A method for quickly extracting a mango planting area boundary, an electronic device, and a storage medium

Abstract

The application discloses a mango planting area boundary extraction method, an electronic device and a storage medium, and belongs to the technical field of mango planting area recognition. In order to solve the problem of fast and accurate recognition of a mango planting area, the application comprises the following steps: establishing a sample data set and dividing the sample data set into a training set and a test set; constructing a multi-scale channel attention mechanism module; constructing a feature extraction module and a down-sampling module; constructing a mango planting area edge recognition model SLMCnet; inputting the training set into the mango planting area edge recognition model SLMCnet for training, obtaining model weights after the training is completed, and then evaluating and optimizing the trained mango planting area edge recognition model by using the test set to obtain an optimal mango planting area boundary extraction segmentation model; and the optimal mango planting area boundary extraction segmentation model is used to segment a target area of the mango planting area. The application can efficiently and accurately realize extraction of the mango planting area in remote sensing images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mango planting area recognition, and in particular to a method for quickly extracting the boundary of a mango planting area, an electronic device and a storage medium. BACKGROUND

[0002] In the traditional process of agricultural insurance for mangoes, data such as the number of insured plants are involved, and the area of the mango planting area is usually estimated by the number of plants through agricultural insurance. Due to the strong subjectivity, the area of the mango planting area is not accurate, resulting in inaccurate insured plant number. Meanwhile, in the remote sensing monitoring of the mango planting area, the problem of displacement and misplacement of massive remote sensing images of multiple frequencies cannot provide real-time and accurate remote sensing monitoring results, making it difficult to implement fine claim settlement tasks. With the development of computer vision technology, the recognition of the mango planting area has higher accuracy and efficiency than the traditional manual recognition method. However, the artificial selection of features requires a large amount of time and money, which leads to the fact that the accuracy of traditional machine learning is not ideal. Deep learning technology can automatically extract features, and the trained model has better recognition accuracy and robustness. However, the current mango planting area recognition method based on deep learning mainly focuses on traditional convolutional neural networks such as Unet and DeeplabV3+, which have large computational overhead and are difficult to meet the actual application requirements. SUMMARY

[0003] In order to solve the problem of fast and accurate recognition of the mango planting area, the present application provides a method for quickly extracting the boundary of a mango planting area, an electronic device and a storage medium.

[0004] The present application is achieved by the following technical solutions:

[0005] A method for quickly extracting the boundary of a mango planting area, comprising the following steps:

[0006] S1. Remote sensing image data of the mango planting area is collected to obtain a mango planting area image, and a sample data set is established;

[0007] S2. The data in the sample data set obtained in step S1 is sequentially cropped according to the pixel arrangement order, and the mango planting area image obtained in step S1 is subjected to data augmentation by means of flipping, color enhancement and rotation, and the augmented image sample is added to the sample data set;

[0008] S3. The sample data set obtained in step S2 is divided into a training set and a test set;

[0009] S4. A multi-scale channel attention mechanism module MSFE is constructed;

[0010] S5. Constructing the feature extraction module SMBU and the down-sampling module SMDU based on the multi-scale channel attention mechanism module MSFE obtained in step S4;

[0011] S6. Constructing the mango planting area edge recognition model SLMCnet based on the feature extraction module SMBU and the down-sampling module SMDU in step S5;

[0012] S7. Inputting the training set obtained in step S3 into the mango planting area edge recognition model SLMCnet obtained in step S6 for training, obtaining the model weight after the training is completed, and then evaluating and optimizing the trained mango planting area edge recognition model SLMCnet by using the test set obtained in step S3 to obtain an optimal mango planting area boundary extraction and segmentation model;

[0013] S8. Using the optimal mango planting area boundary extraction and segmentation model obtained in step S7 to perform mango planting area segmentation on a target region.

[0014] Further, the mango planting area image obtained in step S1 is used to make a mango planting area mask sample data by using the Arcgis software, and a sample data set is constructed.

