Method for underwriting agricultural insurance, electronic device, storage medium, and program product

Abstract

This application provides an underwriting method, electronic device, storage medium, and program product for agricultural insurance, relating to the field of artificial intelligence. The method includes: acquiring drone aerial image data of a target farmland and planting information corresponding to agricultural insurance; inputting the drone aerial image data into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features; performing image segmentation based on the elongated geographical features to obtain a farmland plot segmentation map, the farmland plot segmentation map including at least one crop category; and generating a planting status verification report for the target farmland based on the farmland plot segmentation map, at least one crop category, and planting information, thereby improving the accuracy and efficiency of underwriting work.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more particularly to an underwriting method, electronic device, storage medium, and program product for agricultural insurance. Background Technology

[0002] With the deep integration of agricultural modernization and the digitalization of the insurance industry, agricultural insurance has become an important means of mitigating agricultural production and operational risks and protecting the income of agricultural practitioners. As the core of agricultural insurance risk control, the accuracy of the verification results directly determines the scientific and rational nature of insurance coverage. Accurate verification of farmland planting conditions is a key prerequisite for agricultural insurance underwriting. Automated verification of planting information using technologies such as drone aerial photography and computer vision has become an important development direction in the field of agricultural insurance underwriting.

[0003] In the underwriting process of related agricultural insurance, the verification of farmland planting conditions often adopts a combination of manual on-site inspection and satellite remote sensing image analysis. Some technical solutions also incorporate drone aerial imagery for auxiliary judgment. The specific execution process is as follows: first, aerial or remote sensing image data of farmland is acquired; then, the images are processed through manual annotation or simple image segmentation algorithms to identify the farmland plots and crop types; subsequently, the identification results are manually compared with the planting information provided by the policyholder to finally complete the verification of planting conditions and form an underwriting reference basis. Some algorithms only use single-scale convolutional networks for image feature extraction, which can only identify conventional areal and block-shaped geographical features in the images and are difficult to accurately capture long strip-shaped geographical features such as farmland furrows and strip planting belts.

[0004] In the agricultural insurance underwriting process, the relevant technologies have insufficient accuracy in identifying farmland planting features, especially in capturing long and narrow geographical features, which leads to a large discrepancy between the results of plot segmentation and crop category identification. Summary of the Invention

[0005] This application provides an underwriting method, electronic device, storage medium, and program product for agricultural insurance, in order to solve the problem of large discrepancies between land parcel segmentation and crop category identification results in related technologies.

[0006] Firstly, this application provides an underwriting method for agricultural insurance, comprising:

[0007] Obtain drone aerial imagery data of the target farmland and planting information corresponding to agricultural insurance;

[0008] Drone aerial imagery data is input into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features;

[0009] Image segmentation is performed based on elongated geographical features to obtain farmland plot segmentation maps, which include at least one crop category.

[0010] Based on the farmland plot division map, at least one crop category, and planting information, generate a planting status verification report for the target farmland.

[0011] In one possible implementation, drone aerial imagery data is input into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features, including:

[0012] Preliminary feature extraction is performed on the drone aerial image data to obtain the first feature map;

[0013] The first feature map is input into the offset network in the multi-scale multi-template dynamic strip convolutional network to obtain the rotation factor, which is used to indicate the direction of local features.

[0014] Based on the rotation factor and the length of the predefined bar convolution kernel, the sampling position coordinate offset is generated;

[0015] Based on the offset of the sampling position coordinates, the first feature map is resampled through interpolation to generate a rotated feature map;

[0016] Based on the bar convolution kernel, perform convolution calculation on the rotated feature map, extract the second feature that is consistent with the direction indicated by the rotation factor, and determine the second feature as a long bar geographic feature.

[0017] In one possible implementation, the first feature map is input into an offset network in a multi-scale, multi-template dynamic strip convolutional network to obtain a rotation factor, including:

[0018] The first feature map is input into an offset network, which consists of at least one convolutional layer, at least one batch normalization layer and at least one activation function layer connected in sequence.

[0019] The first feature map is processed by an offset network to output a rotation factor map, wherein the spatial size of the rotation factor map is the same as that of the first feature map and the number of channels is 1.

[0020] Determine the rotation factor based on the rotation factor diagram.

[0021] In one possible implementation, the first feature map is resampled using an interpolation operation based on the sampling position coordinate offset to generate a rotated feature map, including:

[0022] Based on the rotation factor and the length of the bar convolution kernel, the sampling position of each weight point on the bar convolution kernel on the first feature map is calculated using the coordinate transformation formula, and a convolution kernel position index map is generated.

[0023] Based on the convolution kernel position index map, feature values ​​are sampled from the first feature map using bilinear interpolation, and the sampled feature values ​​are arranged in order to generate a rotated feature map.

[0024] In one possible implementation, image segmentation is performed based on elongated geographical features to obtain a farmland plot segmentation map, including:

[0025] Based on the elongated geographical features, a multi-scale, multi-template dynamic strip convolutional network is used for feature enhancement and fusion to obtain a fused feature map, which includes multi-dimensional contextual information.

[0026] The fused feature map is decoded and classified at the pixel level to generate an initial semantic segmentation map;

[0027] The initial semantic segmentation map is post-processed to obtain a farmland plot segmentation map, which includes at least one crop category and plot boundaries.

[0028] In one possible implementation, a verification report on the planting status of the target farmland is generated based on a farmland plot division map, at least one crop category, and planting information, including:

[0029] Based on the farmland plot division map and at least one crop category, calculate the actual planted area for each crop category;

[0030] The actual planted area of ​​at least one crop category and each crop category is compared with the planted area and crop category in the planting information to obtain the comparison results;

[0031] Based on the comparison results, a verification report is generated. The verification report includes at least: plot identification, comparison results of actual and declared crop categories and areas, and anomaly markers.

[0032] Secondly, this application provides an underwriting device for agricultural insurance, comprising:

[0033] The acquisition module is used to acquire drone aerial imagery data of the target farmland and planting information corresponding to agricultural insurance.

[0034] The determination module is used to input UAV aerial imagery data into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features;

[0035] The segmentation module is used to segment the image based on the elongated geographical features to obtain a farmland plot segmentation map, which includes at least one crop category.

[0036] The generation module is used to generate a planting status verification report for the target farmland based on the farmland plot division map, at least one crop category, and planting information.

[0037] In one possible implementation, the determining module is specifically used for:

[0038] Preliminary feature extraction is performed on the drone aerial image data to obtain the first feature map;

[0039] The first feature map is input into the offset network in the multi-scale multi-template dynamic strip convolutional network to obtain the rotation factor, which is used to indicate the direction of local features.

