Farmland plot recognition method based on optical-Ka band SAR feature fusion

Abstract

The present application relates to a farmland plot recognition method based on optical-Ka band SAR feature fusion, which uses an improved ResNet50 network to extract features from optical remote sensing images, uses a complex convolution network to extract features from Ka-SAR images, triggers a gating mechanism when the cloud coverage of the optical remote sensing image exceeds a threshold, and uses a cross-modal attention fusion mechanism for feature fusion if the cloud coverage of the optical remote sensing image does not exceed the threshold; the obtained fusion features are output as farmland plot recognition probability maps based on a UNet++ decoder, and a geometric constraint loss function is used to optimize and iterate the results to obtain the final recognition model. The present application uses a cross-modal attention fusion method for feature fusion, realizes physical-level collaborative fusion of data features, and solves the problems of incomplete recognition of farmland plots in cloudy scenes and ignoring the specific regular geometric properties of farmland, which leads to output results that do not meet the needs of agricultural management.

Description

TECHNICAL FIELD

[0001] The application relates to a farmland plot recognition method based on optical-Ka band SAR feature fusion, and belongs to the technical field of remote sensing image processing and image recognition. BACKGROUND

[0002] With the development of precision agriculture and digital agriculture, high-precision recognition of farmland plots has become a core requirement of agricultural informatization construction. Remote sensing technology has become a main means of farmland monitoring due to its advantages of wide range and periodic observation. At present, the mainstream technologies mainly include optical remote sensing recognition and SAR remote sensing recognition, but these technologies have significant defects in farmland plot recognition. The optical remote sensing recognition technology uses visible light-near infrared bands (such as Landsat-8 OLI and Sentinel-2 MSI) to obtain vegetation indexes (NDVI, EVI, etc.), and realizes plot recognition through threshold segmentation or machine learning, but this method is highly dependent on weather, and high cloud coverage will lead to poor extraction results. The SAR remote sensing recognition technology uses the backscattering coefficient and coherence features of synthetic aperture radar (such as Sentinel-1 C band and TerraSAR-X X band) to realize ground type recognition through polarization decomposition, but this method will cause spot noise to make the texture inside the plot uneven, and side imaging will cause problems such as overlap and shadow, thereby affecting the recognition and extraction accuracy.

[0003] There are also existing technologies for crop recognition using multi-source remote sensing data, such as patent CN110189616A, which discloses a crop mapping method using multi-temporal SAR (GaoFen-3) and optical images (GaoFen-2) as data sources. The method uses the "map" information of the farmland plot structure provided by GaoFen-2 images and the polarization scattering and texture information of the ground features provided by GaoFen-3 images for collaborative processing to accurately distinguish crop types and extract planting areas. Although this technology emphasizes the collaborative processing of SAR and optical images, it does not specifically process the characteristics of Ka-SAR due to its novelty and high resolution, high texture, and rich feature richness. The technology also does not solve the problems of high cloud coverage of optical images and the neglect of the regular geometric properties (orthogonal boundaries, rectangular plots) specific to farmland by the segmentation network, resulting in output results that do not meet the needs of agricultural management. SUMMARY

[0004] The purpose of the present application is to overcome the above-mentioned deficiencies and provide a farmland plot recognition method based on optical-Ka band SAR feature fusion. This method adapts to Ka band SAR satellite images and can better fuse the features of Ka-SAR images and optical images to take advantage of their complementary advantages.

[0005] The technical scheme adopted by the present application is:

[0006] The farmland plot recognition method based on optical-Ka band SAR feature fusion comprises the following steps:

[0007] S1. Preprocess and image register the Ka band SAR image (i.e. Ka-SAR image) and the optical remote sensing image;

[0008] S2. Extract features of the optical remote sensing image by using an improved ResNet50 network: calculate the normalized difference vegetation index (NDVI) of the optical remote sensing image, extend the NDVI channel, then input the improved ResNet50 network backbone for feature extraction, then input the spectral attention module for spectral attention weight calculation, and output the extracted features;

[0009] S3. Extract features of the Ka-SAR image by using a complex convolution network: input the intensity, phase and coherence data of the Ka-SAR image to construct, obtain the corresponding complex features after complex convolution layer processing, and then complete feature extraction through a complex residual module, a scattering mechanism decomposition module, an amplitude conversion module and a noise removal and spot removal module;

[0010] S4. Determine the fusion mechanism of the extracted features according to the cloud coverage of the optical image, when the cloud coverage of the optical remote sensing image exceeds the threshold, trigger the gating mechanism, use the Ka-SAR features as the main feature and the optical image features as the auxiliary feature for feature fusion, if the cloud coverage of the optical remote sensing image does not exceed the threshold, use the cross-modal attention fusion mechanism for feature fusion;

[0011] S5. Output the farmland plot recognition probability map based on the UNet++ decoder of the obtained fusion features, and use a geometric constraint loss function to optimize and iterate the result to obtain the final recognition model.

