Farmland parcel identification method based on optical-Ka frequency band SAR (Synthetic Aperture Radar) feature fusion

Abstract

The invention relates to a farmland parcel identification method based on optical-Ka frequency band SAR feature fusion, and the method comprises the steps: carrying out the feature extraction of an optical remote sensing image through employing an improved ResNet50 network, carrying out the feature extraction of a Ka-SAR image through employing a complex convolution network, triggering a gating mechanism when the cloud coverage rate of the optical remote sensing image exceeds a threshold value, and carrying out the recognition of a farmland parcel. If the cloud coverage rate of the optical remote sensing image does not exceed the threshold value, carrying out feature fusion by adopting a cross-modal attention fusion mechanism; and outputting a farmland plot identification probability graph based on the obtained fusion features based on a UNet + + decoder, and performing optimization and iteration on a result by adopting a geometric constraint loss function to obtain a final identification model. The cross-modal attention fusion method is adopted to perform feature fusion, physical-level collaborative fusion for data features is realized, and meanwhile, the problem that an output result does not meet agricultural management requirements due to the fact that farmland parcels cannot be completely recognized and specific regular geometric attributes of farmland are ignored in a multi-cloud scene is solved.

Description

Technical Field

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

[0002] With the development of precision agriculture and digital agriculture, high-precision identification of farmland plots has become a core requirement for agricultural informatization. Remote sensing technology, with its advantages of large-scale and periodic observation, has become a primary means of farmland monitoring. Currently, mainstream technologies include optical remote sensing and SAR remote sensing, but both have significant limitations in farmland plot identification. Optical remote sensing uses visible-to-near-infrared bands (such as Landsat-8 OLI and Sentinel-2 MSI) to obtain vegetation indices (NDVI, EVI, etc.) and implements plot identification through threshold segmentation or machine learning. However, this method is highly dependent on weather conditions and can degrade extraction results when cloud cover is high. SAR remote sensing uses the backscatter coefficient and coherence characteristics of synthetic aperture radar (such as Sentinel-1 C-band and TerraSAR-X X-band) to identify surface types through polarization decomposition. However, this method can introduce speckle noise, resulting in uneven texture within plots. Side-view imaging can also lead to overlapping and shadowing issues, which can affect identification and extraction accuracy.

[0003] Research on collaborative crop identification using multi-source remote sensing data exists in existing technologies. For example, patent CN110189616A discloses a crop mapping method using multi-temporal SAR (GaoFen-3) and optical imagery (GaoFen-2) as data sources. The method leverages the "map" information of farmland plot structure provided by GaoFen-2 imagery and the polarization scattering and texture information of ground features provided by GaoFen-3 imagery to achieve accurate crop type identification and planted area extraction. While this technology emphasizes the collaborative processing of SAR and optical imagery, due to the novelty of Ka-SAR and its high resolution and rich texture features, it does not specifically address these characteristics. It also does not address issues such as the high cloud coverage of optical imagery and the segmentation network's neglect of regular geometric properties unique to farmland (orthogonal boundaries and rectangular plots), resulting in output that does not meet agricultural management requirements. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a farmland plot identification method based on optical-Ka-band SAR feature fusion. The method adapts Ka-band SAR satellite imagery and can better fuse Ka-SAR image features with optical image features to give play to their complementary advantages.