[0040] Based on the rotation factor and the length of the predefined bar convolution kernel, the sampling position coordinate offset is generated;

[0041] Based on the offset of the sampling position coordinates, the first feature map is resampled through interpolation to generate a rotated feature map;

[0042] Based on the bar convolution kernel, perform convolution calculation on the rotated feature map, extract the second feature that is consistent with the direction indicated by the rotation factor, and determine the second feature as a long bar geographic feature.

[0043] In one possible implementation, the determining module is specifically used for:

[0044] The first feature map is input into an offset network, which consists of at least one convolutional layer, at least one batch normalization layer and at least one activation function layer connected in sequence.

[0045] The first feature map is processed by an offset network to output a rotation factor map, wherein the spatial size of the rotation factor map is the same as that of the first feature map and the number of channels is 1.

[0046] Determine the rotation factor based on the rotation factor diagram.

[0047] In one possible implementation, the determining module is specifically used for:

[0048] Based on the rotation factor and the length of the bar convolution kernel, the sampling position of each weight point on the bar convolution kernel on the first feature map is calculated using the coordinate transformation formula, and a convolution kernel position index map is generated.

[0049] Based on the convolution kernel position index map, feature values ​​are sampled from the first feature map using bilinear interpolation, and the sampled feature values ​​are arranged in order to generate a rotated feature map.

[0050] In one possible implementation, the segmentation module is specifically used for:

[0051] Based on the elongated geographical features, a multi-scale, multi-template dynamic strip convolutional network is used for feature enhancement and fusion to obtain a fused feature map, which includes multi-dimensional contextual information.

[0052] The fused feature map is decoded and classified at the pixel level to generate an initial semantic segmentation map;

[0053] The initial semantic segmentation map is post-processed to obtain a farmland plot segmentation map, which includes at least one crop category and plot boundaries.

[0054] In one possible implementation, the generation module is specifically used for:

[0055] Based on the farmland plot division map and at least one crop category, calculate the actual planted area for each crop category;

[0056] The actual planted area of ​​at least one crop category and each crop category is compared with the planted area and crop category in the planting information to obtain the comparison results;

[0057] Based on the comparison results, a verification report is generated. The verification report includes at least: plot identification, comparison results of actual and declared crop categories and areas, and anomaly markers.

[0058] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0059] The memory stores instructions that the computer executes;

[0060] The processor executes computer-executable instructions stored in memory to implement any of the methods of the first aspect.

[0061] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0062] Fifthly, this application provides a computer program product, including a computer program that, when executed by a computer, implements the method as described in any of the first aspects.

[0063] This application provides an agricultural insurance underwriting method, electronic device, storage medium, and program product. It acquires drone aerial imagery data of the target farmland and corresponding planting information for agricultural insurance. The drone aerial imagery data is input into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features. Image segmentation is performed based on these elongated geographical features to obtain farmland plot segmentation maps, each including at least one crop category. Based on the farmland plot segmentation maps, at least one crop category, and planting information, a planting status verification report for the target farmland is generated. In this way, by combining drone aerial imagery with a multi-scale, multi-template dynamic strip convolutional network to accurately extract elongated geographical features of farmland and complete plot segmentation and crop identification, and then combining this with planting information to generate a verification report, the accuracy and efficiency of farmland planting status verification in agricultural insurance underwriting are improved, and underwriting data verification is standardized. Furthermore, it provides underwriters with detailed decision-making basis, enhancing the accuracy and efficiency of underwriting work. Attached Figure Description

[0064] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0065] Figure 1 This application provides a schematic diagram of the architecture of an underwriting system.

[0066] Figure 2 A flowchart illustrating an underwriting method for agricultural insurance provided in this application. Figure 1 ;

[0067] Figure 3 A flowchart illustrating an underwriting method for agricultural insurance provided in this application. Figure 2 ;

[0068] Figure 4 This application provides a schematic diagram of the structure of a dynamically deformable convolution.

[0069] Figure 5 A flowchart illustrating an underwriting method for agricultural insurance provided in this application. Figure 3 ;

[0070] Figure 6 A schematic diagram illustrating the workflow of a multi-scale, multi-template dynamic strip convolutional network provided in this application;

[0071] Figure 7 A schematic diagram illustrating a multi-scale feature fusion workflow provided in this application;

[0072] Figure 8 A schematic diagram illustrating the workflow of a multi-scale, multi-template dynamic strip convolution provided in this application;

[0073] Figure 9 A schematic diagram of the structure of an underwriting device for agricultural insurance provided in this application;

[0074] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0075] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0076] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0077] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0078] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0079] It should be noted that the agricultural insurance underwriting method, electronic device, storage medium and program product provided in this application can be used in the field of artificial intelligence, or in any field other than artificial intelligence. The application field of the agricultural insurance underwriting method, electronic device, storage medium and program product in this application is not limited.

[0080] With the deep integration of agricultural modernization and the digitalization of the insurance industry, agricultural insurance has become an important means of mitigating agricultural production and operational risks and protecting the income of agricultural practitioners. As the core of agricultural insurance risk control, the accuracy of the verification results directly determines the scientific and rational nature of insurance coverage. Accurate verification of farmland planting conditions is a key prerequisite for agricultural insurance underwriting. Automated verification of planting information using technologies such as drone aerial photography and computer vision has become an important development direction in the field of agricultural insurance underwriting.

[0081] In the underwriting process of related agricultural insurance, the verification of farmland planting conditions often adopts a combination of manual on-site inspection and satellite remote sensing image analysis. Some technical solutions also incorporate drone aerial imagery for auxiliary judgment. The specific execution process is as follows: first, aerial or remote sensing image data of farmland is acquired; then, the images are processed through manual annotation or simple image segmentation algorithms to identify the farmland plots and crop types; subsequently, the identification results are manually compared with the planting information provided by the policyholder to finally complete the verification of planting conditions and form an underwriting reference basis. Some algorithms only use single-scale convolutional networks for image feature extraction, which can only identify conventional areal and block-shaped geographical features in the images and are difficult to accurately capture long strip-shaped geographical features such as farmland furrows and strip planting belts.

[0082] In the agricultural insurance underwriting process, the relevant technologies have insufficient accuracy in identifying farmland planting features, especially in capturing long and narrow geographical features, which leads to a large discrepancy between the results of plot segmentation and crop category identification.

[0083] To address the aforementioned technical issues, this application provides an underwriting method for agricultural insurance. Addressing the need for accuracy and efficiency in verifying farmland planting conditions during agricultural insurance underwriting, this method integrates UAV aerial imagery acquisition technology with computer vision algorithms. By acquiring UAV aerial imagery data of the target farmland and corresponding planting information, a multi-scale, multi-template dynamic strip convolutional network is used to extract features from the aerial images, accurately capturing elongated geographical features in the farmland. Based on these features, image segmentation of farmland plots is performed, and crop categories are identified. Finally, cross-verification is conducted using the insured planting information to generate a planting condition verification report. This achieves automated and precise verification of farmland planting conditions during agricultural insurance underwriting, improving the accuracy and efficiency of the underwriting process.