[0015] Further, the data division ratio of the training set and the test set in step S3 is 8:2.

[0016] Further, the specific implementation method of step S4 includes the following steps:

[0017] S4.1. Constructing a multi-scale attention mechanism MSCA by using 3 strip convolution blocks with different sizes, and the calculation formula is as follows:

[0018] M=Dw_Conv 5×5 (input) (1)

[0019] M7=Dw_Conv 1×7 (Dw_Conv 7×1 (M)) (2)

[0020] M 11 =Dw_Conv 1×11 (Dw_Conv 11×1 (M)) (3)

[0021] M 21 =Dw_Conv 1×21 (Dw_Conv 21×1 (M)) (4)

[0022] Att MSCA =Conv 1×1 (Concat(M7,M11 ,M 21 ) (5)

[0023] wherein, M represents the feature extracted by 5x5 depth convolution kernel, M 11 , M 21 represent the features extracted by 7, 11, 21 size strip depth convolution kernel respectively, Dw_Conv represents depth convolution kernel, Conv represents convolution kernel, Concat represents matrix connection, Att MSCA represents the feature output of multi-scale attention mechanism MSCA;

[0024] S4.2. The channel attention mechanism SE is constructed by global average pooling, full connection layer and Sigmod function, and the calculation formula is as follows:

[0025]

[0026] z′ c =FC1(z c )=W1z c +b1 (7)

[0027] a c =SwishReLu(z’ c ) (8)

[0028]

[0029] Att SE =Sigmod(FC2(SwishReLu(FC1(z c (c))))) (12)

[0030] wherein, x c,i,j is the value of the i-th row and the j-th column of the input feature map in the channel c, H and W are the height and width of the feature map respectively, z c is the feature map after global average pooling;

[0031] z′ c represents the output after the dimension compression of z c by the first full connection layer FC1, W1 is the first weight matrix, and b1 is the first bias term;

[0032] a c represents the output feature of z′ c after the SwishReLu activation function, and the output feature value of z′ c is taken as the input value x in SwishReLu;

[0033] s c represents the output of ac The output is restored to the original dimension, W2 is the second weight matrix, and b2 is the first bias term.

[0034] Att SE represents the feature output of the channel attention mechanism SE; represents s c The weight value is output by the Sigmod activation function.

[0035] S4.3. Construct a multi-scale channel attention mechanism module MSFE by multiplying MSCA and SE, and restore it by 1x1 convolution. The calculation formula is as follows:

[0036]

[0037] Wherein, represents matrix point multiplication, and input represents the input matrix.

[0038] Further, the specific implementation method of step S5 includes the following steps:

[0039] S5.1. Construct a feature extraction module SMBU, which is composed of channel split, channel shuffle operation, 2 1x1 convolution kernels, 1 3x3 deep convolution kernel, 1 MSFE module, 2 batch normalization BN and 2 SwishReLU activation functions. The calculation formula is as follows:

[0040] C, C' = channel split(input) (14)

[0041] S 1_1 = Conv 1×1 (SR(BN(Dw_Conv 3×3 (BN(MSFE(SR(BN(Conv 1×1 (C))))))) (15)

[0042] S SMBU = Channel shuffle(Concat(C',S 1_1 )) (16)

[0043] Wherein, channel split represents channel separation operation, which divides the input feature channel input into equal number of channels C, C' and S 1_1 represents the feature map output after feature extraction of channel C, SR is SwishReLU activation function, BN is batch normalization operation, Channel shuffle is channel shuffle operation, which is used to enhance the information interaction ability, and S SMBU represents the feature map output after connection and channel shuffle of the feature map on channels C, C';

[0044] S5.2. Constructing a down-sampling module SMDU, which is composed of a Channel shuffle operation, a 1x1 convolution kernel, a 3x3 depthwise convolution kernel with a step of 2, an MSFE module, a batch normalization BN and a Swish ReLU activation function, and the calculation formula is as follows:

[0045] S 2_1 = Conv 1×1 (SR(BN(DwConv 3×3,2 (BN(MSFE(SR(BN(Conv 1×1 (input))))))))) (17)

[0046] S 2_2 = MSFE(BN(Dw_Conv 3×3,2 (BN(Conv 1×1 (input))))) (18)

[0047] S SMDU = Channel shuffle(Concat(S 2_1 ,S 2_1 )) (19)

[0048] Wherein, S 2_1 represents the feature map extracted by the right branch of the down-sampling module, S 2_2 represents the feature map extracted by the left branch, and S SMDU represents the output of the down-sampling module SMDU after the output features of the left and right branches are connected and channel shuffled.

[0049] Further, step S6 constructs a mango planting area edge recognition model SLMCnet, which is composed of an input, a 1x1 convolution kernel, a maximum pooling, four feature extraction modules Stage and an output layer, wherein the first feature extraction module Stage1 is composed of one SMBU; the second feature extraction module Stage2 is composed of one SMDU and one SMBU; the third feature extraction module Stage3 is composed of one SMDU and two SMBUs; the fourth feature extraction module Stage4 is composed of one SMDU and one SMBU, then connected with a 5x5 convolution kernel, then the outputs of Stage1, Stage2, Stage3 and Stage4 are subjected to a Concat operation, the output after the Concat operation is connected with a global average pooling, and the output of the global average pooling is connected with a multi-layer perceptron MLP to output a segmentation result.

[0050] Further, step S7 inputs the training set obtained in step S3 into the mango planting area edge recognition model SLMCnet obtained in step S6 for training, and obtains model weights after the training is completed, the model training parameters are that the training batch size is 6, the training times are 200, the optimizer is Adam, the initial learning rate is 0.0005, the learning rate adopts a cosine annealing function, the activation function is SwishReLU, and the loss function is Dice Loss; the obtained model parameters are loaded into the SLMCnet network model, and the average intersection over union mIoU, the accuracy Precision, the recall rate Recall, the F1 score and the training time min / epoch each time are used to evaluate the precision and efficiency of model segmentation by using the test set in step S3.

[0051] An electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the fast mango planting area boundary recognition method when executing the computer program.

[0052] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the fast mango planting area boundary recognition method.

[0053] Advantages of the present application:

[0054] The fast mango planting area boundary recognition method provided by the present application designs a multi-scale channel attention mechanism module MSFE, realizes multi-dimensional and multi-scale feature information extraction, and more effectively extracts the correlation between feature information and target information; the feature extraction module SMBU and the down-sampling module SMDU module are designed, the channels are evenly divided through the channel split operation, the parameter amount of the 1x1 convolution is effectively reduced, the MSFE attention mechanism module is added, and the training efficiency and accuracy of the model are improved; the mango planting area edge recognition model SLMCnet is designed, the number of layers of the feature extraction module SMBU and the down-sampling module SMDU module and the splicing between different sizes are adjusted, and the segmentation ability and training efficiency of the model for different sizes and shapes of features are improved; the present application can efficiently and accurately realize the extraction of the mango planting area in a large amount of remote sensing image with multiple frequencies, improve the agricultural insurance claim efficiency of economic crops, and provide technical support for reducing the economic losses of fruit farmers caused by disasters. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The flowchart of the fast mango planting area boundary recognition method provided by the present application;

[0056] Figure 2 The network mechanism schematic diagram of the SLMCnet model;

[0057] Figure 3 The structure diagram of the MSFE attention mechanism used in the application is shown in the figure;

[0058] Figure 4 The improved SMBU module of the application;

[0059] Figure 5 The improved SMDU module of the application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the application clearer and more understandable, the application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and are not used to limit the application, that is, the described specific embodiments are only a part of the embodiments of the application, but not all the specific embodiments. The components of the specific embodiments of the application described and shown in the drawings herein can be arranged and designed in various different configurations, and the application can also have other embodiments.

