A Method for Extracting Soybean Planting Areas from Remote Sensing Images Based on the SK-UNet Deep Learning Network

Abstract

The present invention relates to a method for extracting soybean planting areas from remote sensing images based on the SK-UNet deep learning network, including: obtaining Sentinel-2 remote sensing images with a resolution of 10m and performing preprocessing to form a remote sensing image dataset; constructing an improved UNet network model, namely the SK-UNet model, as the network for extracting soybean planting area images; obtaining the trained SK-UNet model; obtaining the remote sensing images to be extracted and performing preprocessing; inputting the preprocessed remote sensing images to be extracted into the trained SK-UNet model to obtain the extraction results of soybean planting area images. Based on the improvement of the UNet network, the present invention adds an SKNet Block module after each double-layer convolution. It has good scalability and high extraction accuracy, and the adjustable parameters can be applied to remote sensing satellite images of different sensors; through verification, the classification results of the high-resolution remote sensing images obtained by the present invention are superior to the comparative classification algorithms in terms of evaluation indicators, and the generated classification results can better maintain the smoothness and integrity of the edges.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting soybean planting areas from remote sensing images based on a SK-UNet deep learning network. Background Art

[0002] The planting structure of crops is not only an important basis for growth conditions and yield estimation, but also provides the main reference for the formulation of policy guidelines, the layout of crop structures, and the unified management of farmland resources. Quickly and accurately obtaining the distribution of crops and the agricultural production structure within a large area has great guiding significance for macro-adjusting the industrial structure and obtaining changes in agricultural product market prices.

[0003] The traditional methods for obtaining the soybean planting structure basically involve grass-roots production units reporting layer by layer to higher-level authorities, or using statistical means such as "small-area surveys and large-area estimations". The disadvantages are that they are not only time-consuming and laborious but also costly and inefficient. The "graphical" data collection is often impossible to achieve, and the conditions for information-based agricultural production cannot be met. The emergence of remote sensing images makes up for these disadvantages. They contain a large amount of geometric, spatial, and texture information, so it is possible to better understand the attributes of crops in the images, which helps to determine the specific positions of crops and improve the recognition accuracy. They have the advantages of high efficiency, convenience, low cost, and fast speed, and are suitable for large-scale remote sensing detection.

[0004] Shen Kejian et al. used the 16m satellite images of GF-1 as the data source to provide a feasible solution for the development of domestic satellites. Tian Fuyou et al. used Sentinel-2 satellite images as the data source to extract soybean planting areas and conducted comparative experiments with multi-layer neural networks, support vector machines, and random forest algorithms. The results showed that the classification effect of the multi-layer neural network was the best, and the introduction of the near-infrared band could effectively improve the recognition accuracy, which played the greatest role among all bands and effectively separated soybeans from corn. Song Qian et al. used GF-1 satellite images as the data source to extract soybean planting areas. After comparing Landsat satellite data, it was found that the extraction accuracy of GF-1 satellite data was significantly better than that of Landsat satellite data, proving its advantages over medium-resolution remote sensing images and providing a feasible solution for the selection of complex areas and data sources in the future. However, the main disadvantages in the above studies are as follows: First, the feature extraction ability is limited, and features need to be designed and selected according to specific tasks; second, the robustness to interference factors such as noise, light changes, and occlusion is poor, and misjudgment and missed judgment are likely to occur, mainly because the feature extraction algorithm has weak resistance to interference factors; third, a large amount of manual participation is required, including feature design, parameter adjustment, and model training, etc., and the efficiency is low. Summary of the Invention

[0005] To solve the defect that the remote sensing image extraction technology in the existing soybean planting area is difficult to meet the actual use needs, the purpose of the present invention is to provide a method for extracting soybean planting areas from remote sensing images more accurately, achieving better results in terms of efficiency, speed, applicability and accuracy, which is a method for extracting soybean planting areas from remote sensing images based on the SK-UNet deep learning network.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions: A method for extracting soybean planting areas from remote sensing images based on the SK-UNet deep learning network, the method comprises the following steps in sequence:

[0007] (1) Obtain Sentinel-2 remote sensing images with a resolution of 10m and perform preprocessing, use ArcGIS software to draw the vector of the soybean area, obtain the original image data and label data of the soybean area, the size of the original image data of the soybean area is 256×256×3, the size of the label data is 256×256×1, and the original image data and label data of the soybean area constitute a remote sensing image dataset;

[0008] (2) Take the UNet network as the basic network, embed the SKNetBlock module in the basic convolutional layers of its encoder and decoder, and construct an improved UNet network model, namely the SK-UNet model, as the soybean planting area image extraction network;

[0009] (3) Input the remote sensing image dataset into the SK-UNet model for training to obtain the trained SK-UNet model;

[0010] (4) Obtain the remote sensing image to be extracted and perform preprocessing;

[0011] (5) Input the preprocessed remote sensing image to be extracted into the trained SK-UNet model to obtain the extraction result of the soybean planting area image.

[0012] The specific steps of step (2) include the following steps:

[0013] (2a) Extract features based on the UNet network as the basic framework, and set that the Unet network includes an encoder, a decoder and a skip connection part;

[0014] (2b) The encoder is set to consist of a 3×3 convolutional layer, a ReLU function, and a 2×2 max pooling layer. The encoder is used to extract image feature information through the convolutional layer and perform four downsamplings. After each pooling operation, the feature image becomes smaller and the number of channels doubles. The decoder performs upsampling through a 2×2 transposed convolution, and the decoder part completes four upsamplings. After each upsampling, the size of the feature image increases and the number of channels is reduced by half. Then, through the skip connection part, the low-level details and high-level semantics of the feature map are combined. Finally, through Softmax as the classification layer, the picture pixels are segmented into two categories, one is the background and the other is the extracted soybean area.

[0015] (2c) Set the SKNet Block module:

[0016] (2d) Embed the SKNet Block module in the basic convolutional layers of the encoder and decoder to obtain the SK-UNet model. The SKNet Block module is used to integrate feature maps of different scales, dynamically adjust the convolution kernel size, and automatically obtain the importance of each feature channel through learning with the UNet network.

[0017] The specific steps of step (3) are as follows:

[0018] (3a) Determine the hyperparameters during the training process and initialize the parameters of the SK-UNet model. The Batch size is 4, the learning rate is 0.0001, the parameters of the Adam optimizer are Beta1 = 0.5, Beta2 = 0.999, and Epochs = 100.

[0019] (3b) Input the remote sensing image dataset into the SK-UNet model.

[0020] (3c) Use the SK-UNet model to perform forward propagation on the current training data.

[0021] (3d) Calculate the loss and backpropagate it to the SK-UNet model.

[0022] (3e) Use the Adam optimizer to update the parameters of the SK-UNet model according to the loss value of the SK-UNet model. Repeat steps (3b) to (3d) until the loss value is less than a predetermined threshold.

[0023] The specific content of step (2c) is as follows: The SKNet Block module adopts a non-linear multi-core information aggregation method to achieve the adaptive receptive field size of neurons. It consists of three operators: Split, Fuse, and Select operators. Among them, the Split operator generates multiple paths with different kernel sizes, corresponding to different neuron receptive field sizes. The Fuse operator combines and aggregates the information from multiple paths to obtain a global comprehensive representation of the selection weights. The Select operator aggregates the feature maps of different kernel sizes according to the selection weights;

[0024] The Split operator passes the original feature map through a 3×3 grouped / depth convolution and a 3×3 dilated convolution respectively to generate two feature maps U1 and U2, and then adds these two feature maps to generate the feature map U;

[0025] The Fuse operator uses global average pooling to embed global information, thereby generating the statistical information of the channels. The c-th element of the feature S is calculated by shrinking U in the spatial dimension H×W:

[0026]

[0027] In the formula, F gp represents average pooling of the feature, U c is the c-th dimension feature of the input feature map, W and H represent the width and height of the feature, and i and j are the coordinates of the feature map in the spatial dimension;

[0028] A compact feature z is established to achieve precise and adaptive selection guidance, which is realized through a simple fully connected layer. The formula is as follows:

[0029]

[0030] In the formula, δ is the ReLU function, B represents batch normalization, W ∈ R d×C is the weight matrix. To study the influence of d on the model efficiency, the reduction rate r is used to control its value:

[0031] d = max(C / r, L)

[0032] In the formula, L represents the minimum value of d, C is the C-th dimension feature of the input, and r is a hyperparameter;

[0033] The Select operator passes the generated feature z through a c and b c two functions, and multiplies the generated function values with the feature maps U1 and U2:

[0034]

[0035] In the formula, A, B ∈ RC×d , these two matrices are trained end-to-end, Z ∈ R d×1 , after the softmax operation, the obtained a ∈ R C×1 , z is a compact feature, and a is a soft attention vector;

[0036] The final feature map V is obtained through the attention weights on different kernels:

[0037] V c = a c ·U1 + b c ·U2, a c + b c = 1

[0038] wherein, U1, U2 ∈ R H×W×C , a, b ∈ R C×1 .

[0039] As can be seen from the above technical solutions, the beneficial effects of the present invention are as follows: First, compared with the prior art, the present invention can more accurately extract the soybean planting area from remote sensing images. By designing for the crop spatial distribution information extraction task and the characteristics of high-resolution remote sensing data, good results are achieved in terms of efficiency, speed, applicability, and accuracy. Second, based on the improvement of the UNet network, the SKNet Block module is added after each double-layer convolution. It has good scalability and high extraction accuracy, and can be adjusted to apply to remote sensing satellite images of different sensors. Third, through verification, the high-resolution remote sensing image classification results obtained by the present invention are superior to the comparative classification algorithm in terms of evaluation indicators, and the generated classification results can better maintain the smoothness and integrity of the edges. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is the flowchart of the method of the present invention;

[0041] Figure 2 is the schematic diagram of the data label making process;

[0042] Figure 3 is the schematic diagram of the set transformation included in data augmentation;

[0043] Figure 4 is the prediction result map obtained by using the Sentinel-2 dataset as the training set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] As Figure 1 shown, a method for extracting the soybean planting area of remote sensing images based on the SK-UNet deep learning network, the method includes the following steps in sequence:

[0045] (1) Obtain Sentinel-2 remote sensing images with a resolution of 10m and perform preprocessing. Use ArcGIS software to outline the vector of the soybean area, obtaining the original image data and label data of the soybean area. The size of the original image data of the soybean area is 256×256×3, and the size of the label data is 256×256×1. The original image data and label data of the soybean area constitute the remote sensing image dataset;

[0046] (2) Use the UNet network as the basic network, embed the SKNetBlock module in the basic convolutional layers of its encoder and decoder, and construct an improved UNet network model, namely the SK-UNet model, as the image extraction network for the soybean planting area;

[0047] (3) Input the remote sensing image dataset into the SK-UNet model for training to obtain the trained SK-UNet model;

[0048] (4) Obtain the remote sensing image to be extracted and perform preprocessing;

[0049] (5) Input the preprocessed remote sensing image to be extracted into the trained SK-UNet model to obtain the image extraction result of the soybean planting area.

