on the importance parameter distribution, importance weighting processing is performed on the local depth feature of the crop, and the local depth feature of the crop is connected to the output of the importance weighting processing to form a high-frequency enhanced feature of the crop corresponding to the local spectrum graph of the crop.

[0131] It can be selected that, in the above-mentioned step S143, the specific process of fusing each of the high-frequency enhanced features of the crop is not restricted. For example, in a feasible implementation, in order to ensure the accuracy of fusion, the above-mentioned step S143 can further include step S143a, step S143b, step S143c and step S143d. The specific content of each step is as follows.

[0132] Step S143a: traverse the plurality of crop high-frequency enhancement features in a direction from small to large in size of the plurality of crop local spectrograms to form the currently traversed crop high-frequency enhancement features.

[0133] In an embodiment of the present invention, multiple crop high-frequency enhancement features can be traversed in the direction from small to large in size of the multiple crop local spectrum graphs to form the currently traversed crop high-frequency enhancement features, that is, traversing in the direction from less semantic information to more.

[0134] Step S143b: If the currently traversed crop high-frequency enhancement feature belongs to the first crop high-frequency enhancement feature, the crop high-frequency enhancement feature is used as the first local fusion feature.

[0135] In an embodiment of the present invention, after forming the currently traversed crop high-frequency enhancement feature, if the currently traversed crop high-frequency enhancement feature belongs to the first crop high-frequency enhancement feature, the crop high-frequency enhancement feature is used as the first local fusion feature.

[0136] Step S143c: If the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, the currently traversed crop high-frequency enhancement feature is fused and output based on the previous local fusion feature to form the current local fusion feature.

[0137] In an embodiment of the present invention, after forming the currently traversed high-frequency enhancement feature of the crop, if the currently traversed high-frequency enhancement feature of the crop does not belong to the first high-frequency enhancement feature of the crop (such as the second high-frequency enhancement feature of the crop), the currently traversed high-frequency enhancement feature of the crop is fused and output based on the previous local fusion feature to form the current local fusion feature. That is to say, by fusing features of adjacent sizes in sequence, it is possible to avoid the problem of directly fusing features with large differences, which may lead to a decrease in fusion accuracy.

[0138] Step S143d: determining the crop local image feature based on the last formed local fusion feature.

[0139] In the embodiment of the present invention, after the last local fusion feature is formed, the crop local image feature may be determined based on the last local fusion feature formed. For example, the last local fusion feature may be used as the crop local image feature.

[0140] Optionally, in the above step S143c, the specific method of forming the current local fusion feature is not limited. For example, in a feasible implementation, on the basis of ensuring the reliability of fusion, in order to achieve efficient fusion of features, the above step S143c may further include the following contents:

[0141] In the first step, if the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, then based on the size of the currently traversed crop high-frequency enhancement feature, the previous local fusion feature is upsampled to form a local upsampled feature (the same size as the currently traversed crop high-frequency enhancement feature), and the currently traversed crop high-frequency enhancement feature and the local upsampled feature are convolutionally fused (i.e., first spliced, then convolved, and in the case of compressed size, the features are interacted, and the compressed size is the same as the size of the currently traversed crop high-frequency enhancement feature) to form a high-frequency convolution fusion feature; in the second step, an importance parameter distribution can be formed based on the high-frequency convolution fusion feature mapping, and, based on the importance parameter distribution, an importance weighting process is performed on the currently traversed crop high-frequency enhancement feature. The specific processing process is as described above; in the third step, the currently traversed crop high-frequency enhancement feature can be connected to the output of the importance weighting process (such as an addition operation) to form the current local fusion feature.

[0142] Based on the above scheme, compared with the fusion method through attention, the computational overhead can be lowered to a certain extent, and high-efficiency feature fusion can be achieved.

[0143] In the fifth part, it should be explained with respect to step S150 that the specific process of analyzing the spread of pests and diseases between two adjacent sub-regions is not limited and can be selected according to actual conditions.

[0144] For example, in one feasible implementation, the crop-related image features corresponding to two adjacent sub-regions can be spliced, and then convolution, pooling and full connection processing can be performed to form corresponding fusion features. Then, the fusion features can be output processed (such as full connection and activation) to form the corresponding pest and disease transmission situation.

[0145] For example, in another feasible implementation, in order to improve the reliability of the analyzed pest and disease transmission situation, the above-mentioned step S150 may further include step S151, step S152, step S153, step S154 and step S155, and the specific content of each step is as follows.