[0061] Therefore, the detailed description of the specific embodiments of the application provided in the drawings below is not intended to limit the scope of the claimed application, but only represents selected specific embodiments of the application. Based on the specific embodiments of the application, all other specific embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0062] In order to further understand the invention content, characteristics and effects of the application, the following specific embodiments are exemplified, and the drawings are Figure 1 -Appendix Figure 5 The detailed description is as follows:

[0063] Example 1:

[0064] A fast extraction method for identifying the boundary of mango planting area, comprising the following steps:

[0065] S1. Remote sensing image data of mango planting area is collected to obtain mango planting area image, and sample data set is established;

[0066] Further, the mango planting area image obtained in step S1 is used to make a mango planting area mask sample data by Arcgis software, and a sample data set is constructed;

[0067] S2. The data in the sample data set obtained in step S1 is cut according to the pixel arrangement order in turn, and the mango planting area image obtained in step S1 is subjected to data expansion by means of flipping, color enhancement and rotation, and the expanded image sample is added to the sample data set;

[0068] Further, the sample data set obtained in step S2 is cropped in turn according to the pixel arrangement order to 512x512 size;

[0069] S3. The sample data set obtained in step S2 is divided into a training set and a test set;

[0070] Further, the data division ratio of the training set and the test set in step S3 is 8:2;

[0071] S4. Constructing a multi-scale channel attention mechanism module MSFE;

[0072] Further, the specific implementation method of step S4 includes the following steps:

[0073] S4.1. Constructing a multi-scale attention mechanism MSCA through 3 different size strip convolution blocks, the calculation formula is as follows:

[0074] M = Dw_Conv 5×5 (input) (1)

[0075] M7 = Dw_Conv 1×7 (Dw_Conv 7×1 (M)) (2)

[0076] M 11 = Dw_Conv 1×11 (Dw_Conv 11×1 (M)) (3)

[0077] M 21 = Dw_Conv 1×21 (Dw_Conv 21×1 (M)) (4)

[0078] Att MSCA = Conv 1×1 (Concat(M7,M 11 ,M 21 ) (5)

[0079] Wherein, M represents the features extracted by the 5x5 depth convolution kernel, M7, M 11 , M 21 respectively represent the features extracted by the strip depth convolution kernel of 7, 11, 21 size, Dw_Conv represents the depth convolution kernel, Conv represents the convolution kernel, Concat represents the matrix connection, Att MSCA represents the feature output of the multi-scale attention mechanism MSCA;

[0080] S4.2. Channel attention mechanism SE is constructed by global average pooling, full connection layer and Sigmod function, and the calculation formula is as follows:

[0081]

[0082] z′ c =FC1(z c )=W1z c +b1 (7)

[0083] a c =SwishReLu(z’ c ) (8)

[0084]

[0085] s c =FC2(a c )=W2a c +b2 (10)

[0086]

[0087] Att SE =Sigmod(FC2(SwishReLu(FC1(z c (c))))) (12)

[0088] Wherein, x c,i,j is the value of the input feature map on the i-th row and the j-th column in the channel c, H and w are the height and width of the feature map respectively, z c is the feature map after global average pooling;

[0089] z′ c represents the output of z c dimension compression by the first full connection layer FC1, W1 is the first weight matrix, and b1 is the first bias term;

[0090] a c represents the output feature of z′ c after the SwishReLu activation function, and the output feature value of z′ c is taken as the input value x in SwishReLu;

[0091] s c represents the output of a c restored to the original dimension by the second full connection layer FC2, W2 is the second weight matrix, and b2 is the first bias term;

[0092] Att SE represents the feature output of channel attention mechanism SE; s cThe weight value is output through a Sigmod activation function;

[0093] S4.3. Construct a multi-scale channel attention mechanism module MSFE by multiplying the MSCA and the SE, and restore it through 1x1 convolution, the calculation formula is as follows:

[0094]

[0095] Wherein, represents matrix point multiplication, and input represents an input matrix.