[0084] Below, in conjunction with Figure 1 The architecture of the underwriting system will be illustrated with an example.

[0085] Figure 1 This is a schematic diagram of the architecture of an underwriting system provided in an embodiment of this application. Please refer to [link / reference]. Figure 1 , Figure 1This may include an underwriting system. The system comprises a data acquisition unit 101, a data analysis unit 102, and a report generation unit 103. Specifically, this system is applied to automated agricultural insurance underwriting scenarios, automatically verifying the planting status of insured farmland by processing drone aerial imagery.

[0086] The data acquisition unit 101 can be used to collect and receive the raw data required for underwriting and perform preliminary processing to provide standardized input for subsequent analysis.

[0087] The functions of the data acquisition unit can include data acquisition and reception, data preprocessing, and data output.

[0088] Data acquisition and reception can involve obtaining drone aerial imagery data of the target farmland. This data can be RGB imagery, multispectral imagery, or multi-source data fused with a digital elevation model (DEM).

[0089] Data preprocessing can be a process of standardizing the acquired raw images, including but not limited to geometric correction, radiometric correction, image stitching, and normalization, to eliminate the effects caused by differences in shooting angle, lighting conditions, or sensors.

[0090] Data output can be the transmission of preprocessed, standardized image data to the data analysis unit 102. Simultaneously, this unit can also receive and associate planting information from the agricultural insurance policy information corresponding to the target farmland, such as the crop type, planting area, and plot location declared by the policyholder, and provide this information to subsequent units in conjunction with or in connection with it.

[0091] The data analysis unit 102 can be used to intelligently extract features from images and make segmentation decisions. This unit typically contains multiple sub-modules, implemented in the form of software algorithms or hardware acceleration modules.

[0092] Multiple sub-modules may include a feature extraction sub-module, a dynamic bar convolution sub-module, and a segmentation decision sub-module.

[0093] The feature extraction submodule can be used to receive image data from the data acquisition unit 101.

[0094] The dynamic strip convolution submodule can be used to further process the first feature map. Through offset network, coordinate transformation and resampling, and multi-scale multi-template fusion, it determines the long strip geographical features that integrate multi-scale multi-template information.

[0095] The segmentation decision submodule can be used to input extracted elongated geographic features as high-level data into a semantic segmentation network for upsampling and pixel-level classification. The submodule ultimately outputs a farmland plot segmentation map, where each pixel is assigned a category label.

[0096] The report generation unit 103 can be used to transform data analysis results into decision reports that can be directly used for underwriting business.

[0097] After underwriters initiate a task, data acquisition unit 101 automatically or manually imports drone images and policy information of the target farmland. The processed data flows into data analysis unit 102, where feature extraction and segmentation using a multi-scale, multi-template dynamic strip convolutional network accurately identifies the crop area. Finally, report generation unit 103 compares the segmentation results with the policy information to generate a verification report, thereby achieving automated and high-precision underwriting of farmland planting conditions, significantly reducing the cost and time of manual surveys.

[0098] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0099] Figure 2 A flowchart illustrating an underwriting method for agricultural insurance provided in this application. Figure 1 ,like Figure 2 As shown, the method includes:

[0100] S201. Obtain drone aerial imagery data of the target farmland and planting information corresponding to agricultural insurance.

[0101] The target farmland can refer to the specific farmland area corresponding to the agricultural insurance purchased by the policyholder, and its boundaries can be clearly defined.

[0102] Drone aerial imagery data can be visual data obtained by drones taking aerial photos of target farmland from all directions and multiple angles.

[0103] Drone aerial imagery data can include key information such as the terrain of the target farmland and the crop growth status, and the aerial resolution and coverage can be flexibly adjusted according to underwriting requirements.

[0104] Planting information can come from agricultural insurance application materials submitted by the policyholder, and can be any kind of information related to farmland planting.

[0105] Planting information may include, but is not limited to, the types of crops planted, the planting area, the planting cycle, and the location of the insured plots.

[0106] Planting information can be used for cross-verification with the actual planting situation in farmland.

[0107] It can receive real-time aerial image data of target farmland transmitted by drones via wireless communication links, and at the same time connect to the agricultural insurance system to automatically retrieve the electronic planting information submitted by the insured. It can simultaneously acquire two types of data, and after acquisition, it uses Gaussian filtering algorithm for noise reduction and missing value interpolation method to supplement incomplete data.

[0108] Historical and real-time aerial image data of the target farmland can be obtained through third-party drone aerial photography service platforms, and planting information of the target farmland can be retrieved through the farmland planting database of the agricultural and rural departments.

[0109] S202. Input the drone aerial imagery data into a multi-scale, multi-template dynamic strip convolutional network to determine the elongated geographical features.

[0110] Multi-scale, multi-template dynamic strip convolutional networks can be a type of convolutional network that possesses multi-scale feature capture capabilities and dynamic adaptability.

[0111] Multi-scale, multi-template dynamic strip convolutional networks can be used to process drone aerial image data, accurately extracting long strip geographical features of different sizes and shapes from the images, overcoming the shortcomings of traditional convolutional networks in capturing subtle long strip features.

[0112] Elongated geographical features can be geographical or planting-related features that appear as elongated strips in aerial images of the target farmland.

[0113] Long, narrow geographical features, including but not limited to farmland furrows, strip-shaped planting belts, and field ridges, are key features for distinguishing farmland plots and identifying planting layouts.

[0114] The acquired drone aerial image data can be preprocessed, and then the preprocessed image data can be input into a preset multi-scale multi-template dynamic strip convolutional network. Through the multi-scale convolutional kernel and dynamic template matching of the network, various strip-shaped geographical features in the image can be identified and extracted, and the strip-shaped geographical features can be output.

[0115] Optionally, the preprocessed UAV aerial imagery data can be directly input into the trained multi-scale, multi-template dynamic strip convolutional network. The network extracts strip-shaped geographic features of different sizes through three strip convolutional kernels of different scales, and then fuses the three types of features through a dynamic fusion module to output strip-shaped geographic features.

[0116] Optionally, the drone aerial image data can be segmented into multiple sub-regions. Each sub-region is then input into a multi-scale, multi-template dynamic strip convolutional network. The template parameters of the network are dynamically adjusted according to the image characteristics of different sub-regions. The elongated geographical features of each sub-region are extracted, and finally, the features of each sub-region are stitched together to obtain the elongated geographical features of the target farmland.

[0117] S203. Perform image segmentation based on the elongated geographical features to obtain a farmland plot segmentation map.

[0118] The farmland plot subdivision map includes at least one crop category.