[0005] The technical solution adopted by the present invention is: The farmland plot recognition method based on optical-Ka band SAR feature fusion includes the following steps: S1. Preprocess and register Ka-band SAR images (i.e., Ka-SAR images) and optical remote sensing images; S2. Feature extraction from optical remote sensing images using an improved ResNet50 network: The normalized difference vegetation index (NDVI) of the optical remote sensing images is calculated and the NDVI channel is expanded. This data is then fed into the improved ResNet50 network backbone for feature extraction. This data is then fed into the spectral attention module for spectral attention weight calculation, and the extracted features are output. S3. Feature extraction from Ka-SAR images using a complex convolutional network: The intensity, phase, and coherence data of the Ka-SAR image are input and constructed. After processing through a complex convolutional layer to obtain the corresponding complex features, feature extraction is completed through a complex residual module, a scattering mechanism decomposition module, an amplitude conversion module, and a noise reduction and speckle removal module. S4. Determine the fusion mechanism for extracted features based on the cloud coverage of the optical image. When the cloud coverage of the optical remote sensing image exceeds a threshold, a gating mechanism is triggered, and feature fusion is performed using Ka-SAR features as the primary and optical image features as the auxiliary. If the cloud coverage of the optical remote sensing image does not exceed the threshold, feature fusion is performed using a cross-modal attention fusion mechanism. S5. The obtained fusion features are used to output the farmland parcel recognition probability map based on the UNet++ decoder. The results are optimized and iterated using the geometric constraint loss function to obtain the final recognition model.

[0006] In the above method, the image registration described in step S1 uses a sub-pixel registration transformation model 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: , In the formula x’ 、 y’ To register the target image coordinates, x 、 y is the coordinate of the image to be registered, t x 、 t y is the translation parameter, a 、 b 、 c 、 d is the rotation parameter.

[0007] The NDVI channel expansion described in step S2 inputs the Red, Green, Blue, and Nir bands of the optical image, then calculates the Normalized Difference Vegetation Index (NDVI). NDVI is then combined with the Red, Green, and Blue bands to form four-band data that is input into the ResNet50 network backbone to improve the optical image farmland feature extraction effect. The NDVI calculation formula is as follows: .

[0008] The improved ResNet50 network backbone processes the input consisting of NDVI, Red, Green, and Blue in sequence through an initial convolutional layer, a maximum pooling layer, and a four-stage residual processing (stage 1: maintaining resolution, stage 2: downsampling, stage 3: high-level semantics, stage 4: maintaining resolution).

[0009] The spectral attention module uses the output result of the improved ResNet50 network backbone as the input of the spectral attention module, and performs global average pooling, full connection dimensionality reduction, ReLU activation, full connection dimensionality increase, and Sigmoid activation processing in sequence.

[0010] The complex residual module described in step S3 takes the complex features obtained after the Ka-SAR image passes through the complex convolution layer as input, and obtains the output after successively passing through the complex convolution layer, complex batch normalization, complex ReLU activation, short chain connection, complex weighting, and complex ReLU activation.

[0011] The scattering mechanism decomposition module calculates the scattering power based on the complex features 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. After calculating their weights respectively, the characteristic components are calculated, and finally the characteristic components are spliced ​​and output.

[0012] The amplitude conversion module separates the real and imaginary parts of the input complex number features, generates a real tensor and an imaginary tensor, and then performs a square root operation to output the amplitude feature.

[0013] The noise reduction and freckle removal module inputs the amplitude-converted features into the noise estimation network, generates a noise estimation intensity map, and then outputs the noise reduction features through convolution and weighted processing.

[0014] The fusion mechanism described in step S4 triggers a 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 feature fusion. The formula is as follows: , In the formula α is the optical feature weight, β To compensate for the strength, Mcloud is the cloud coverage, F opt For optical imaging, F SAR This is a Ka-SAR image.

[0015] 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, and the formula is as follows: , , , In the formula Q 、 K 、 V Respectively represent the query, key, and value spaces projected after the feature input, W q 、 W k 、 W v They are query projection matrix, key projection matrix, and value projection matrix respectively. F opt 、 F SAR are optical remote sensing image features and SAR image features, LN () is the normalization layer; After feature projection, the attention weight is calculated using the following formula: , , In the formula S is the similarity, c k is the scaling factor, Q’ 、 K’ Query features Q Key Features K The result after dimension reshaping; After calculating the attention weights, the weights are combined for feature weighting and fusion. The core formula is as follows: , In the formula F fuse is the fused feature, c is the fusion coefficient, which is initially the default value and is subsequently optimized according to different stages of model training. V’ is the value feature to be weighted, V The result after dimension reshaping.

