A crop classification method based on semantic super edge three driving

Abstract

The application discloses a crop classification method based on semantic super-resolution edge three driving, and belongs to the technical field of intelligent agriculture. In order to solve the problem of high-precision classification of crops in medium and low resolution remote sensing images. The application comprises the following steps: preparing multi-crop samples; building a feature extraction backbone network Deep_HR to obtain high-level semantic features and low-level detail features extracted by the backbone network; designing a multi-scale feature fusion module STD_ASPP to output STD_ASPP fusion features; establishing a semantic segmentation head based on a segmentation decoder SEG_Decoder and a super-resolution head based on a super-resolution decoder SR_Decoder; inputting the low-level detail features extracted by the backbone network and the STD_ASPP fusion features into the semantic segmentation head and the super-resolution head at the same time; training in combination with four kinds of losses; obtaining a trained crop classification model based on semantic super-resolution edge three driving; and outputting a final crop classification result.

Description

Technical Field

[0001] This invention belongs to the field of smart agriculture technology, specifically relating to a crop classification method based on semantic super-resolution edge three-drive. Background Technology

[0002] With the increasing global demand for food security and the rapid development of precision agriculture technology, achieving efficient and accurate crop monitoring and classification has become a core focus of agricultural policies in various countries. It is necessary to fully utilize remote sensing imagery and advanced machine learning methods to construct a crop monitoring system to provide a scientific basis for food security decisions. This places unprecedented demands on the accuracy and robustness of crop classification technology, especially in low-resolution remote sensing imagery processing scenarios, where overcoming data quality limitations to achieve accurate classification has become a key challenge.

[0003] Against this backdrop, the deep integration of deep learning and remote sensing image analysis has become the mainstream technical approach for crop classification. Semantic segmentation models based on convolutional neural networks (CNNs) and the Deeplab series of networks have made significant progress in crop classification tasks due to their powerful feature extraction capabilities, becoming the core framework for current technological applications. However, limited by the acquisition conditions of remote sensing images, existing technologies still face insurmountable performance bottlenecks: mainstream remote sensing satellite images such as Landsat and Sentinel-2 are limited by sensor hardware, with spatial resolution mostly maintained at the level of 10m~30m / pixel. This results in severely blurred details of small plots in the field, crop boundaries, and mixed-crop areas. These detailed features are precisely the key basis for distinguishing different crop species and judging crop growth status, directly affecting the final accuracy of semantic segmentation.

[0004] The rise of super-resolution technology has provided an effective solution to this problem. Its ability to reconstruct high-resolution details from low-resolution images through algorithmic models can directly compensate for the performance deficiencies of sensor hardware, laying the foundation for improved semantic segmentation accuracy. Numerous studies have confirmed the clear value of super-resolution modules in enhancing the performance of semantic segmentation of remote sensing images: high-quality super-resolution reconstruction can restore key details such as crop textures and plot boundaries, enabling segmentation models to more accurately capture category differences, especially in complex scenarios such as distinguishing similar crops and identifying small plots.

[0005] Despite the enormous potential of super-resolution technology, existing technologies still suffer from three major shortcomings in crop semantic segmentation applications, severely hindering the improvement of classification accuracy: insufficient synergy between super-resolution and semantic segmentation tasks, limiting the effectiveness of detail enhancement; single-dimensional feature extraction, resulting in insufficient semantic and edge information mining; and the lack of multi-feature fusion mechanisms, failing to fully unleash the synergistic optimization efficiency. To address these shortcomings, it is urgent to construct a novel crop semantic segmentation technology system that breaks through existing technological bottlenecks through deep coupling of super-resolution and segmentation tasks, accurate extraction of multi-dimensional features, and the design of efficient fusion mechanisms. Summary of the Invention

[0006] The problem this invention aims to solve is how to achieve high-precision classification of crops in medium- and low-resolution remote sensing images, and proposes a crop classification method based on semantic super-resolution edge three-drive.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A crop classification method based on semantic super-resolution edge tri-drive includes the following steps:

[0009] S1. Sample making: Select remote sensing images of crops with low cloud cover and good atmospheric transmittance during the key stages of sowing, growth, maturity or excavation, and make samples of various crops.