[0012] In the above method, the image registration of step S1 uses a sub-pixel registration transformation model to register the Ka-SAR image and the corresponding optical image data, and ensures that the row and column numbers of the two images are consistent, and the transformation model is as follows:

[0013] ,

[0014] In the formula, x’ 、 y’ is the coordinate of the registration target image, x 、 y is the coordinate of the image to be registered, t x 、 t y is a translation parameter, a 、 b ,c , d is a rotation parameter.

[0015] The NDVI channel expansion in step S2 inputs the Red, Green, Blue and Nir bands of the optical image, and then calculates the NDVI (normalized vegetation index). The NDVI and the Red, Green and Blue bands are combined into 4-band data input into the improved ResNet50 network trunk to improve the farmland feature extraction effect of the optical image. The NDVI calculation formula is as follows:

[0016] .

[0017] The improved ResNet50 network trunk sequentially performs initial convolution layer processing, maximum pooling layer processing and four-stage residual processing (stage 1: maintaining resolution, stage 2: downsampling, stage 3: high-level semantics, and stage 4: maintaining resolution) on the input composed of NDVI, Red, Green and Blue.

[0018] The spectral attention module takes the output of the improved ResNet50 network trunk as input, and sequentially performs global average pooling, fully connected dimension reduction, ReLU activation, fully connected dimension increase and Sigmoid activation processing.

[0019] The complex residual module in step S3 takes the complex feature obtained by the complex convolution layer of the Ka-SAR image as input, and sequentially performs complex convolution layer processing, complex batch normalization, complex ReLU activation, short chain connection, complex weighting and complex ReLU activation to obtain the output.

[0020] The scattering mechanism decomposition module calculates the scattering power based on the complex feature output by the complex residual module, and then decomposes it into surface scattering branch, volume scattering branch and secondary scattering branch after complex batch normalization processing. The feature components are calculated after calculating their weights, and finally the feature components are spliced and output.

[0021] The amplitude conversion module separates the input complex feature into real and imaginary parts to generate real and imaginary tensors, and then performs square root operation to output the amplitude feature.

[0022] The noise reduction and spot removal module inputs the feature after amplitude conversion into a noise estimation network to generate a noise estimation intensity map, and then outputs the noise reduction feature after convolution and weighting processing.

[0023] The fusion mechanism in step S4 triggers the gating mechanism when the cloud coverage of the optical remote sensing image exceeds a certain threshold, and sets the Ka-SAR feature as the dominant feature for fusion, as follows:

[0024] ,

[0025] wherein α is the optical feature weight, β is the compensation intensity, M cloud is the cloud coverage, F opt is the optical image, F SAR is the Ka-SAR image.

[0026] When the cloud coverage of the optical remote sensing image is not over the threshold value, cross-modal attention fusion is used for feature projection, and the formula is as follows:

[0027] ,

[0028] ,

[0029] ,

[0030] wherein Q , K , V respectively represent the projected query, key, and value spaces after feature input, W q , W k , W v are respectively the query projection matrix, the key projection matrix, and the value projection matrix, F opt , F SAR are respectively the optical remote sensing image feature and the SAR image feature, LN () is a normalization layer;

[0031] After feature projection, the attention weight is calculated, and the calculation formula is as follows:

[0032] ,

[0033] ,

[0034] wherein S is the similarity, c k is the scaling factor, Q’ , K’ is the dimensionality-reshaped result of the query feature Q and the key feature K .

[0035] After calculating the attention weight, feature weighting and fusion are performed in combination with the weight, and the core formula is as follows:

[0036] ,

[0037] wherein F fuse is the fused feature, γ is the fusion coefficient, initially set as a default value and then optimized according to different stages of model training, V’ is the value feature to be weighted, and is the dimensionally reshaped result of the value feature V .