[0016] In step S5, the geometric constraint loss function is used to optimize the boundary quality of the segmentation map based on the fused features. The loss function is constructed as follows: , , In the formula L CE is the cross entropy loss, L smooth 、 L geometry They are smoothing loss function and geometric loss function respectively. l 1 is the smoothed loss weight, l 2 is the geometric loss weight, N is the total number of image pixels, i 、 j is the pixel position index, y i,j Pixels i 、 j The true label, p i,j Pixels i 、 j The predicted probability of , , , , In the formula L dir 、 L ar They are direction loss and aspect ratio loss, N is the total number of image pixels, i 、 j is the pixel position index, x , y Represent the horizontal and vertical directions of the probability graph, respectively. P is the predicted probability map, K represents the number of straight line segments detected in the plot boundaries extracted from the probability map, k is the line segment index, i k is the direction angle of the line segment, R is a set of connected regions, r is the plot index, w r 、 h r Represents the width and height of the plot respectively, r Ideal aspect ratio.

[0017] Another object of the present invention is to provide a farmland plot identification device based on optical-Ka band SAR feature fusion, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, the farmland plot identification method based on optical-Ka band SAR feature fusion as described above is implemented.

[0018] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned farmland plot identification method based on optical-Ka band SAR feature fusion.

[0019] The beneficial effects of the present invention are: (1) This invention addresses the heterogeneity between optical image features and Ka-SAR image features, establishes a cross-modal physical association model, and adopts a cross-modal attention fusion method for feature fusion, achieving physical-level collaborative fusion of data features. This solves the problem that simple fusion schemes are unable to fully exploit the complementary advantages caused by the essential differences between optical reflection characteristics and SAR scattering mechanisms. (2) The cloud gating mechanism is used. When the optical image is severely obscured by clouds, the complex fusion module is automatically bypassed and the original SAR feature weighted output is directly used, which solves the problem of incomplete identification of farmland plots in cloudy scenes. (3) The geometric constraint loss function is used to optimize the boundary quality of the segmentation map, which can better identify farmland plots. This solves the problem that the general segmentation network ignores the regular geometric properties unique to farmland (orthogonal boundaries, rectangular plots), resulting in output results that do not meet agricultural management requirements. (4) When extracting features from optical remote sensing images, a spectral attention module is added to the ResNet50 architecture, and NDVI (normalized difference vegetation index) is used as an independent channel input to enhance the model's sensitivity to vegetation. (5) Complex convolution + scattering decomposition is used to extract Ka-SAR image features, which can be adapted to Ka-band SAR satellite images, thereby facilitating a better fusion of Ka-SAR image features with optical image features. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention; Figure 2 A feature extraction network for optical remote sensing images of the present invention; Figure 3 The feature extraction network of Ka-SAR image of the present invention; Figure 4Schematic diagram of Ka-SAR and optical remote sensing images of the study area of ​​the embodiment; (a) is the Ka-SAR image of area 1, (b) is the optical remote sensing image of area 1, (c) is the Ka-SAR image of area 2, and (d) is the optical remote sensing image of area 2; Figure 5 This is the farmland extraction result diagram of the study area of ​​the embodiment, a is the farmland identification result of area 1, and b is the farmland identification result of area 2. DETAILED DESCRIPTION

[0021] The present invention is further described below with reference to specific embodiments.

[0022] Example 1 A method for identifying farmland plots based on optical-Ka band SAR feature fusion includes the following steps: S1. Preprocess and register Ka-band SAR images (i.e., Ka-SAR images) and optical remote sensing images: First, the Ka-band SAR image is subjected to radiometric calibration and multi-look processing, while the optical image is also preprocessed with radiometric calibration, atmospheric correction, and cloud mask generation. The Ka-SAR image is then registered with the corresponding optical image using a sub-pixel registration transformation model, ensuring that the number of rows and columns in the two images is consistent. The transformation model is as follows: , In the formula x’ 、 y’ To register the target image coordinates, x 、 y is the coordinate of the image to be registered, t x 、 t y is the translation parameter, a 、 b 、 c 、 d is the rotation parameter.