[0096] S5. Construct a feature extraction module SMBU and a down-sampling module SMDU based on the multi-scale channel attention mechanism module MSFE obtained in step S4;

[0097] Further, the specific implementation method of step S5 includes the following steps:

[0098] S5.1. Construct a feature extraction module SMBU, which is composed of a channel split, a channel shuffle operation, two 1x1 convolution kernels, one 3x3 depth convolution kernel, one MSFE module, two batch normalization BN and two SwishReLU activation functions, the calculation formula is as follows:

[0099] C, C' = channel split(input) (14)

[0100] S 1_1 = Conv 1×1 (SR(BN(Dw_Conv 3×3 (BN(MSFE(SR(BN(Conv 1×1 (C))))))) (15)

[0101] S SMBU = Channel shuffle(Concat(C',S 1_1 )) (16)

[0102] Wherein, channel split represents a channel separation operation, which separates the input of the input feature channel into equal number of channels C, C' and S 1_1 represents the feature map output after feature extraction of channel C, SR is the SwishReLU activation function, BN is the batch normalization operation, Channel shuffle is the channel shuffle operation, which is used to enhance the information interaction ability, S SMBU represents the feature map output after connection and channel shuffle of the feature map on channel C, C';

[0103] S5.2. Constructing a down-sampling module SMDU, which is composed of a Channel shuffle operation, a 1x1 convolution kernel, a 3x3 depthwise convolution kernel with a step of 2, an MSFE module, a batch normalization BN and a Swish ReLU activation function, and the calculation formula is as follows:

[0104] S 2_1 = Conv 1×1 (SR(BN(DwConv 3×3,2 (BN(MSFE(SR(BN(Conv 1×1 (input))))))))) (17)

[0105] S 2_2 = MSFE(BN(Dw_Conv 3×3,2 (BN(Conv 1×1 (input))))) (18)

[0106] S SMDU = Channel shuffle(Concat(S 2_1 ,S 2_1 )) (19)

[0107] wherein, S 2_1 represents the feature map extracted by the right branch of the down-sampling module, S 2_2 represents the feature map extracted by the left branch, and S SMDU represents the output of the down-sampling module SMDU after the output features of the left and right branches are connected and channel shuffled.

[0108] S6. Based on the feature extraction module SMBU and the down-sampling module SMDU in step S5, a mango planting area edge recognition model SLMCnet is constructed.

[0109] Further, the step S6 of constructing the mango planting area edge recognition model SLMCnet is composed of an input, a 1x1 convolution kernel, a maximum pooling, four feature extraction modules Stage and an output layer, wherein the first feature extraction module Stage1 is composed of one SMBU; the second feature extraction module Stage2 is composed of one SMDU and one SMBU; the third feature extraction module Stage3 is composed of one SMDU and two SMBUs; and the fourth feature extraction module Stage4 is composed of one SMDU and one SMBU, and then connected with a 5x5 convolution kernel. Then, the outputs of Stage1, Stage2, Stage3 and Stage4 are subjected to a Concat operation, and the output after the Concat operation is connected with a global average pooling, and the output of the global average pooling is connected with a multi-layer perception MLP to output a segmentation result.

[0110] S7. Input the training set obtained in step S3 into the mango planting area edge recognition model SLMCnet obtained in step S6 for training, and obtain the model weight after the training is completed. Then, the test set obtained in step S3 is used to evaluate and optimize the trained mango planting area edge recognition model SLMCnet, and an optimal mango planting area boundary extraction and segmentation model is obtained.

[0111] Further, in step S7, the training set obtained in step S3 is input into the mango planting area edge recognition model SLMCnet obtained in step S6 for training, and the model weight is obtained after the training is completed. The model training parameters are that the training batch size is 6, the training times are 200, the optimizer is Adam, the initial learning rate is 0.0005, the learning rate uses a cosine annealing function, the activation function is SwishReLU, and the loss function is Dice Loss. The obtained model parameters are loaded into the SLMCnet network model, and the test set in step S3 is used to evaluate the precision and efficiency of the model segmentation by using the average intersection over union mIoU, the accuracy Precision, the recall Recall, the F1 score, and the training time min / epoch each time.