[0038] In step S5, the boundary quality of the segmentation map is optimized by using a geometric constraint loss function for the fused feature, and the loss function is constituted as follows:

[0039] ,

[0040] ,

[0041] wherein L CE is the cross-entropy loss, L smooth , L geometry are respectively a smoothing loss function and a geometric loss function, λ 1 is a smoothing loss weight, λ 2 is a geometric loss weight, N is the total number of pixels of the image, i , j is a pixel position index, y i,j is a true label of a pixel i , j , p i,j is a predicted probability of a pixel i , j .

[0042] ,

[0043] ,

[0044] ,

[0045] ,

[0046] wherein L dir , L ar are respectively a direction loss and an aspect ratio loss, N is the total number of pixels of the image,i 、 j is a pixel position index, x , y respectively represent the horizontal and vertical directions of the probability map, P is a predicted probability map, K represents the number of detected straight line segments in the extracted parcel boundary in the probability map, k is a line segment index, θ k is a line segment direction angle, R is a connected region set, r is a parcel index, w r 、 h r respectively represent the width and height of the parcel, ρ is an ideal aspect ratio.

[0047] Another object of the present application is to provide an agricultural field parcel recognition device based on optical-Ka band SAR feature fusion, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the agricultural field parcel recognition method based on optical-Ka band SAR feature fusion as described above.

[0048] A computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the steps of the agricultural field parcel recognition method based on optical-Ka band SAR feature fusion as described above.

[0049] The beneficial effects of the present application are:

[0050] (1) The present application is aimed at the heterogeneous characteristics of optical image features and Ka-SAR image features, establishes a cross-modal physical correlation model, and adopts a cross-modal attention fusion method for feature fusion, realizes physical-level collaborative fusion of data features, solves the problem that the simple fusion scheme cannot play the complementary advantages due to the essential difference between optical reflection characteristics and SAR scattering mechanism;

[0051] (2) The cloud gate mechanism is adopted, when the optical image is seriously blocked by clouds, the complex fusion module is automatically bypassed, and the original SAR feature is directly weighted and output, solving the problem that the agricultural field parcel cannot be completely recognized in the cloudy scene;

[0052] (3) The geometric constraint loss function is adopted to optimize the boundary quality of the segmentation graph, which can better identify the agricultural field parcel, and solve the problem that the general segmentation network ignores the specific regular geometric properties (orthogonal boundary, rectangular parcel) of the agricultural field, resulting in that the output result does not meet the needs of agricultural management, etc.

[0053] (4) Adding a spectral attention module to the ResNet50 architecture during feature extraction of optical remote sensing images, combining NDVI (normalized vegetation index) as an independent channel input to enhance the sensitivity of the model to vegetation;

[0054] (5) Using complex convolution + scattering decomposition for feature extraction of Ka-SAR images, which can adapt to Ka-band SAR satellite images, thereby facilitating better fusion of Ka-SAR image features and optical image features. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a flowchart of the method of the present application;

[0056] Figure 2 is the feature extraction network of the optical remote sensing image of the present application;

[0057] Figure 3 is the feature extraction network of the Ka-SAR image of the present application;

[0058] Figure 4 is a schematic diagram of the Ka-SAR and optical remote sensing images of the research area in the embodiment, (a) is the Ka-SAR image of region 1, (b) is the optical remote sensing image of region 1, (c) is the Ka-SAR image of region 2, and (d) is the optical remote sensing image of region 2;

[0059] Figure 5 is a farmland extraction result diagram of the research area in the embodiment, a is the farmland recognition result of region 1, and b is the farmland recognition result of region 2. DETAILED DESCRIPTION

[0060] The present application will be further described below in conjunction with specific embodiments.

[0061] Embodiment 1: A farmland plot recognition method based on optical-Ka band SAR feature fusion, comprising the following steps:

[0062] S1. Preprocessing and image registration of Ka-band SAR images (i.e. Ka-SAR images) and optical remote sensing images:

[0063] First, perform radiation calibration and multi-view processing on the Ka-band SAR image, and perform radiation calibration, atmospheric correction, and cloud mask generation on the optical image. Subsequently, a sub-pixel registration transformation model is used to perform image registration on the Ka-SAR image and the corresponding optical image data, and ensure that the number of rows and columns of the two images is consistent. The transformation model is as follows:

[0064] ,

[0065] In the formula, x’ , y’To register the target image coordinates, x , y The coordinates of the image to be registered. t x , t y For translation parameters, a , b , c , d These are rotation parameters.