[0146] Step S151 : For every two adjacent sub-regions, performing a difference operation on the crop-related image features corresponding to the two adjacent sub-regions to obtain crop image difference features.

[0147] In this embodiment of the present invention, for every two adjacent sub-regions, a difference operation is performed on the crop-related image features corresponding to the two adjacent sub-regions to obtain a crop image difference feature. Specifically, for each feature parameter in the crop image difference feature, the feature parameter is equal to the difference between the feature parameters at the same position in the crop-related image features corresponding to the two adjacent sub-regions. This means that a bitwise subtraction operation is performed on the two crop-related image features. This difference operation can thus be used to calculate feature change information between the two sub-regions. For example, if a pest or disease in sub-region A spreads to sub-region B, sub-region A will experience it earlier, while sub-region B will experience it later. Therefore, the crop growth in sub-region A is worse than that in sub-region B.

[0148] Step S152 , performing convolution fusion on the crop-associated image features corresponding to two adjacent sub-regions and the crop image difference features to form crop convolution fusion features corresponding to the two adjacent sub-regions.

[0149] In an embodiment of the present invention, after obtaining the crop image difference features, the crop-associated image features corresponding to two adjacent sub-regions and the crop image difference features can be convolved and fused to form crop convolution fused features corresponding to the two adjacent sub-regions. In other words, the crop-associated image features corresponding to the two adjacent sub-regions and the crop image difference features can be concatenated and then convolved to compress the features, thereby obtaining a crop convolution fused feature that has the same size as the crop-associated image features.

[0150] Step S153: performing full connection processing on the crop convolution fusion feature to form a crop fully connected feature.

[0151] In an embodiment of the present invention, after forming the crop convolution fusion feature, the crop convolution fusion feature may be subjected to full connectivity processing to form a crop fully connected feature. The size of the crop fully connected feature is determined based on the number of pest and disease transmission types. For example, if the size of the crop fully connected feature is 1*3, the number of pest and disease transmission types is 3.

[0152] Step S154: activating the fully connected features of the crop to form a propagation type probability distribution.

[0153] In an embodiment of the present invention, after forming the fully connected crop features, the fully connected crop features can be activated (e.g., via a softmax function) to form a propagation type probability distribution (e.g., also of size 1*3). The probability values ​​in the propagation type probability distribution are used to characterize the likelihood of a corresponding pest or disease propagation type, including propagation from a first subregion to a second subregion, propagation from a second subregion to a first subregion, and no propagation between the first and second subregions (e.g., propagation from other subregions to the first and second subregions).

[0154] Step S155 : determining the pest and disease propagation situation between two adjacent sub-regions based on the pest and disease propagation type corresponding to the probability with the maximum value in the propagation type probability distribution.

[0155] In an embodiment of the present invention, after forming the propagation type probability distribution, the pest and disease propagation situation between two adjacent sub-regions can be determined based on the pest and disease propagation type corresponding to the probability with the maximum value in the propagation type probability distribution, for example, from the first sub-region to the second sub-region. In this way, the pest and disease propagation situation can be effectively determined.

[0156] In the sixth part, it is necessary to explain step S160 that the specific process of obtaining the predicted path of pest and disease transmission in the target area is not limited and can be selected according to actual conditions.

[0157] For example, in one feasible implementation, a pest and disease transmission network can be constructed based on the pest and disease transmission situation between every two adjacent sub-regions in the target region, such as Figure 8 As shown. In this way, in the pest and disease transmission network, we can first determine the sub-region that is the source of the transmission, that is, no pest and disease transmission situation represents the transmission from other sub-regions to this sub-region, such as Figure 8 If there are multiple sub-regions F in the network, it indicates that there are multiple transmission sources. Then, for each sub-region belonging to the transmission source, we can use this sub-region as the starting point and traverse the other sub-regions in the transmission direction of the pest transmission network in turn until there are no other sub-regions that can be transmitted (such as Figure 8 In this way, at least one predicted path for the spread of pests and diseases is formed.

[0158] In addition, during the process of pest and disease control, if the path length between the sub-area that needs to be controlled and the sub-area that belongs to the source of transmission is shorter, the control intensity can be greater; conversely, if the path length between the sub-area that needs to be controlled and the sub-area that belongs to the source of transmission is longer, the control intensity can be smaller.