[0112] Further, the specific calculation formula is as follows:

[0113]

[0114] wherein k represents the number of classifications, p ij represents that i is predicted as j, which is a false negative FN; p ji represents that j is predicted as i, which is a false positive FP; p ii represents that i is predicted as i, which is a true positive TP;

[0115]

[0116] wherein TP represents a positive sample that is correctly predicted in classification, i.e., a true positive, and FP represents a negative sample that is incorrectly predicted in classification, i.e., a false positive;

[0117]

[0118] wherein TP represents a positive sample that is correctly predicted in classification, i.e., a true positive, and FN represents a positive sample that is incorrectly predicted in classification, i.e., a false negative;

[0119]

[0120] wherein Precision represents the accuracy, and Recall represents the recall.

[0121] S8. The optimal mango planting area boundary extraction and segmentation model obtained in step S7 is used to perform mango planting area segmentation of a target region.

[0122] For the purpose of understanding the technical effects of the present application, the effects of the method of the embodiment of the present application and the prior art are compared as follows:

[0123] Table 1 Comparison table of effects of the method of the embodiment of the present application and the prior art

[0124]

[0125] As can be seen from Table 1, compared with other disclosed methods, the method proposed in the present application has better improved model precision, and all evaluation indexes are optimal, the average intersection over union (mIoU) of mango planting area boundary segmentation reaches 91.3%, the single body accuracy (Precision) of mango planting area boundary segmentation reaches 93.2%, and the mango planting area boundary segmentation efficiency reaches 5.07 min / epoch.

[0126] In summary: the embodiment fully utilizes the designed multi-scale channel attention mechanism module MSFE, realizes multi-dimensional and multi-scale feature information extraction, and more effectively extracts the correlation between feature information and target information; the feature extraction module SMBU and the down-sampling module SMDU module are designed, the channel is evenly divided through the channel split operation, the parameter amount of the 1x1 convolution is effectively reduced, the MSFE attention mechanism module is added, and the training efficiency and precision of the model are improved; the mango planting area edge recognition model SLMCnet is designed, the number of layers of the feature extraction module SMBU and the down-sampling module SMDU module and the splicing between different sizes are adjusted, and the segmentation ability and training efficiency of the model for different sizes and shape features are improved.

[0127] Embodiment 2:

[0128] An electronic device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of the fast mango planting area boundary extraction and recognition method of embodiment 1.

[0129] The computer device of the present application can be a device including a processor and a memory, such as a single-chip microcomputer including a central processing unit. Moreover, the processor is used to execute the computer program stored in the memory to realize the steps of the fast mango planting area boundary extraction and recognition method.

[0130] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0131] The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs required by at least one function (such as a sound playing function, an image playing function, etc.), etc.; and the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0132] Embodiment 3:

[0133] A computer readable storage medium, having stored thereon a computer program, the computer program being executed by a processor to implement the method of quickly extracting the boundary of the mango planting area according to the embodiment 1.

[0134] The computer readable storage medium of the present application can be any form of storage medium readable by the processor of the computer device, including but not limited to non-volatile memory, volatile memory, ferroelectric memory, etc., and the computer readable storage medium has stored thereon a computer program, when the processor of the computer device reads and executes the computer program stored in the memory, the steps of the method of quickly extracting the boundary of the mango planting area can be implemented.

[0135] The computer readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB (Universal Serial Bus) flash disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a ROM (Read-Only Memory), a RAM (Random Access Memory), an electrical carrier signal, a telecommunications signal, and a software distribution medium, and the like. It should be noted that the computer readable medium contains contents that can be appropriately added, deleted or modified according to requirements of legislation and patent practices in a jurisdiction, for example, in some jurisdictions, according to legislation and patent practices, the computer readable medium does not include an electrical carrier signal and a telecommunications signal.