[0119] Image segmentation refers to dividing an image into multiple non-overlapping regions based on the differences in features within the image, with each region corresponding to a specific target or category.

[0120] Farmland plot segmentation maps are images obtained by segmenting drone aerial images based on elongated geographical features. They can clearly show the plot boundaries of the target farmland and mark at least one crop category corresponding to each plot, intuitively reflecting the actual planting layout of the farmland.

[0121] Crop category can refer to the various types of crops grown in the target farmland, such as wheat, corn, rice, cotton, etc.

[0122] The aerial imagery can be segmented using elongated geographical features as the boundary, and an image segmentation algorithm can be used to divide the aerial imagery into multiple independent farmland plots. At the same time, by combining the image texture, color and other features of each plot after segmentation, a crop category recognition model can be used to identify the crop category corresponding to each plot. Finally, the plot boundaries and crop categories are marked to generate a farmland plot segmentation map.

[0123] Optionally, an image segmentation algorithm based on semantic segmentation can be used to segment drone aerial images using elongated geographical features as semantic labels to divide them into independent farmland plots. At the same time, a CNN crop recognition model is used to identify the image of each plot, determine the crop category of each plot, and mark the plot boundaries and crop categories on the image to generate a farmland plot segmentation map.

[0124] Optionally, a threshold segmentation combined with edge detection algorithm can be used. First, the grayscale threshold of the elongated geographical features is used as a basis to initially segment the boundaries of farmland plots. Then, the edge detection algorithm is used to optimize the boundary accuracy. Crop category identification adopts the template matching method, which matches the images of each plot with the preset standard image templates of various crops to determine the crop category and finally generate a farmland plot segmentation map.

[0125] S204. Based on the farmland plot division map, at least one crop category, and planting information, generate a planting status verification report for the target farmland.

[0126] The planting status verification report can include details of farmland plot division, crop category identification results, planting information comparison results, verification conclusions, etc., and can be directly used as a reference for agricultural insurance underwriting decisions.

[0127] Information such as the number of plots, area of ​​each plot, and crop type can be extracted from the farmland plot segmentation map. Then, it can be compared item by item with the planting information submitted by the policyholder to analyze the consistency between the two, mark the differences, and finally combine the comparison results to generate a planting status verification report that includes verification details, difference explanations, and verification conclusions.

[0128] Optionally, the system can statistically analyze the total planting area, planting area of ​​various crops, and distribution area on the farmland plot segmentation map, and compare them item by item with the total planting area, crop type, and planting area of ​​various crops in the planting information submitted by the policyholder. If the deviation of each data is within the preset threshold, the verification is deemed to have passed; if the deviation exceeds the threshold, the deviation points are marked and the reasons are analyzed. Finally, a planting status verification report containing comparison details, deviation explanations, and verification conclusions is generated, which can be exported in PDF format.

[0129] Optionally, a tiered verification model can be adopted. First, the crop types and total planting area are compared. Then, the area of ​​each plot, the crop distribution, and the location of the insured plot are compared. Based on the verification results, a tiered verification conclusion is given, and a planting verification report containing tiered verification details, verification basis, and underwriting recommendations is generated and simultaneously pushed to the underwriting personnel's terminal for review.

[0130] Optionally, a verification report on the planting status of the target farmland can be generated based on the farmland plot division map, at least one crop category, and planting information in the following manner: Calculate the actual planting area of ​​each crop category based on the farmland plot division map and at least one crop category; compare the actual planting area of ​​at least one crop category and each crop category with the planting area and crop category in the planting information to obtain the comparison results; and generate a verification report based on the comparison results.

[0131] The verification report should include at least: plot identification, comparison results of actual and declared crop categories and areas, and anomaly markers.

[0132] This embodiment provides an agricultural insurance underwriting method that acquires drone aerial imagery data of the target farmland and corresponding planting information for agricultural insurance. The drone aerial imagery data is input into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features. Image segmentation is performed based on these elongated geographical features to obtain farmland plot segmentation maps, each including at least one crop category. Based on the farmland plot segmentation maps, at least one crop category, and planting information, a planting status verification report for the target farmland is generated. This method, by combining drone aerial imagery with a multi-scale, multi-template dynamic strip convolutional network to accurately extract elongated geographical features of farmland and complete plot segmentation and crop identification, and then combining this with planting information to generate a verification report, not only improves the accuracy and efficiency of farmland planting status verification in agricultural insurance underwriting and standardizes underwriting data verification, but also provides underwriters with detailed decision-making basis, enhancing the accuracy and efficiency of underwriting work.

[0133] Below, in conjunction with Figure 3 This paper explains the process of inputting UAV aerial imagery data into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features.

[0134] Figure 3 A flowchart illustrating an underwriting method for agricultural insurance provided in this application. Figure 2 ,like Figure 3 As shown, in this embodiment... Figure 2 Based on the examples, a method for underwriting agricultural insurance is described in detail, which includes:

[0135] S301. Perform preliminary feature extraction on the drone aerial image data to obtain the first feature map.

[0136] Preliminary feature extraction can refer to performing basic operations such as convolution and normalization on the original UAV aerial image data to filter out basic shallow features such as texture, edge, and grayscale in the image.

[0137] The first feature map can be the feature matrix output after preliminary feature extraction of drone aerial image data.

[0138] The first feature map can contain information such as the basic edges and textures of the farmland image.

[0139] The drone aerial image data can be preprocessed, and the preprocessed image data can be input into a preset basic convolution. The shallow features of the image can be extracted through the basic convolution operation, and the first feature map can be output.

[0140] S302. Input the first feature map into the offset network in the multi-scale multi-template dynamic strip convolutional network to obtain the rotation factor.

[0141] The rotation factor is used to indicate the orientation of local features.

[0142] Multi-scale, multi-template dynamic strip convolutional networks can be optimized convolutional networks for extracting long strip geographical features of farmland.

[0143] The multi-scale, multi-template dynamic strip convolutional network can include an offset network, a strip convolution kernel generation module, a feature resampling module, and a convolution calculation module, which can dynamically adapt to the extraction needs of long strip geographic features of different directions and sizes.

[0144] An offset network can be a subnetwork of a multi-scale, multi-template dynamic strip convolutional network.

[0145] Offset networks can consist of convolutional layers, batch normalization (BN) layers, and sigmoid layers.

[0146] The offset network can be used to predict the rotation factor corresponding to each position on the first feature map, thereby achieving dynamic orientation adaptation of the strip convolution kernel.

[0147] The rotation factor can be an element value from a feature map output by the offset network, which has the same size as the first feature map and has 1 channel.

[0148] The range of values ​​for the rotation factor is limited to [0, π].

[0149] The rotation factor can be used to represent the angle by which the strip convolution kernel needs to rotate counterclockwise when passing through a location, thus indicating the local orientation of the elongated geographical feature of farmland at that location.