[0023] S2. Use the improved ResNet50 network to extract features from optical remote sensing images: 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 combined as an independent channel input to enhance the model's sensitivity to vegetation. The spectral attention weight calculation formula is as follows: , , In the formula s is the normalization function, d is a nonlinear activation function,W 1 is the dimensionality reduction weight matrix, W 2 is the reconstruction weight matrix, z is the channel statistics vector, H 、 W are the feature map height and width respectively, x ( i,j ) is the eigenvalue at position (i, j).

[0024] NDVI channel expansion is to input the Red, Green, Blue, and Nir bands of the optical image, then calculate the Normalized Difference Vegetation Index (NDVI). NDVI is then combined with the Red, Green, and Blue bands to form four-band data and input into the ResNet50 network backbone to improve the optical image farmland feature extraction effect. The NDVI calculation formula is as follows: .

[0025] The improved ResNet50 network backbone processes the input consisting of NDVI, Red, Green, and Blue in sequence through the initial convolution layer, the maximum pooling layer, and the four-stage residual processing (stage 1: maintaining resolution, stage 2: downsampling, stage 3: high-level semantics, stage 4: maintaining resolution).

[0026] The spectral attention module takes the output of the improved ResNet50 network backbone as the input of the spectral attention module, and performs global average pooling, full connection dimensionality reduction, ReLU activation, full connection dimensionality increase, and Sigmoid activation processing in sequence.

[0027] S3. Feature extraction of Ka-SAR images using complex convolutional networks: The intensity, phase, and coherence data of the Ka-SAR image are input and constructed. After being processed by the complex convolution layer to obtain the corresponding complex features, the feature extraction is completed by the complex residual module, scattering mechanism decomposition module, amplitude conversion module, and noise reduction and speckle removal module. The complex residual module takes the complex features obtained by passing the Ka-SAR image through the complex convolution layer as input, and sequentially undergoes the complex convolution layer, complex batch normalization, complex ReLU activation, short chain connection (a processing method that directly transfers the input features to the output end through identity mapping or linear transformation and fuses them with the main path output), complex weighting, and complex ReLU activation to obtain the output.

[0028] The scattering mechanism decomposition module calculates the scattering power based on the complex features 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. After calculating their weights respectively, the characteristic components are calculated, and finally the characteristic components are spliced ​​and output.

[0029] The amplitude conversion module separates the real and imaginary parts of the input complex number features, generates a real tensor and an imaginary tensor, and then performs a square root operation to output the amplitude feature.

[0030] The noise reduction and freckle removal module inputs the amplitude-converted features into the noise estimation network, generates a noise estimation intensity map, and then outputs the noise reduction features through convolution and weighted processing.

[0031] S4. The fusion mechanism for extracting features is determined based on the cloud coverage of the optical image. When the cloud coverage of the optical remote sensing image exceeds a threshold, the gating mechanism is triggered, and Ka-SAR features are used as the main feature, and optical image features are used as 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: When the cloud coverage of the optical remote sensing image exceeds a certain threshold, the gating mechanism is triggered and the Ka-SAR feature is set as the dominant feature fusion. The formula is as follows: , In the formula α is the optical feature weight, β To compensate for the strength, M cloud is the cloud coverage, F opt For optical imaging, F SAR This is a Ka-SAR image.

[0032] 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, and the formula is as follows: , , , In the formula Q 、 K 、 V Respectively represent the query, key, and value spaces projected after the feature input, W q 、 W k 、 W v They are query projection matrix, key projection matrix, and value projection matrix respectively. F opt 、 F SAR are optical remote sensing image features and SAR image features, LN () is the normalization layer; After feature projection, the attention weight is calculated using the following formula: , , In the formula S is the similarity, c k is the scaling factor, Q’ 、 K’ Query features Q Key Features K The result after dimension reshaping; After calculating the attention weights, the weights are combined for feature weighting and fusion. The core formula is as follows: , In the formula F fuse is the fused feature, c is the fusion coefficient, which is initially the default value and is subsequently optimized according to different stages of model training. V’ is the value feature to be weighted, V The result after dimension reshaping.