[0136] It should be noted that the terms "first" and "second" and the like, such relational terms are used only to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0137] Although the present application has been described above with reference to specific embodiments, various modifications can be made to it without departing from the scope of the present application, and equivalent components can be substituted therefor. In particular, each feature in the specific embodiments disclosed by the present application can be used in any combination with each other, provided that there is no structural conflict. The fact that these combinations are not described in the present specification is merely due to the consideration of omitting the description and saving resources, and therefore the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for fast extraction of mango plantation area boundary identification, characterized by, The method comprises the following steps: S1. Remote sensing image data of mango planting area is collected to obtain a mango planting area image, and a sample data set is established; S2. The data in the sample data set obtained in step S1 is sequentially cropped according to the pixel arrangement order, and the mango planting area image obtained in step S1 is subjected to data expansion through flipping, color enhancement and rotation mode, and the expanded image sample is added to the sample data set; S3. The sample data set obtained in step S2 is divided into a training set and a test set; S4. A multi-scale channel attention mechanism module MSFE is constructed; The specific implementation method of step S4 comprises the following steps: S4.

1. A multi-scale attention mechanism MSCA is constructed through three different size strip convolution blocks, and the calculation formula is as follows: M = Dw_Conv 5×5 (input) (1) M7 = Dw_Conv 1×7 (Dw_Conv 7×1 (M)) (2) M 11 = Dw_Conv 1×11 (Dw_Conv 11×1 (M)) (3) M 21 = Dw_Conv 1×21 (Dw_Conv 21×1 (M)) (4) Att MSCA = Conv 1×1 (Concat(M7, M 11 , M 21 )) (5) where M represents the feature extracted by a 5x5 depth convolution kernel, M 11 , M 21 represent the features extracted by 7, 11, and 21 size strip depth convolution kernels respectively, Dw_Conv represents a depth convolution kernel, Conv represents a convolution kernel, Concat represents matrix connection, Att MSCA represents the feature output of the multi-scale attention mechanism MSCA. S4.

2. A channel attention mechanism SE is constructed through global average pooling, a fully connected layer and a Sigmod function; S4.

3. A multi-scale channel attention mechanism module MSFE is constructed by multiplying MSCA and SE, and is restored through 1×1 convolution; S5. A feature extraction module SMBU and a down-sampling module SMDU are constructed based on the multi-scale channel attention mechanism module MSFE obtained in step S4; The specific implementation method of step S5 comprises the following steps: S5.

1. The feature extraction module SMBU is constructed and comprises a channel split, a channel shuffle operation, two 1×1 convolution kernels, one 3×3 deep convolution kernel, one MSFE module, two batch normalization BNs and two SwishReLU activation functions; S5.

2. The down-sampling module SMDU is constructed and comprises a channel shuffle operation, a 1×1 convolution kernel, a 3×3 deep convolution kernel with a step of 2, an MSFE module, a batch normalization BN and a SwishReLU activation function; S6. A mango planting area edge recognition model SLMCnet is constructed based on the feature extraction module SMBU and the down-sampling module SMDU in step S5; The mango planting area edge recognition model SLMCnet constructed in step S6 comprises an input, a 1×1 convolution kernel, a maximum pooling, four feature extraction modules Stage and an output layer, wherein the first feature extraction module Stage1 comprises one SMBU; the second feature extraction module Stage2 comprises one SMDU and one SMBU; the third feature extraction module Stage3 comprises one SMDU and two SMBUs; and the fourth feature extraction module Stage4 comprises one SMDU and one SMBU, and then connects a 5×5 convolution kernel, and then the outputs of Stage1, Stage2, Stage3 and Stage4 are subjected to a Concat operation, the output after the Concat operation is connected with a global average pooling, and the output of the global average pooling is connected with a multi-layer perception MLP to output a segmentation result; S7. The training set obtained in step S3 is input into the mango planting area edge recognition model SLMCnet obtained in step S6 for training, and after the training is completed, the model weight is obtained. Then, the test set obtained in step S3 is used to evaluate and optimize the trained mango planting area edge recognition model SLMCnet, and the optimal mango planting area boundary extraction segmentation model is obtained. S8. The optimal mango planting area boundary extraction segmentation model obtained in step S7 is used to segment the target area of the mango planting area.