[0050] The specific content of step (1) is as follows: Obtain Sentinel-2 remote sensing image data, and perform preprocessing on it as needed through ENVI software. Use ArcGIS software to label the soybeans in the study area with Sentinel-2 data as a reference, create a vector file on the original image through Create New Shapefile, outline the vector of the soybean area, and at the same time keep the spatial reference consistent with the remote sensing image. Then, complete the labeling of the soybeans in the study area in sequence. Among them, the soybeans are white and set to 1, and the non-soybeans (background) are black and set to 0. Then, crop it into sample label data of 256×256 size. At the same time, save the original remote sensing image as a JPG format output. This image is a 3-channel 8-bit JPG image, and then crop it into original image data of 256×256 size. The file names of the original image data and the label data correspond one by one. In order to avoid distortion of features during the data augmentation process, the data augmentation method used is geometric transformation. Through the data augmentation method, 5000 remote sensing image datasets with a fixed size of 256 pixels × 256 pixels are obtained, and the remote sensing image dataset is divided into a training set and a validation set according to a ratio of 9:1.

[0051] The specific content of step (2) includes the following steps:

[0052] (2a) Extract features based on the UNet network as the basic framework. Set the Unet network to include an encoder, a decoder, and a skip connection part;

[0053] (2b) The encoder is set to consist of a 3×3 convolutional layer, a ReLU function, and a 2×2 max pooling layer. The encoder is used to extract image feature information through the convolutional layer and perform four downsamplings. After each pooling operation, the feature image becomes smaller and the number of channels doubles. The decoder performs upsampling through 2×2 transposed convolution, and the decoder part completes four upsamplings. After each upsampling, the size of the feature image increases and the number of channels is reduced by half. Then, through the skip connection part, the low-level details and high-level semantics of the feature map are combined. Finally, through Softmax as the classification layer, the picture pixels are segmented into two categories, one is the background and the other is the extracted soybean area;

[0054] (2c) Set the SKNet Block module; the grouped convolution and global average pooling in the SKNet Block module can reduce the number of model parameters and computational complexity, thereby improving the running efficiency and practicality of the model. Its adaptive convolution module can adaptively adjust the convolution kernel size according to the size of the input feature map, so as to adapt to feature maps of different sizes and improve the adaptability of the model.

[0055] (2d) Embed the SKNet Block module in the basic convolutional layers of the encoder and decoder to obtain the SK-UNet model. The SKNet Block module is used to integrate feature maps of different scales, dynamically adjust the convolution kernel size, and automatically obtain the importance of each feature channel through learning by the UNet network.

[0056] Step (3) specifically includes the following steps:

[0057] (3a) Determine the hyperparameters during the training process and initialize the parameters of the SK-UNet model. The Batch size is 4, the learning rate is 0.0001, the parameter Beta1 of the Adam optimizer is 0.5, Beta2 is 0.999, and Epochs is 100;

[0058] (3b) Input the remote sensing image dataset into the SK-UNet model;

[0059] (3c) Use the SK-UNet model to perform forward propagation on the current training data;

[0060] (3d) Calculate the loss and backpropagate it to the SK-UNet model;

[0061] (3e) Use the Adam optimizer to update the parameters of the SK-UNet model according to the loss value of the SK-UNet model, and repeat steps (3b) to (3d) until the loss value is less than a predetermined threshold.

[0062] The specific content of step (2c) is as follows: The SKNet Block module adopts a non-linear multi-core information aggregation method to achieve the adaptive receptive field size of neurons. It consists of three operators: Split, Fuse, and Select operators. Among them, the Split operator generates multiple paths with different kernel sizes, corresponding to different neuron receptive field sizes. The Fuse operator combines and aggregates the information from multiple paths to obtain a global comprehensive representation of the selection weights. The Select operator aggregates the feature maps of different kernel sizes according to the selection weights;

[0063] The Split operator passes the original feature map through a 3×3 grouped / depth convolution and a 3×3 dilated convolution respectively to generate two feature maps U1 and U2, and then adds these two feature maps to generate the feature map U;

[0064] The Fuse operator uses global average pooling to embed global information, thereby generating the statistical information of the channels. The c-th element of the feature S is calculated by shrinking U in the spatial dimension H×W:

[0065]