[0033] S5. The fused features are used to generate a farmland parcel recognition probability map based on the UNet++ decoder. The geometric constraint loss function is used to optimize and iterate the results to obtain the final recognition model: The loss function is constructed as follows: , , In the formula L CE is the cross entropy loss, L smooth 、 L geometry They are smoothing loss function and geometric loss function respectively. l 1 is the smoothed loss weight, l 2 is the geometric loss weight, N is the total number of image pixels, i 、 j is the pixel position index, y i,j Pixels i 、 j The true label, p i,j Pixels i 、 j The predicted probability of , , , , In the formula L dir 、 L ar They are direction loss and aspect ratio loss respectively. N is the total number of image pixels, i 、 j is the pixel position index, x , y Represent the horizontal and vertical directions of the probability graph, respectively. P is the predicted probability map, K represents the number of straight line segments detected in the plot boundaries extracted from the probability map, k is the line segment index, i k is the direction angle of the line segment, R is a set of connected regions, r is the plot index, w r 、 h r Represents the width and height of the plot respectively, r Ideal aspect ratio.

[0034] The trained model formed by the Ka-SAR image of a certain area and its corresponding optical image input method is used to obtain the farmland identification results. The optical and Ka-SAR images of some areas are selected to show the following. Figure 4 The results of farmland plot extraction using the method of this embodiment are shown in Figure 5 shown.

[0035] The recognition accuracy indicators of the method of the present invention are calculated and compared with those of HBGNet and UNet. As shown in Table 1, the method used in the present invention outperforms the traditional methods in all accuracy indicators. In the case of cloud cover, the degree of accuracy reduction is the least, and it can better cope with the recognition and extraction of farmland plots in cloudy conditions.

[0036] Table 1 .

[0037] Example 2 A farmland plot identification device based on optical-Ka band SAR feature fusion includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the farmland plot identification method based on optical-Ka band SAR feature fusion as described in Example 1 above is implemented.

[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the farmland plot identification method based on optical-Ka band SAR feature fusion as described in Example 1 above.

[0039] The above is a further description of the present invention in conjunction with specific embodiments, and the protection scope of the present invention is not limited thereto.

Claims

1. A farmland plot recognition method based on optical-Ka band SAR feature fusion is characterized by: The steps are as follows: S1. Preprocess and register Ka-SAR and optical remote sensing images. S2. Feature extraction from optical remote sensing images using an improved ResNet50 network: The optical remote sensing images are used to calculate the Normalized Difference Vegetation Index (NDVI) and perform NDVI channel expansion. These images are then fed into the improved ResNet50 backbone for feature extraction. The spectral attention module then calculates spectral attention weights and outputs the extracted features. S3. Feature extraction from Ka-SAR images using a complex convolutional network: The intensity, phase, and coherence data of the Ka-SAR image are input and constructed. After processing through a complex convolutional layer to obtain the corresponding complex features, feature extraction is completed through a complex residual module, a scattering mechanism decomposition module, an amplitude conversion module, and a noise reduction and speckle removal module. S4. Determine the fusion mechanism for extracted features based on the cloud coverage of the optical image. When the cloud coverage of the optical remote sensing image exceeds a threshold, a gating mechanism is triggered, and feature fusion is performed using Ka-SAR features as the primary and optical image features as the auxiliary. If the cloud coverage of the optical remote sensing image does not exceed the threshold, feature fusion is performed using a cross-modal attention fusion mechanism. S5. The obtained fusion features are used to output the farmland parcel recognition probability map based on the UNet++ decoder. The results are optimized and iterated using the geometric constraint loss function to obtain the final recognition model.

2. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: The image registration described in step S1 uses a sub-pixel registration transformation model 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: ， In the formula x’ 、 y’ To register the target image coordinates, x 、 y is the coordinate of the image to be registered, t x 、 t y is the translation parameter, a 、 b 、 c 、 d is the rotation parameter.

3. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: The NDVI channel expansion described in step S2 inputs the Red, Green, Blue, and Nir bands of the optical image, then calculates the Normalized Difference Vegetation Index (NDVI). NDVI is synthesized with the Red, Green, and Blue bands to form four-band data and input into the ResNet50 network backbone to improve the optical image farmland feature extraction effect. The NDVI calculation formula is as follows: 。 4. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: The improved ResNet50 network backbone processes the input consisting of NDVI, Red, Green, and Blue in sequence through an initial convolutional layer, a maximum pooling layer, and a four-stage residual process.

5. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: The spectral attention module uses the output result of the improved ResNet50 network backbone as the input of the spectral attention module, and performs global average pooling, full connection dimensionality reduction, ReLU activation, full connection dimensionality increase, and Sigmoid activation processing in sequence.

6. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: The complex residual module described in step S3 takes the complex features obtained by passing the Ka-SAR image through the complex convolution layer as input, and sequentially undergoes the complex convolution layer, complex batch normalization, complex ReLU activation, short chain connection, complex weighting, and complex ReLU activation to obtain the output; The scattering mechanism decomposition module calculates the scattering power based on the complex features 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. After calculating the weights of each branch, the characteristic components are calculated, and finally the characteristic components are spliced ​​and output. The amplitude conversion module separates the real and imaginary parts of the input complex number features to generate real tensors and imaginary tensors, and then performs square root operation to output amplitude features; The noise reduction and freckle removal module inputs the amplitude-converted features into the noise estimation network, generates a noise estimation intensity map, and then outputs the noise reduction features through convolution and weighted processing.

7. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: The fusion mechanism described in step S4 triggers a 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 feature fusion. The formula is as follows: ， In the formula α is the optical feature weight, β To compensate for the strength, M cloud is the cloud coverage, F opt For optical imaging, F SAR It is 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, and the formula is as follows: ， ， ， In the formula Q 、 K 、 V Respectively represent the query, key, and value spaces projected after the feature input, W q 、 W k 、 W v They are query projection matrix, key projection matrix, and value projection matrix respectively. F opt 、 F SAR are optical remote sensing image features and SAR image features, LN () is the normalization layer; After feature projection, the attention weight is calculated using the following formula: ， ， In the formula S is the similarity, c k is the scaling factor, Q’ 、 K’ Query features Q Key Features K The result after dimension reshaping; After calculating the attention weights, the weights are combined for feature weighting and fusion. The core formula is as follows: ， In the formula F fuse is the fused feature, γ is the fusion coefficient, which is initially the default value and is subsequently optimized according to different stages of model training. V’ is the value feature to be weighted, V The result after dimension reshaping.

8. The farmland plot identification method based on optical-Ka band SAR feature fusion according to claim 1 is characterized in that: In step S5, the geometric constraint loss function is used to optimize the boundary quality of the segmentation map based on the fused features. The loss function is constructed as follows: ， ， In the formula L CE is the cross entropy loss, L smooth 、 L geometry They are smoothing loss function and geometric loss function respectively. λ 1 is the smoothed 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 Pixels i 、 j The true label, p i,j Pixels i 、 j The predicted probability of ， ， ， ， In the formula L dir 、 L ar They are direction loss and aspect ratio loss, N is the total number of image pixels, i 、 j is the pixel position index, x , y Represent the horizontal and vertical directions of the probability graph, respectively. P is the predicted probability map, K represents the number of straight line segments detected in the plot boundaries extracted from the probability map, k is the line segment index, θ k is the direction angle of the line segment, R is a set of connected regions, r is the plot index, w r 、 h r Represents the width and height of the plot respectively, ρ Ideal aspect ratio.

9. A farmland plot identification device based on optical-Ka-band SAR feature fusion, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the farmland plot identification method based on optical-Ka band SAR feature fusion as described in any one of claims 1 to 8 is implemented.

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

Images