[0066] S2. Feature extraction from optical remote sensing images is performed using an improved ResNet50 network:

[0067] An improved ResNet50 architecture is used to extract features from optical remote sensing images. Based on the original ResNet50 architecture, a spectral attention module is added, and NDVI (Normalized Difference Vegetation Index) is used as an independent channel input to enhance the model's sensitivity to vegetation. The formula for calculating the spectral attention weights is as follows:

[0068] ,

[0069] ,

[0070] In the formula σ For normalization function, δ It is a non-linear activation function. W 1 is the dimension-reduced weight matrix. W 2 is for reconstructing the weight matrix. z For channel statistics vectors, H , W These represent the feature map height and width, respectively. x ( i,j ) represents the eigenvalue at position (i,j).

[0071] NDVI channel expansion involves inputting the Red, Green, Blue, and Nir bands of optical imagery, calculating the NDVI (Normalized Difference Vegetation Index), and then combining the NDVI with the Red, Green, and Blue bands to form a 4-band data set, which is then input into the ResNet50 network backbone to improve the extraction of farmland features from optical imagery. The NDVI calculation formula is as follows:

[0072] .

[0073] The improved ResNet50 backbone processes the input consisting of NDVI, Red, Green and Blue into an initial convolutional layer, a max pooling layer, and a four-stage residual processing (stage 1: preserving resolution, stage 2: downsampling, stage 3: high-level semantics, stage 4: preserving resolution).

[0074] The spectral attention module is to take the improved ResNet50 network backbone output result as the input of the spectral attention module, and sequentially perform global average pooling, full connection dimension reduction, ReLU activation, full connection dimension increase, and Sigmoid activation processing.

[0075] S3. A complex convolutional network is used to extract features from the Ka-SAR image:

[0076] The intensity, phase, and coherence data of the Ka-SAR image are input to construct, and after the corresponding complex features are obtained through the complex convolutional layer processing, the feature extraction is completed through the complex residual module, the scattering mechanism decomposition module, the amplitude conversion module, and the noise reduction and spot removal module;

[0077] The complex residual module takes the complex features obtained after the Ka-SAR image passes through the complex convolutional layer as input, and sequentially performs complex convolutional layer, complex batch normalization, complex ReLU activation, short chain connection (processing method for directly transmitting input features to the output end through identity mapping or linear transformation, and fusing with the output of the main path), complex weighting, and complex ReLU activation to obtain output.

[0078] The scattering mechanism decomposition module calculates the scattering power based on the complex features output by the complex residual module, and then performs complex batch normalization processing, decomposes into surface scattering branch, volume scattering branch, and secondary scattering branch, calculates the weight of each branch, and finally splices and outputs the feature components.

[0079] The amplitude conversion module separates the input complex features into real and imaginary parts to generate real and imaginary tensors, and then performs square root operation to output amplitude features.

[0080] The noise reduction and spot removal module inputs the features after amplitude conversion into a noise estimation network, generates a noise estimation intensity map, and outputs noise reduction features after convolution and weighting processing.

[0081] S4. The fusion mechanism of the extracted features is determined according to the cloud coverage of the optical image. When the cloud coverage of the optical remote sensing image exceeds the threshold, the gating mechanism is triggered, and the Ka-SAR features are used as the main features, and the optical image features are used as the auxiliary features for feature fusion. If the cloud coverage of the optical remote sensing image does not exceed the threshold, the cross-modal attention fusion mechanism is used for feature fusion:

[0082] When the cloud coverage of the optical remote sensing image exceeds a certain threshold, the gating mechanism is triggered, and the Ka-SAR features are set as the main features for feature fusion, and the formula is as follows:

[0083] ,

[0084] In the formula,α is the weight of optical feature, β is the compensation intensity, M cloud is the cloud coverage, F opt is the optical image, F SAR is the Ka-SAR image.

[0085] When the cloud coverage of the optical remote sensing image is not over the threshold, cross-modal attention fusion is used for feature projection, and the formula is as follows:

[0086] ,

[0087] ,

[0088] ,

[0089] In the formula, Q , K , V respectively represent the projected query, key, and value spaces after feature input, W q , W k , W v are the query projection matrix, key projection matrix, and value projection matrix, F opt , F SAR are the optical remote sensing image feature and SAR image feature, LN () is the normalization layer;

[0090] After feature projection, the attention weight is calculated, and the calculation formula is as follows:

[0091] ,

[0092] ,

[0093] In the formula, S is the similarity, c k is the scaling factor, Q’ , K’ is the dimensionality restoration result of the query feature Q and the key feature K .