2. A method for rapid extraction of mango plantation area boundary identification as claimed in claim 1 wherein, In step S1, the obtained mango planting area image is used to make a mango planting area mask sample data through Arcgis software, and a sample data set is constructed.

3. A method for rapid extraction of mango plantation area boundary identification as claimed in claim 2, wherein, In step S3, the data division ratio of the training set and the test set is 8:

2.

4. The method according to claim 3, wherein, S4.

2. The calculation formula is as follows: Att SE = Sigmod(FC2(SwishReLu(FC1(z c (c))))) (12) wherein x c,i,j is the value of the input feature map at the i-th row and j-th column on channel c, H and W are the height and width of the feature map, respectively, z c is the feature map after global average pooling; z' c represents passing z through a first fully connected layer FC1 c dimensionally compressed output, W1 is a first weight matrix, and b1 is a first bias term; a c representing z' c output feature value of z' c the output feature value of z' s c denotes passing a through a second fully connected layer FC2 c output restored to the original dimension, W2 is a second weight matrix, and b2 is a second bias term; Att SE representing the feature output of the channel attention mechanism SE; represents s c output weight values through the Sigmod activation function S4.

3. The calculation formula is as follows: wherein, denotes a matrix point multiplication, and input denotes an input matrix.

5. The method according to claim 4, wherein, S5.

1. The calculation formula is as follows: C, C' = channel split(input) (14) S 1_1 = Conv 1×1 (SR(BN(Dw_Conv 3×3 (BN(MSFE(SR(BN(Conv 1×1 (C))))))))) (15) S SMBU = Channel shuffle(Concat(C', S 1_1 )) (16) Wherein, the channel split represents a channel separation operation, and separates the input of the input feature channel into an equal number of channels C, C' 1_1 The feature map output after the channel C passes through the feature extraction, SR is the SwishReLU activation function, BN is the batch normalization operation, and Channel shuffle is the channel shuffle operation, which is used to enhance the information interaction ability, S SMBU The feature map output after the channel C, C' passes through the connection and channel shuffle S5.

2. The calculation formula is as follows: S 2_1 = Conv 1×1 (SR(BN(DwConv 3×3,2 (BN(MSFE(SR(BN(Conv 1×1 (input))))))))) (17) S 2_2 = MSFE(BN(Dw_Conv 3×3,2 (BN(Conv 1×1 (input))))) (18) S SMDU = Channel shuffle(Concat(S 2_1 ,S 2_2 )) (19) wherein S 2_1 represents the feature map extracted by the right branch of the down-sampling module, S 2_2 represents the feature map extracted by the left branch, S SMDU represents the output of the down-sampling module SMDU after the output features of the left and right branches are connected and channel shuffled.

6. A method for fast extraction of mango plantation area boundary identification as claimed in claim 5 wherein, In step S7, the training set obtained in step S3 is input into the mango planting area edge recognition model SLMCnet obtained in step S6 for training, and after the training is completed, the model weight is obtained. The model training parameters are that the training batch size is 6, the training times are 200, the optimizer is Adam, the initial learning rate is 0.0005, the learning rate uses a cosine annealing function, the activation function is SwishReLU, and the loss function is Dice Loss. The obtained model parameters are loaded into the SLMCnet network model, and the test set in step S3 is used to evaluate the precision and efficiency of the model segmentation by using the average intersection over union mIoU, the accuracy Precision, the recall Recall, the F1 score, and the training time min / epoch each time.

7. An electronic device, comprising: The computer program is executed by the processor to realize the steps of the method for quickly extracting the mango planting area boundary recognition method according to any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method for quickly extracting the mango planting area boundary recognition method according to any one of claims 1-6.

Images