[0150] The first feature map can be input into an offset network, which can consist of two 3×3 convolutional layers, one BN layer, and one Sigmoid layer. The first convolutional layer compresses the number of channels in the first feature map to 16, the second convolutional layer further reduces the number of channels to 1, the BN layer normalizes the feature values, and finally the Sigmoid layer maps the output values ​​to the [0,π] interval, generating a rotation factor map with the same size as the first feature map. Each element in the map is the rotation factor at the corresponding position.

[0151] Optionally, the first feature map can be input into an offset network in a multi-scale, multi-template dynamic strip convolutional network to obtain a rotation factor: input the first feature map into an offset network; process the first feature map through the offset network to output a rotation factor map; determine the rotation factor based on the rotation factor map.

[0152] The offset network consists of at least one convolutional layer, at least one batch normalization layer, and at least one activation function layer connected in sequence.

[0153] Convolutional layers can be used to compress and abstract features from the first feature map, ultimately compressing the multi-channel input feature map into a single-channel feature.

[0154] Batch normalization layers can be used to normalize the feature values ​​of the convolution output, stabilize the training process, and avoid gradient vanishing or exploding.

[0155] The activation function layer can be used with the Sigmoid function to map the output value to the interval [0,π], ensuring that the rotation factor is a valid angle value.

[0156] The spatial dimensions of the rotation factor map are the same as those of the first feature map, and the number of channels is 1.

[0157] The rotation factor map provides pixel-by-pixel directional instructions for the subsequent rotation of the bar convolution kernel, enabling the network to dynamically adapt to long bar features in different directions, thereby accurately extracting key geographical information such as furrows and ridges in farmland.

[0158] S303. Generate sampling position coordinate offset based on the rotation factor and the length of the predefined bar convolution kernel.

[0159] The length of the bar convolution kernel can be a predefined one-dimensional length of the bar convolution kernel, and its value is the area of ​​the square convolution kernel. It is the core parameter for calculating the coordinate offset of the sampling position.

[0160] The sampling position coordinate offset can be calculated using a geometric coordinate transformation formula based on the rotation factor and the length of the bar convolution kernel, representing the offset between the target sampling position and the original position of each weight on the first feature map.

[0161] The length L of the predefined bar convolution kernel can be determined. For each position (i,j) on the rotation factor map, corresponding to the rotation angle α, the target coordinates (x,y) of the k-th weight (1≤k≤L) on the bar convolution kernel after rotation can be calculated according to the formula, with that position as the center:

[0162] x = a + (k-(L+1) / 2)×sinα

[0163] y = b + (k-(L+1) / 2)×cosα

[0164] Where a and b are the center coordinates of the convolution kernel sliding to that position, and the calculated x and y are subject to boundary constraints: 0 when x < 0, H when x > H; 0 when y < 0, W when y > W, where H and W are the height and width of the first feature map.

[0165] The difference between the target coordinates and the original coordinates can be calculated to obtain the sampling position coordinate offset, and finally a coordinate offset tensor with dimensions of 2×L×H×W is generated.

[0166] Here, 2 can correspond to the x and y dimensions, L can be the convolution kernel length, and H and W can be the feature map sizes.

[0167] S304. Based on the offset of the sampling position coordinates, the first feature map is resampled through interpolation to generate a rotated feature map.

[0168] Interpolation can be a method that calculates the feature value at a non-integer coordinate by weighting the feature values ​​of adjacent integer coordinate points for non-integer coordinates obtained from the offset of the sampling position.

[0169] Resampling is a process that reselects feature values ​​on the first feature map based on the offset of the sampling position, and constructs a new feature distribution. The core is to convert the geometric operation of rotating the bar convolution kernel into the adjustment of the sampling position of the feature map.

[0170] The rotated feature map is a feature matrix generated after resampling the first feature map. Its dimensions are (H×L)×W×C, where H and W can be the size of the first feature map, L can be the length of the bar convolution kernel, and C can be the number of channels. It is equivalent to the feature distribution obtained by rotating the bar convolution kernel and sampling it on the original feature map.

[0171] The sampling position coordinate offset tensor can be broadcast along the channel dimension to match the number of channels C of the first feature map. Then, for each channel of the first feature map, the corresponding non-integer sampling coordinate is found according to the coordinate offset value, and the feature value at that coordinate is calculated using bilinear interpolation. Finally, the L interpolation results at each position are concatenated along the height dimension to convert the original H×W×C first feature map into a (H×L)×W×C rotated feature map, thus completing the reconstruction of the feature distribution and enabling the rotated feature map to match the sliding requirements of the strip convolution kernel.

[0172] Optionally, the first feature map can be resampled using interpolation based on the sampling position coordinate offset to generate a rotated feature map: based on the rotation factor and the length of the bar convolution kernel, the sampling position of each weight on the bar convolution kernel on the first feature map is calculated using the coordinate transformation formula to generate a convolution kernel position index map; based on the convolution kernel position index map, feature values ​​are sampled from the first feature map using bilinear interpolation, and the multiple sampled feature values ​​are arranged in order to generate a rotated feature map.

[0173] Traditional convolution uses a fixed grid for sampling, but this application uses a rotatable line segment for sampling. This line segment has multiple sampling points and can be rotated at any angle to capture features from different directions.

[0174] S305. Based on the bar convolution kernel, perform convolution calculation on the rotated feature map, extract the second feature that is consistent with the direction indicated by the rotation factor, and determine the second feature as a long bar geographical feature.

[0175] The strip convolution kernel can abandon the traditional square convolution kernel design. It is a one-dimensional strip convolution kernel based on the nn.Conv2d function of the PyTorch framework. The kernel_size attribute is set to (L,1) (L is the preset length), which is specially adapted for the extraction of long strip geographical features.

[0176] The second feature is the feature output after performing bar convolution on the rotated feature map, which accurately matches the direction of the long strip-shaped geographical features indicated by the rotation factor, including core feature information such as farmland furrows, field ridges, and strip-shaped planting belts.

[0177] The elongated geographical features can ultimately be extracted as a set of features that characterize the elongated geographical elements of farmland (ridges, furrows, field ridges, strip planting belts, etc.).

[0178] A strip convolution kernel can be initialized, and the convolution stride can be set to be the same as the length L of the strip convolution kernel. Then, the rotated feature map is input into the strip convolution layer to perform two-dimensional convolution calculation. During the convolution process, only the long strip features that match the direction indicated by the rotation factor are extracted. Finally, the feature map output by the convolution is restored to its dimensions to obtain the second feature, and the second feature is determined as the long strip geographical feature.

[0179] Below, in conjunction with Figure 4 This paper illustrates the process of inputting drone aerial imagery data into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features.