[0094] After calculating the attention weight, feature weighting and fusion are combined with the weight, and the core formula is as follows:

[0095] ,

[0096] wherein F fuse is the fused feature, γ is the fusion coefficient, initially set as a default value, and then optimized according to different stages of model training, V’ is the value feature to be weighted, and is the value feature V after dimension reshaping.

[0097] S5. The obtained fused feature is outputted to a farmland plot recognition probability map based on a UNet++ decoder, a geometric constraint loss function is used to optimize and iterate the result, and a final recognition model is obtained:

[0098] The loss function is composed as follows:

[0099] ,

[0100] ,

[0101] wherein L CE is the cross-entropy loss, L smooth , L geometry are a smoothing loss function and a geometric loss function, λ 1 is the smoothing loss weight, λ 2 is the geometric loss weight, N is the total number of image pixels, i , j is the pixel position index, y i,j is the true label of the pixel i , j p i,j is the predicted probability of the pixel i , j

[0102] ,

[0103] ,

[0104] ,

[0105] ,

[0106] wherein L dir , L ar are a direction loss and an aspect ratio loss,​​N is the total number of pixels of the image, i , j is the pixel position index, x , y respectively represent the horizontal and vertical directions of the probability map, P is the predicted probability map, K represents the number of straight line segments detected in the extracted field boundary in the probability map, k is the line segment index, θ k is the line segment direction angle, R is the connected region set, r is the field index, w r , h r respectively represent the field width and height, ρ is the ideal aspect ratio.

[0107] Select a certain area Ka-SAR image and its corresponding optical image input method to form a trained model to obtain the farmland field recognition result, select part of the optical and Ka-SAR image as shown in Figure 4 ; the result of extracting farmland field by the method of the embodiment is shown in Figure 5 .

[0108] The recognition accuracy indicators of the method of the application and HBGNet, UNet are calculated and compared, as shown in Table 1, the method used in the application is better than the traditional method in each accuracy indicator, among which the accuracy reduction degree is the least in the case of cloud shielding, which can better cope with the recognition and extraction of farmland field under the condition of cloud.

[0109] Table 1

[0110] .

[0111] Embodiment 2: A farmland field recognition device based on optical-Ka band SAR feature fusion, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the farmland field recognition method based on optical-Ka band SAR feature fusion as described in Embodiment 1 above.

[0112] A computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the steps of the farmland field recognition method based on optical-Ka band SAR feature fusion as described in Embodiment 1 above.

[0113] The above is a further description of the application in conjunction with specific embodiments, and the protection scope of the application is not limited thereto.

Claims

1. A farmland plot recognition method based on optical-Ka band SAR feature fusion, characterized in that, The steps comprise the following: S1. Preprocess and image register the Ka-SAR image and the optical remote sensing image; S2. Extract features from the optical remote sensing image using an improved ResNet50 network: calculate the normalized vegetation index of the optical remote sensing image and perform NDVI channel expansion, then input the improved ResNet50 network backbone for feature extraction, and then input the spectral attention module for spectral attention weight calculation, and output the extracted features; S3. Extract features from the Ka-SAR image using a complex convolution network: input the intensity, phase, and coherence data of the Ka-SAR image to construct, process through a complex convolution layer to obtain corresponding complex features, and then complete feature extraction through a complex residual module, a scattering mechanism decomposition module, an amplitude conversion module, and a noise removal and spot removal module; S4. Determine the fusion mechanism of the extracted features according to the cloud coverage of the optical image. When the cloud coverage of the optical remote sensing image exceeds the threshold, trigger the gating mechanism, use the Ka-SAR features as the main feature, and use the optical image features as the auxiliary feature for feature fusion. If the cloud coverage of the optical remote sensing image does not exceed the threshold, use the cross-modal attention fusion mechanism for feature fusion; S5. Output the farmland plot recognition probability map based on the UNet++ decoder of the obtained fused features, and use a geometric constraint loss function to optimize and iterate the results to obtain the final recognition model.

2. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, The image registration in step S1 uses a sub-pixel registration transformation model to register the Ka-SAR image and the corresponding optical image data, and ensures that the number of rows and columns of the two images is consistent. The transformation model is as follows: ， wherein x’ , y’ is a registration target image coordinate, x , y is a to-be-registered image coordinate, t x , t y is a translation parameter, a , b , c , d is a rotation parameter.

3. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, The NDVI channel expansion in step S2 inputs the Red, Green, Blue, and Nir bands of the optical image, then calculates the normalized vegetation index NDVI, and combines NDVI with the Red, Green, and Blue bands into 4-band data to input the ResNet50 network backbone, in order to improve the farmland feature extraction effect of the optical image. The NDVI calculation formula is as follows: 。 4. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, The improved ResNet50 network backbone processes the input composed of NDVI, Red, Green, and Blue through an initial convolution layer, a maximum pooling layer, and a four-stage residual processing in turn.

5. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, The spectral attention module inputs the output of the improved ResNet50 network backbone, and performs global average pooling, full connection dimension reduction, ReLU activation, full connection dimension increase, and Sigmoid activation processing in turn.

6. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, The complex residual module in step S3 inputs the complex features obtained by the Ka-SAR image through the complex convolution layer, and then performs complex convolution, complex batch normalization, complex ReLU activation, short chain connection, complex weighting, and complex ReLU activation in turn to obtain the output. The scattering mechanism decomposition module calculates the scattering power based on the complex feature output by the complex residual module, and then decomposes the scattering power into surface scattering branch, volume scattering branch and secondary scattering branch after complex batch normalization processing, calculates the feature components after calculating the weights of the branches, and finally splices and outputs the feature components. The amplitude conversion module separates the input complex feature into real part and imaginary part, generates real part tensor and imaginary part tensor, and then performs square root operation to output amplitude feature. The noise reduction and speckle removal module inputs the feature after amplitude conversion into a noise estimation network, generates a noise estimation intensity map, and then outputs a noise reduction feature after convolution and weighting processing.

7. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, The fusion mechanism in step S4 triggers the gating mechanism when the cloud coverage of the optical remote sensing image exceeds a certain threshold, and sets the Ka-SAR feature as the dominant for feature fusion, with the formula as follows: ， wherein α is an optical feature weight, β is a compensation intensity, M cloud is a cloud coverage, F opt is an optical image, F SAR is a Ka-SAR image; When the cloud coverage of the optical remote sensing image does not exceed the threshold, cross-modal attention fusion is first used for feature projection, with the formula as follows: ， ， ， In the formula Q , K , V respectively represent the query, key, value space projected after the feature input, W q , W k , W v respectively are the query projection matrix, key projection matrix, value projection matrix, F opt , F SAR respectively are the optical remote sensing image features and SAR image features, LN () is a normalization layer; After feature projection, the attention weight is calculated, with the formula as follows: ， ， wherein S is a similarity, c k is a scaling factor, Q’ , K’ is a query feature Q and a key feature K after dimension reshaping. After calculating the attention weight, the feature is weighted and fused combined with the weight, with the core formula as follows: ， In the formula F fuse is the fused feature, γ is the fusion coefficient, initially set as a default value, and then continuously optimized according to different stages of model training, V’ is the value feature to be weighted, and is the value feature V after dimension reshaping.

8. The farmland plot recognition method based on optical-Ka band SAR feature fusion according to claim 1, characterized in that, In step S5, the geometric constraint loss function is used to optimize the boundary quality of the segmentation map for the fused feature, with the loss function composition as follows: ， ， In the formula L CE is the cross-entropy loss, L smooth , L geometry are a smoothing loss function and a geometric loss function, respectively, λ 1 is a smoothing loss weight, λ 2 is a geometric loss weight, N is the total number of image pixels, i , j is a pixel position index, y i,j is a true label of a pixel i , j , p i,j is a predicted probability of a pixel i , j . ， ， ， ， where L dir , L ar are the orientation loss and the aspect ratio loss, respectively, N is the total number of pixels in the image, i , j is the pixel position index, x , y are the horizontal and vertical directions of the probability map, respectively, P is the predicted probability map, K is the number of detected line segments in the extracted parcel boundary in the probability map, k is the line segment index, θ k is the line segment orientation angle, R is the connected region set, r is the parcel index, w r , h r are the parcel width and height, respectively, ρ is the ideal aspect ratio.

9. An agricultural field parcel recognition device based on optical-Ka band SAR feature fusion, comprising a memory, a processor and a computer program stored on the memory and capable of running on the processor, characterized in that, The processor executes the program to realize the farmland plot recognition method based on optical-Ka band SAR feature fusion according to any one of claims 1-8.

10. A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, realizes the steps of the farmland plot recognition method based on optical-Ka band SAR feature fusion according to any one of claims 1-8.

Images