[0159] In summary, the pest and disease transmission path prediction method and system provided by the present invention first obtains a global crop image and a local crop image; secondly, performs a first semantic encoding on the global crop image to form a global crop image feature; then, performs a second semantic encoding on the local crop image to form a local crop image feature; further, based on the global crop image feature, the local crop image features are respectively subjected to association semantic mining to form a crop association image feature; further, based on the corresponding crop association image feature, the pest and disease transmission situation between two adjacent sub-regions is analyzed; finally, based on the pest and disease transmission situation between each two adjacent sub-regions, the pest and disease transmission prediction path is obtained. Based on the above method, on the one hand, since the local crop images of the sub-regions are subjected to multi-scale mining and fusion in the frequency domain, the detailed information in the local crop images can be fully mined, and these detailed information are closely related to the growth of the crop (pests and diseases directly affect the growth). Therefore, based on the mined detailed information, the pest and disease situation can be reliably predicted. On the other hand, since the features of the sub-regions will also be associated and mined based on the global image of the crops in the target area, the obtained crop-associated image features can not only characterize the detailed information, but also characterize the overall situation of the crops. In this way, the reliability of the analyzed pest and disease transmission situation can be further improved. Therefore, the reliability of the pest and disease transmission prediction path obtained by further analysis can be guaranteed, thereby improving the relatively low accuracy of the pest and disease transmission path prediction in the existing technology, so that relevant personnel or equipment can focus on the pest and disease transmission sub-region according to the pest and disease transmission prediction path.

[0160] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for predicting the transmission path of pests and diseases, characterized in that: include: Acquire a global crop image corresponding to the crops in the target area and a local crop image corresponding to the crops in each sub-area included in the target area; performing a first semantic encoding on the crop global image to form a crop global image feature; performing a second semantic coding on each of the crop local images to form a crop local image feature corresponding to each of the crop local images, wherein the second semantic coding is different from the first semantic coding and includes multi-scale mining and fusion in the frequency domain; Based on the crop global image features, performing association semantic mining on each of the crop local image features to form each crop associated image feature; For every two adjacent sub-regions, based on the crop-related image features corresponding to the two adjacent sub-regions, the disease and insect pest transmission situation between the two adjacent sub-regions is analyzed; Based on the pest and disease transmission conditions between every two adjacent sub-regions in the target region, a predicted pest and disease transmission path in the target region is obtained.

2. The method for predicting the spread of pests and diseases according to claim 1, wherein: The step of performing second semantic coding on each of the crop partial images to form crop partial image features corresponding to each of the crop partial images includes: For each of the crop local images, performing Fourier transforms of multiple sizes on the crop local image to form multiple crop local frequency spectra corresponding to the crop local image, wherein the multiple crop local frequency spectra have different sizes; Semantically encoding each of the crop local spectrum graphs, and semantically enhancing the high-frequency features during the semantic encoding process to form high-frequency enhanced features of each crop; Each of the crop high-frequency enhanced features is fused to form a crop local image feature.

3. The method for predicting the spread of pests and diseases according to claim 2, wherein: The step of semantically encoding each of the crop local spectrograms and semantically enhancing the high-frequency features during the semantic encoding process to form the high-frequency enhanced features of each crop includes: Masking low-frequency information on the crop local spectrum to form a crop high-frequency spectrum; Performing depth convolution on the crop local spectrum map and the crop high-frequency spectrum map respectively to form crop local depth features corresponding to the crop local spectrum map and crop high-frequency depth features corresponding to the crop high-frequency spectrum map; Based on the high-frequency depth features of the crop, semantic enhancement of the high-frequency features is performed on the local depth features of the crop to form a high-frequency enhanced feature of the crop corresponding to the local spectrum graph of the crop.

4. The method for predicting the spread of pests and diseases according to claim 3, wherein: The step of semantically enhancing the high-frequency features of the local depth features of the crop based on the high-frequency depth features of the crop to form the high-frequency enhanced features of the crop corresponding to the local spectrum of the crop includes: performing an upsampling operation on the current crop high-frequency depth feature according to the size of the previous crop high-frequency depth feature to form a crop high-frequency upsampled feature, wherein after sorting the multiple crop local spectrograms corresponding to the crop local image in a direction from large to small in size, the corresponding crop high-frequency depth features are sorted according to the sorting result to form a corresponding order relationship; Based on the previous crop high-frequency depth feature, attention processing is performed on the crop high-frequency upsampled feature to form a crop high-frequency attention feature; Based on the crop local depth feature, downsampling the crop high-frequency attention feature to form a crop high-frequency downsampling feature; An importance parameter distribution is formed based on the high-frequency downsampling feature map of the crop, and based on the importance parameter distribution, importance weighting processing is performed on the local depth features of the crop, and the local depth features of the crop are connected to the output of the importance weighting processing to form the high-frequency enhanced features of the crop corresponding to the local spectrum map of the crop.