[0180] Figure 4 This application provides a schematic diagram of a dynamically deformable convolution structure, such as... Figure 4 As shown, Figure 4 Includes the following steps:

[0181] Step 1: For the original drone aerial images with a resolution of 1024×1024, including roads, farmland, low-rise buildings, etc., use a pre-trained lightweight CNN to perform preliminary feature extraction, and obtain the first feature map with 64 channels and a size of 256×256.

[0182] Step 2: Input the first feature map into the offset network of the multi-scale multi-template dynamic strip convolutional network. This network consists of 3 layers of 3×3 convolutional layers and finally outputs a rotation factor map with a size of 1×256×256.

[0183] The α_xy value at each spatial location (ranging from [-π / 2, π / 2]) precisely indicates the principal direction of the local features at that location.

[0184] Step 3: Based on the predefined strip convolution kernel length L=9, and combined with the rotation factor α_xy at each position, calculate the coordinate offset of the dynamic sampling position.

[0185] For the initial horizontally arranged set of sampling points {(x_0+i,y_0) | i=0,…,L-1}, the coordinates (x',y') of the new sampling points are obtained through rotation transformation. The difference between the new sampling points and the initial points is the coordinate offset (Δx, Δy).

[0186] Step 4: Using bilinear interpolation, the first feature map is resampled based on the sampling position coordinate offset. The pixel values ​​of the original feature map are mapped to the rotated sampling positions to generate a rotated feature map aligned with the local feature direction.

[0187] Its size is consistent with the first feature map (64×256×256), ensuring that the features of the road and other long strip structures are fully aligned in the direction of the convolution kernel.

[0188] Step 5: Perform convolution calculation on the rotated feature map using a predefined bar convolution kernel (size 1×9, number of channels 64). This convolution operation extracts features in the dimension that is completely consistent with the direction indicated by the rotation factor, and finally obtains an output feature map with 32 channels and a size of 256×256.

[0189] By performing threshold segmentation and connected component analysis on the output feature map, the elongated road features in the UAV imagery can be obtained.

[0190] The implementation details of each step in this application embodiment can be found in the description of the corresponding steps or operations in the above method embodiments; repeated content will not be repeated.

[0191] This embodiment provides an underwriting method for agricultural insurance. First, a first feature map is obtained by performing preliminary feature extraction on UAV aerial imagery data. This first feature map is then input into an offset network within a multi-scale, multi-template dynamic strip convolutional network to obtain a rotation factor, which indicates the direction of local features. Based on the rotation factor and the length of a predefined strip convolutional kernel, a sampling position coordinate offset is generated. Based on the sampling position coordinate offset, the first feature map is resampled using interpolation to generate a rotated feature map. Finally, a convolution calculation is performed on the rotated feature map using the strip convolutional kernel to extract a second feature consistent with the direction indicated by the rotation factor, and this second feature is identified as a long strip-shaped geographical feature. In this way, by dynamically rotating the strip convolutional kernel and aligning it with the direction of local features, long strip-shaped geographical features such as roads and rivers in UAV imagery can be extracted more accurately and efficiently, improving recognition accuracy and directional adaptability.

[0192] Below, in conjunction with Figure 5 This paper explains the process of segmenting images based on elongated geographical features to obtain farmland plot segmentation maps.

[0193] Figure 5 A flowchart illustrating an underwriting method for agricultural insurance provided in this application. Figure 3 ,like Figure 5 As shown, in this embodiment... Figure 2 Based on the examples, a method for underwriting agricultural insurance is described in detail, which includes:

[0194] S501. Based on the elongated geographical features, a multi-scale, multi-template dynamic strip convolutional network is used to enhance and fuse features, resulting in a fused feature map.

[0195] The fused feature map includes multi-dimensional contextual information.

[0196] The offset network can be modified to expand its output channels to four, enabling it to generate four sets of rotation factor templates for the input elongated geographic feature map. These templates are then converted into four sets of convolutional kernel position index maps by the position index generation unit. The four sets of index maps are assigned to each channel of the input feature map in a cyclic or maximum response priority manner, so that the predefined horizontal bar convolutional kernels can dynamically rotate and perform convolution on each channel according to the corresponding index map, generating a rotated feature map with orientation information. This is then enhanced with two-dimensional convolution to improve details. Rotational feature maps of different scales are extracted by multi-size bar convolutional kernels, and all features are then fused by methods such as concatenation, element-wise addition, or cross-scale attention weighting to obtain a fused feature map containing multi-scale and multi-directional information.

[0197] Below, in conjunction with Figure 6 The workflow of multi-scale, multi-template dynamic strip convolutional networks is explained.

[0198] Figure 6 A schematic diagram illustrating the workflow of a multi-scale, multi-template dynamic strip convolutional network provided in this application is shown below. Figure 6 As shown, Figure 6 Includes the following steps:

[0199] Step 1: The input feature map can be fed into the offset network, which outputs four sets of rotation factor templates (α, β, γ, δ). These templates are the basis for the subsequent rotation of the convolutional kernels, and each set of templates corresponds to a rotation parameter of a certain angle.

[0200] Step 2: The four sets of rotation factor templates are converted into convolution kernel position index maps and assigned to each channel of the input feature map in a cyclic manner. For example, channel 1 matches the α template, channel 2 matches the β template, channel 3 matches the γ template, channel 4 matches the δ template, channel 5 returns to the α template, and so on, ensuring that each channel has an independent rotation angle.

[0201] Step 3: In each channel, a predefined horizontal strip convolution kernel is rotated according to the assigned rotation factor template, and then convolved with the input features of that channel to generate a rotated feature channel with orientation information. This step, through channel-by-channel rotational convolution, allows the model to capture long strip features in different directions.

[0202] Step 4: After the rotation feature channels of all channels are integrated, they are added element-by-element or spliced ​​together ( Figure 6 The feature fusion is performed using a method that is indicated by "Add or Concat" to achieve deep separability, ultimately resulting in a long strip-shaped geographic feature.

[0203] Below, in conjunction with Figure 7 The workflow of multi-scale feature fusion is explained.

[0204] Figure 7 This application provides a schematic diagram of a multi-scale feature fusion workflow, such as... Figure 7 As shown, Figure 7 Includes the following steps:

[0205] Step 1: Input feature map;

[0206] Step 2: The input feature map is fed in parallel into three dynamic strip convolution branches with identical structures but different kernel sizes:

[0207] Branch 1: Kernel size L = kernel_size2 (base size)

[0208] Branch 2: Kernel size L = kernel_size² + 6 (medium size)

[0209] Branch 3: Kernel size L = kernel_size² + 12 (large scale)

[0210] The dynamic strip convolution module of each branch dynamically rotates the strip convolution kernel on the corresponding channel according to the rotation factor generated by the offset network, thereby capturing long strip features of different scales and directions.