[0066] In the formula, F gp represents average pooling of the feature, U c is the c-th dimension feature of the input feature map, W and H represent the width and height of the feature, and i and j are the coordinates of the feature map in the spatial dimension;

[0067] A compact feature z is established to achieve precise and adaptive selection guidance, which is realized through a simple fully connected layer. The formula is as follows:

[0068]

[0069] In the formula, δ is the ReLU function, B represents batch normalization, W∈R d×C is the weight matrix. To study the influence of d on the model efficiency, the reduction rate r is used to control its value:

[0070] d = max(C / r, L)

[0071] In the formula, L represents the minimum value of d, C is the C-th dimension feature of the input, and r is a hyperparameter;

[0072] The Select operator passes the generated feature z through a c and b c two functions, and multiplies the generated function values with the feature maps U1 and U2:

[0073]

[0074] In the formula, A, B∈RC×d , These two matrices are trained end-to-end, Z ∈ R d×1 , After the softmax operation, the obtained a ∈ R C×1 , z is a compact feature, and a is a soft attention vector;

[0075] The final feature map V is obtained through attention weights on different kernels:

[0076] V c = a c ·U1 + b c ·U2, a c + b c = 1

[0077] where U1, U2 ∈ R H×W×C , a, b ∈ R C×1 .

[0078] The following combines simulation experiments to further illustrate the effects of the present invention:

[0079] The computer hardware environment for the experiments of the present invention is a 13th Gen Intel(R) Core(TM) i7-13700K processor, 32.0 Gb of memory, an NVIDIA GeForce RTX 3070 graphics card, with the GPU acceleration library using CUDA 11.8 and the deep learning framework using Pytorch. All subsequent training and testing experiments are based on this machine.

[0080] To verify the effectiveness of the present invention, the SegNet network, the Deeplabv3+ network, and the UNet network are introduced for comparison. The present invention uses five popular criteria, namely accuracy, mean intersection over union, recall, overall accuracy, and F1-score, to evaluate the performance of the proposed model.

[0081] Content and result analysis of the simulation experiment:

[0082] In the selection of the network, the UNet network can undoubtedly be given priority. Its U-shaped network structure performs layer-by-layer convolution and pooling on the image to obtain a more accurate segmentation result. In terms of data input, there is no need for excessive data processing and data augmentation, and the original data can be directly input, reducing the workload and improving efficiency. At the same time, the skip-connection structure it adopts can avoid information loss during the segmentation process and improve the robustness of the network. The SegNet network structure is relatively simple, using convolutional layers and pooling layers to complete feature extraction and upsampling, with small computational complexity, fast training speed, and high computational efficiency, having great advantages in real-time segmentation tasks. Facing the large amount of data in remote sensing images, the SegNet network can still perform well with fewer parameters, and the edge information it retains shows excellent performance in the segmentation of the object contours in remote sensing images. The DeepLabv3+ network uses dilated convolution and Atrous Spatial Pyramid Pooling techniques to increase the receptive field while maintaining the resolution, achieving a better classification effect. Combining the analysis of the above network structures, the UNet, SegNet, and DeepLabv3+ networks are used for comparison to explore the advantages of the SK-UNet model. In the experiment, the training set and the test set are kept the same, trained with Sentinel-2 data, and also predicted with Sentinel-2 data.