5. The method for predicting the spread of pests and diseases according to claim 2, wherein: The step of fusing each of the crop high-frequency enhanced features to form crop local image features includes: Traversing the plurality of crop high-frequency enhancement features in a direction from small to large in size of the plurality of crop local frequency spectra to form a currently traversed crop high-frequency enhancement feature; If the currently traversed crop high-frequency enhancement feature belongs to the first crop high-frequency enhancement feature, then the crop high-frequency enhancement feature is used as the first local fusion feature; If the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, then based on the previous local fusion feature, the currently traversed crop high-frequency enhancement feature is fused and output to form the current local fusion feature; Based on the last formed local fusion feature, the crop local image feature is determined.

6. The method for predicting the spread of pests and diseases according to claim 5, wherein: If the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, the step of fusing and outputting the currently traversed crop high-frequency enhancement feature based on the previous local fusion feature to form the current local fusion feature includes: If the currently traversed crop high-frequency enhancement feature does not belong to the first crop high-frequency enhancement feature, then based on the size of the currently traversed crop high-frequency enhancement feature, the previous local fusion feature is upsampled to form a local upsampled feature, and the currently traversed crop high-frequency enhancement feature and the local upsampled feature are convoluted and fused to form a high-frequency convolution fusion feature; forming an importance parameter distribution based on the high-frequency convolution fusion feature map, and performing importance weighting processing on the currently traversed high-frequency enhanced features of the crop based on the importance parameter distribution; The currently traversed high-frequency enhanced features of the crop are connected to the output of the importance weighting process to form the current local fusion features.

7. The method for predicting the spread of pests and diseases according to any one of claims 1 to 6, wherein: The step of performing association semantic mining on each of the local crop image features based on the crop global image features to form each crop association image feature includes: Performing multiple segmentations on the crop global image features to form multiple segmentation sets, wherein the number of crop segmentation image features included in each segmentation set is equal to the number of the crop local image features; Based on the feature correlation between the crop segmentation image features and the crop local image features, determining a segmentation set that best matches the multiple crop local image features from the multiple segmentation sets, and determining the segmentation set as a target segmentation set; Based on each crop segmentation image feature in the target segmentation set, attention processing is performed on each corresponding crop local image feature to form each crop-related image feature.

8. The method for predicting the spread path of pests and diseases according to any one of claims 1 to 6, wherein: The step of performing a first semantic encoding on the crop global image to form crop global image features includes: Performing Fourier transform on the crop global image to form a crop global spectrum map; Performing depth convolution on the crop global spectrum map to form a crop global depth feature; Self-attention processing is performed on the crop global depth features to form crop global image features.

9. The method for predicting the spread of pests and diseases according to any one of claims 1 to 6, wherein: The step of analyzing the pest and disease transmission situation between each two adjacent sub-regions based on the crop-related image features corresponding to the two adjacent sub-regions includes: For every two adjacent sub-regions, performing a difference operation on the crop-related image features corresponding to the two adjacent sub-regions to obtain the crop image difference features; Performing convolution fusion on the crop-related image features corresponding to two adjacent sub-regions and the crop image difference features to form crop convolution fusion features corresponding to the two adjacent sub-regions; Performing full connection processing on the crop convolution fusion feature to form a crop fully connected feature, wherein the size of the crop fully connected feature is determined based on the number of pest and disease transmission types; activating the fully connected features of the crop to form a propagation type probability distribution, wherein the probability values ​​in the propagation type probability distribution are used to represent the likelihood of a corresponding pest and disease propagation type, the pest and disease propagation type including propagation from the first sub-region to the second sub-region, propagation from the second sub-region to the first sub-region, and no propagation between the first sub-region and the second sub-region; The pest and disease transmission situation between two adjacent sub-areas is determined based on the pest and disease transmission type corresponding to the probability with the maximum value in the transmission type probability distribution.

10. A pest and disease transmission path prediction system, characterized in that: It includes a processor and a memory, the memory is used to store a computer program, and the processor is used to execute the computer program to implement the method for predicting the transmission path of pests and diseases according to any one of claims 1 to 9.

Images