[0211] Step 3: The convolution results of the three branches are fused element-wise to generate an attention map. This attention map highlights significant elongated feature regions at different scales, effectively providing dynamic weights to the original feature map.

[0212] Step 4: The attention map is element-wise multiplied with the original input feature map to enhance the features of the corresponding salient regions in the original feature map and suppress the features of the background regions. Then, a 1×1 convolution (Conv_1x1) is used to adjust the number of channels to match the input requirements of the subsequent network, and finally the output feature map is obtained.

[0213] By using multi-scale parallel convolution and attention weighting, this module can simultaneously capture elongated geographical features of different thicknesses and directions, which not only enhances the expressive power of the features but also suppresses irrelevant backgrounds through the attention mechanism, thereby improving the accuracy of subsequent semantic segmentation tasks.

[0214] Below, in conjunction with Figure 8 The structure of multi-scale, multi-template dynamic strip convolution is explained.

[0215] Figure 8 This application provides a schematic diagram of the workflow for a multi-scale, multi-template dynamic strip convolution, as shown below. Figure 8 As shown, Figure 8 Includes the following steps:

[0216] Step 1: Input feature map;

[0217] Step 2: The input feature map is fed in parallel into three branches with the same structure but different kernel sizes:

[0218] Branch 1: Kernel size L = kernel_size2 (base size)

[0219] Branch 2: Kernel size L = kernel_size² + 6 (medium size)

[0220] Branch 3: Kernel size L = kernel_size² + 12 (large scale)

[0221] Each branch uses a multi-template feature fusion module instead of the original dynamic strip convolution module. Each module generates four sets of rotation factor templates, allowing the strip convolution kernel to rotate dynamically on different channels, thus capturing long strip features in multiple directions at the same scale.

[0222] Step 3: The outputs of the three branches are merged by element-wise addition to generate an attention map;

[0223] Step 4: The attention map is element-wise multiplied with the original input feature map to enhance the salient feature regions and suppress background interference. Then, the number of channels is adjusted by a 1×1 convolution (Conv_1x1) to adapt it to the input requirements of subsequent networks, and finally the optimized feature map is output.

[0224] S502. Decode and classify the fused feature map at the pixel level to generate an initial semantic segmentation map.

[0225] Decoding can be an operation relative to the encoding stage. It restores the high-dimensional, low-resolution feature map to the same size as the input image through upsampling, feature fusion, and other methods.

[0226] Pixel-level classification can refer to classifying each pixel in a feature map and assigning a unique semantic label to each pixel.

[0227] The initial semantic segmentation map can be the direct output of pixel-level classification, containing semantic category information for each pixel, but it may have issues such as blurred details or inaccurate boundaries, requiring further optimization.

[0228] The fused feature map can be upsampled (e.g., by bilinear interpolation or transposed convolution) to map its low-resolution high-dimensional features back to the same size as the original input image. A 1×1 convolution is used to adjust the number of channels in the fused feature map, matching the number of feature channels with the number of categories in the subsequent classification head, while reducing computational complexity. The decoded high-resolution feature map is then fed into the classification head, where a linear transformation is performed on the feature vector of each pixel, resulting in an output dimension equal to the number of categories in the target task. A Softmax activation function is applied to the output of the classification head, converting the vector of each pixel into a probability distribution corresponding to its category, ensuring that the sum of all category probabilities is 1. The category corresponding to the maximum value of the probability distribution for each pixel is taken as the semantic label for that pixel. The labels of all pixels are then integrated into an image, yielding the initial semantic segmentation map.

[0229] S503. Post-process the initial semantic segmentation map to obtain the farmland plot segmentation map.

[0230] Farmland plot partitioning maps include at least one crop category and plot boundaries.

[0231] Post-processing can address the shortcomings of the initial segmentation image by using a series of morphological operations, optimization algorithms, or rule constraints to correct segmentation errors, refine boundaries, and improve the consistency of results.

[0232] Farmland plot segmentation maps can be the final output after post-processing optimization, which can clearly and accurately delineate the complete boundaries of individual farmland plots, meeting the accuracy requirements of practical applications such as farmland surveying and crop monitoring.

[0233] Through this post-processing step, problems such as noise and blurred boundaries in the initial semantic segmentation map are effectively corrected. The final generated farmland plot segmentation map can accurately reflect the real farmland distribution and provide reliable data support for agricultural applications such as farmland area statistics and crop growth monitoring.

[0234] The implementation details of each step in this application embodiment can be found in the description of the corresponding steps or operations in the above method embodiments; repeated content will not be repeated.

[0235] This embodiment provides an agricultural insurance underwriting method. Based on elongated geographical features, a multi-scale, multi-template dynamic strip convolutional network is used for feature enhancement and fusion to obtain a fused feature map. The fused feature map includes multi-dimensional contextual information. The fused feature map is then decoded and classified at the pixel level to generate an initial semantic segmentation map. Post-processing of the initial semantic segmentation map yields a farmland plot segmentation map, which includes at least one crop category and plot boundaries. By dynamically rotating the strip convolutional kernel to align with local feature directions, elongated geographical features such as roads and rivers in UAV imagery can be extracted more accurately and efficiently, improving recognition accuracy and directional adaptability. This method accurately captures multi-scale, multi-directional elongated features of farmland plots. Through the fusion of multi-dimensional contextual information and post-processing optimization, a high-precision farmland plot segmentation map containing crop categories and clear plot boundaries is generated, significantly improving the accuracy and efficiency of farmland surveying and crop monitoring.

[0236] Figure 9 A schematic diagram of an underwriting device for agricultural insurance provided in this application is shown below. Figure 9 As shown, the agricultural insurance underwriting device 900 provided in this embodiment includes an acquisition module 901, a determination module 902, a segmentation module 903, and a generation module 904.

[0237] The acquisition module 901 is used to acquire drone aerial image data of the target farmland and planting information corresponding to agricultural insurance.

[0238] The determination module 902 is used to input UAV aerial image data into a multi-scale, multi-template dynamic strip convolutional network to determine the elongated geographical features;

[0239] The segmentation module 903 is used to segment the image based on the elongated geographical features to obtain a farmland plot segmentation map, which includes at least one crop category.

[0240] The generation module 904 is used to generate a planting status verification report for the target farmland based on the farmland plot division map, at least one crop category, and planting information.

[0241] In one possible implementation, the determining module 902 is specifically used for:

[0242] Preliminary feature extraction is performed on the drone aerial image data to obtain the first feature map;

[0243] The first feature map is input into the offset network in the multi-scale multi-template dynamic strip convolutional network to obtain the rotation factor, which is used to indicate the direction of local features.

[0244] Based on the rotation factor and the length of the predefined bar convolution kernel, the sampling position coordinate offset is generated;

[0245] Based on the offset of the sampling position coordinates, the first feature map is resampled through interpolation to generate a rotated feature map;

[0246] Based on the bar convolution kernel, perform convolution calculation on the rotated feature map, extract the second feature that is consistent with the direction indicated by the rotation factor, and determine the second feature as a long bar geographic feature.