[0083] By observing the test results of the four networks, the prediction results of the UNet network are relatively close to the sample label data, but there are still omissions in the edge areas of soybeans. A comprehensive comparison shows that the segmentation results of the SegNet network failed to identify large areas of soybeans, and the phenomenon of soybeans being misclassified as other objects is serious, resulting in large fragmented areas on the classification result map. Compared with the label, the distribution of soybeans is more scattered. Observing the classification result map of the Deeplabv3+ network, the edge of the soybean planting area presents a more rounded effect, especially the recognition of soybeans in edge details is poor. Overall, the phenomenon of misclassifying other objects as soybeans is more obvious. For areas where soybean crops and other crops are intercropped, it is common to misclassify all areas as soybeans. By observing the segmentation effect of the SK-UNet model, as shown in Table 1, a balance is maintained between the U-Net, SegNet and Deeplabv3+ networks, and the classification results are closest to the true labels. At the same time, it can be concluded from the evaluation indicators that the accuracy of the SK-UNet model can reach 93.17%, which is 3 percentage points, 7 percentage points and 5 percentage points higher than the U-Net, SegNet and Deeplabv3+ networks respectively. The average intersection-over-union ratio can reach nearly 83%, while the average intersection-over-union ratios of the other three network models are much lower than the results of the SK-UNet model, all below 80%. The overall segmentation quality of the SK-UNet model is the highest, and it is the best soybean planting area extraction network.

[0084] Table 1 Accuracy evaluation of three networks

[0085]

[0086]

[0087] like Figure 2 As shown in the figure, ArcGIS software was used to label the soybeans in the study area with Sentinel-2 data as a reference. The vector file was created on the original image through the Create New Shapefile tool to outline the vector of the soybean area while keeping the spatial reference consistent with the remote sensing image. Then the soybeans in the study area were labeled one by one, with soybeans in white and set to 1, and non-soybeans (background) in black and set to 0. At this point, the data labels were completed.

[0088] like Figure 3As shown in the figure, in order to avoid feature distortion during data augmentation, the data augmentation method adopted is geometric transformation, which includes brightness enhancement, vertical flipping, horizontal flipping, rotation by 90°, rotation by 180°, rotation by 270°, and random rotation. A total of 5000 remote sensing image datasets with a fixed size of 256 pixels × 256 pixels are obtained, and the remote sensing image datasets are divided into a training set and a validation set according to a ratio of 9:1.

[0089] As Figure 4 shown in the figure, through the training of the remote sensing image dataset, classification result maps of four networks, namely SK-UNet, U-Net, SegNet, and Deeplabv3+, for the image to be predicted are obtained. After comparing them with the labels, the best extraction network for the soybean planting area is selected.

[0090] In summary, the present invention can more accurately extract the soybean planting area from remote sensing images. It is designed for the crop spatial distribution information extraction task and the characteristics of high-resolution remote sensing data, and achieves good results in terms of efficiency, speed, applicability, and accuracy. The present invention is based on the improvement of the UNet network, and an SKNet Block module is added after each double-layer convolution. It has good scalability and high extraction accuracy, and can be adjusted to apply to remote sensing satellite images of different sensors. Through verification, the classification results of the high-resolution remote sensing images obtained by the present invention are superior to the comparative classification algorithms in terms of evaluation indicators, and the generated classification results can better maintain the smoothness and integrity of the edges.