[0247] In one possible implementation, the determining module 902 is specifically used for:

[0248] The first feature map is input into an offset network, which consists of at least one convolutional layer, at least one batch normalization layer and at least one activation function layer connected in sequence.

[0249] The first feature map is processed by an offset network to output a rotation factor map, wherein the spatial size of the rotation factor map is the same as that of the first feature map and the number of channels is 1.

[0250] Determine the rotation factor based on the rotation factor diagram.

[0251] In one possible implementation, the determining module 902 is specifically used for:

[0252] Based on the rotation factor and the length of the bar convolution kernel, the sampling position of each weight point on the bar convolution kernel on the first feature map is calculated using the coordinate transformation formula, and a convolution kernel position index map is generated.

[0253] Based on the convolution kernel position index map, feature values ​​are sampled from the first feature map using bilinear interpolation, and the sampled feature values ​​are arranged in order to generate a rotated feature map.

[0254] In one possible implementation, the segmentation module 903 is specifically used for:

[0255] Based on the elongated geographical features, a multi-scale, multi-template dynamic strip convolutional network is used for feature enhancement and fusion to obtain a fused feature map, which includes multi-dimensional contextual information.

[0256] The fused feature map is decoded and classified at the pixel level to generate an initial semantic segmentation map;

[0257] The initial semantic segmentation map is post-processed to obtain a farmland plot segmentation map, which includes at least one crop category and plot boundaries.

[0258] In one possible implementation, the generation module 904 is specifically used for:

[0259] Based on the farmland plot division map and at least one crop category, calculate the actual planted area for each crop category;

[0260] The actual planted area of ​​at least one crop category and each crop category is compared with the planted area and crop category in the planting information to obtain the comparison results;

[0261] Based on the comparison results, a verification report is generated. The verification report includes at least: plot identification, comparison results of actual and declared crop categories and areas, and anomaly markers.

[0262] This embodiment provides an underwriting device for agricultural insurance, which can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0263] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Please refer to... Figure 10 The electronic device 1000 may include: a memory 1001, a processor 1002, and a transceiver 1003.

[0264] Memory 1001 is used to store program instructions;

[0265] The processor 1002 is used to execute the program instructions stored in the memory so that the electronic device 1000 performs the above-described method.

[0266] Transceiver 1003 may include a transmitter and / or a receiver. The transmitter may also be referred to as a transmitter, transmitter, transmitting port, or transmitting interface, and the receiver may also be referred to as a receiver, receiving port, or receiving interface, etc. Exemplarily, memory 1001, processor 1002, and transceiver 1003 are interconnected via bus 1004.

[0267] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0268] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0269] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0270] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0271] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0272] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0273] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0274] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0275] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0276] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0277] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0278] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An underwriting method of agricultural insurance, characterized by, include: Obtain drone aerial imagery data of the target farmland and planting information corresponding to agricultural insurance; The drone aerial image data is input into a multi-scale, multi-template dynamic strip convolutional network to determine the elongated geographical features; Image segmentation is performed based on the elongated geographical features to obtain a farmland plot segmentation map, which includes at least one crop category; Based on the farmland plot division map, the at least one crop category, and the planting information, a planting status verification report for the target farmland is generated.

2. The method according to claim 1, characterized in that, The drone aerial imagery data is input into a multi-scale, multi-template dynamic strip convolutional network to determine elongated geographical features, including: Preliminary feature extraction is performed on the drone aerial image data to obtain a first feature map; The first feature map is input into the offset network in the multi-scale multi-template dynamic strip convolutional network to obtain a rotation factor, which is used to indicate the direction of local features. Based on the rotation factor and the predefined length of the bar convolution kernel, the sampling position coordinate offset is generated; Based on the sampling position coordinate offset, the first feature map is resampled through interpolation to generate a rotated feature map; Based on the strip convolution kernel, perform convolution calculation on the rotated feature map, extract a second feature that is consistent with the direction indicated by the rotation factor, and determine the second feature as the strip-shaped geographical feature.

3. The method according to claim 2, characterized in that, The first feature map is input into the offset network in the multi-scale, multi-template dynamic strip convolutional network to obtain the rotation factor, including: The first feature map is input into an offset network, which is composed of at least one convolutional layer, at least one batch normalization layer and at least one activation function layer connected in sequence. The first feature map is processed by the offset network to output a rotation factor map, wherein the spatial size of the rotation factor map is the same as that of the first feature map, and the number of channels is 1. The rotation factor is determined based on the rotation factor diagram.

4. The method according to claim 2, characterized in that, Based on the sampling position coordinate offset, the first feature map is resampled through interpolation to generate a rotated feature map, including: Based on the rotation factor and the length of the bar convolution kernel, the sampling position of each weight point on the bar convolution kernel on the first feature map is calculated using the coordinate transformation formula, and a convolution kernel position index map is generated. Based on the convolution kernel position index map, feature values ​​are sampled from the first feature map using bilinear interpolation, and the sampled feature values ​​are arranged in order to generate a rotated feature map.

5. The method according to any one of claims 1-4, characterized in that, Image segmentation is performed based on the elongated geographical features to obtain a farmland plot segmentation map, including: Based on the elongated geographical features, the multi-scale, multi-template dynamic strip convolutional network is used to enhance and fuse features to obtain a fused feature map, which includes multi-dimensional contextual information. The fused feature map is decoded and classified at the pixel level to generate an initial semantic segmentation map; The initial semantic segmentation map is post-processed to obtain a farmland plot segmentation map, which includes at least one crop category and plot boundaries.

6. The method according to claim 5, characterized in that, Based on the farmland plot division map, the at least one crop category, and the planting information, a planting status verification report for the target farmland is generated, including: Based on the farmland plot division map and the at least one crop category, calculate the actual planted area of ​​each crop category; The actual planting area of ​​the at least one crop category and each crop category is compared with the planting area and crop category in the planting information to obtain the comparison result; Based on the comparison results, a verification report is generated, which includes at least: plot identification, comparison results of actual and declared crop categories and areas, and anomaly markers.

7. An underwriting device for agricultural insurance, characterized in that, include: The acquisition module is used to acquire drone aerial imagery data of the target farmland and planting information corresponding to agricultural insurance. The determination module is used to input the UAV aerial image data into a multi-scale, multi-template dynamic strip convolutional network to determine the elongated geographical features; The segmentation module is used to segment the image based on the elongated geographical features to obtain a farmland plot segmentation map, wherein the farmland plot segmentation map includes at least one crop category; The generation module is used to generate a planting status verification report for the target farmland based on the farmland plot division map, the at least one crop category, and the planting information.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.

Images