[0091] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting soybean planting areas from remote sensing images based on the SK-UNet deep learning network, characterized in that: The method includes the following steps in sequence: (1) Obtain Sentinel-2 remote sensing images with a resolution of 10m and perform preprocessing. Use ArcGIS software to draw the vector of the soybean area to obtain the original image data and label data of the soybean area. The size of the original image data of the soybean area is 256×256×3, and the size of the label data is 256×256×1. The original image data and label data of the soybean area constitute a remote sensing image dataset; (2) Take the UNet network as the basic network, embed the SKNet Block module in the basic convolutional layers of its encoder and decoder, and construct an improved UNet network model, namely the SK-UNet model, as the image extraction network for the soybean planting area; (3) Input the remote sensing image dataset into the SK-UNet model for training to obtain the trained SK-UNet model; (4) Obtain the remote sensing image to be extracted and perform preprocessing; (5) Input the preprocessed remote sensing image to be extracted into the trained SK-UNet model to obtain the extraction result of the soybean planting area image; The SKNet Block module adopts a non-linear multi-core information aggregation method to realize the adaptive receptive field size of neurons. It consists of three operators: Split, Fuse and Select operators; among them, the Split operator generates multiple paths with different kernel sizes, corresponding to different neuron receptive field sizes, and the Fuse operator combines and aggregates the information from multiple paths to obtain a global comprehensive representation of the selection weights, and the Select operator aggregates the feature maps of different kernel sizes according to the selection weights; The Split operator generates two feature maps U1 and U2 by passing the original feature map through a 3×3 grouped / depth convolution and a 3×3 dilated convolution respectively, and then adds these two feature maps to generate the feature map U; The Fuse operator uses global average pooling to embed global information, thereby generating the statistical information of the channels. The c-th element of the feature S is calculated by shrinking U through the spatial dimension H×W: where F gp represents average pooling of features, U c is the c-th dimensional feature of the input feature map, W and H represent the width and height of the feature, and i and j are the coordinates of the feature map in the spatial dimension; Establish a compact feature z to achieve precise and adaptive selection guidance, which is realized through a simple fully connected layer. The formula is as follows: where δ is the ReLU function, B represents batch normalization, and W ∈ R d×C is the weight matrix. To study the effect of d on the model efficiency, the reduction rate r is used to control its value: In the formula, L represents the minimum value of d, C is the C-th dimensional feature of the input, and r is a hyperparameter; The Select operator passes the generated feature z through functions a c and b c and multiplies the generated function values with the feature maps U1 and U2: where \(A, B\in R\) C×d , these two matrices are trained end-to-end, \(Z\in R\) d×1 , after the softmax operation, the resulting \(a\in R\) C×1 , \(z\) is the compact feature, and \(a\) is the soft attention vector; The final feature map V is obtained through the attention weights on different kernels: V c = a c ·U1 + b c ·U2, a c + b c = 1 where \(U1, U2\in R\) H×W×C , \(a, b\in R\) C×1 .

2. The method for extracting soybean planting areas from remote sensing images based on the SK-UNet deep learning network according to claim 1, wherein: The specific steps of step (2) include the following steps: (2a) Extract features based on the UNet network as the basic framework. It is set that the Unet network includes an encoder, a decoder and a skip connection part; (2b) The encoder is set to consist of a 3×3 convolutional layer, a ReLU function, and a 2×2 max pooling layer. The encoder is used to extract image feature information through the convolutional layer and perform four downsamplings. After each pooling operation, the size of the feature image becomes smaller and the number of channels doubles. The decoder performs upsampling through a 2×2 transposed convolution, and the decoder part completes four upsamplings. After each upsampling, the size of the feature image increases and the number of channels is reduced by half. Then, through the skip connection part, the low-level details of the feature map are combined with the high-level semantics. Finally, through Softmax as the classification layer, the picture pixels are segmented into two categories, one is the background, and the other is the extracted soybean area; (2c) Set the SKNet Block module: (2d) Embed the SKNet Block module in the basic convolutional layers of the encoder and decoder to obtain the SK-UNet model. The SKNet Block module is used to integrate feature maps of different scales, dynamically adjust the size of the convolutional kernel, and automatically obtain the importance of each feature channel through learning with the UNet network.

3. The method for extracting soybean planting areas from remote sensing images based on the SK-UNet deep learning network according to claim 1, wherein: (3) The specific steps of step (3) are as follows: (3a) Determine the hyperparameters during the training process and initialize the parameters of the SK-UNet model. The Batch size is 4, the learning rate is 0.0001, the parameter Beta1 of the Adam optimizer is 0.5, Beta2 is 0.999, and Epochs is 100; (3b) Input the remote sensing image dataset into the SK-UNet model; (3c) Use the SK-UNet model to perform forward propagation on the current training data; (3d) Calculate the loss and backpropagate it to the SK-UNet model; (3e) Use the Adam optimizer to update the parameters of the SK-UNet model according to the loss value of the SK-UNet model. Repeat steps (3b) to (3d) until the loss value is less than a predetermined threshold.

Images