[0010] S2. Feature Extraction: Combining the advantages of HRNet and Deeplabv3+ type backbone networks, a feature extraction backbone network Deep_HR is built to separate high-level semantic features and low-level detail features from the multi-class crop samples in step S1, thereby obtaining the high-level semantic features extracted by the backbone network and the low-level detail features extracted by the backbone network.

[0011] S3. Module Design: Based on the characteristics of small crop samples, the multi-scale feature fusion module STD_ASPP is designed by combining the edge enhancement module EE_Block and the high resolution preservation module HRK_Block. The high-level semantic features extracted by the backbone network obtained in step S2 are input into the multi-scale feature fusion module STD_ASPP, and the STD_ASPP fused features are output.

[0012] S4. Multi-task head construction: Establish a semantic segmentation head based on the segmentation decoder SEG_Decoder and a super-resolution head based on the super-resolution decoder SR_Decoder. Input the low-level detail features extracted by the backbone network obtained in step S2 and the STD_ASPP fusion features obtained in step S3 into the semantic segmentation head and the super-resolution head at the same time. Combine the cross-entropy loss CE Loss, Dice coefficient loss Dice Loss, mean squared error loss MSE Loss and feature affinity loss FA Loss for training to obtain the trained crop classification model based on semantic super-resolution edge three-drive, and output the final crop classification result;

[0013] S5. Result Inference: Collect untrained crop remote sensing images of the same resolution, input them into the trained crop classification model based on semantic super-resolution edge three-drive constructed in steps S2-S4, perform crop inference prediction, and output crop classification results.

[0014] Furthermore, the specific implementation method for step S1 of creating multiple crop samples includes the following steps:

[0015] S1.1. Use annotation tools to draw vector-format polygonal crop surfaces on the acquired remote sensing images of crops;

[0016] S1.2. Convert the vector format to raster labels. At the same resolution, the raster labels are mapped one-to-one to the pixels of the crop remote sensing image.

[0017] S1.3. Combine the collected crop remote sensing image samples and raster labels into a unified format to obtain the dataset;

[0018] S1.4. Apply data augmentation techniques such as rotation, mirroring, scaling, sharpening, contrast adjustment, perspective adjustment, and gamma correction to the dataset obtained in step S1.3 to obtain multiple crop samples.

[0019] Furthermore, the specific implementation method of step S2 includes the following steps:

[0020] S2.1. Construct the feature extraction backbone network Deep_HR, including the initial feature extraction module Stem, the basic module Basic Block, the downsampling module, the fusion convolution module, and the low-level feature extraction module;

[0021] The feature extraction backbone network Deep_HR consists of three branches: Branch 1 is Branch1, Branch 2 is Branch2, and Branch 3 is Branch3.

[0022] S2.2. The initial feature extraction module Stem is constructed by sequentially connecting a 3×3 convolution with a stride of 2, a batch normalization layer (BN), an activation layer (ReLU), a 3×3 convolution with a stride of 2, a batch normalization layer (BN), and an activation layer (ReLU).

[0023] S2.3. The basic module Basic Block consists of a main branch and a stride-determining residual branch. The main branch is composed of a 3×3 convolutional layer with a given stride, a BN layer, a ReLU layer, and a 3×3 convolutional layer with a stride of 1 connected in sequence.

[0024] The condition for determining the residual branch by step size is that the step size is not 1 or the input and output channels are not equal. If the condition is not met, the residual branch remains unchanged. If the condition is met, it goes through a 3×3 convolutional layer and a BN layer with a step size of 1 as the output of the step size residual branch.

[0025] The outputs of the main branch and the residual branch are passed through a ReLU layer and used as the output of the basic module.

[0026] S2.4. The constructed downsampling module is divided into downsampling module 1 and downsampling module 2. Downsampling module 1 is applied to the input of branch 2 in the HRNet backbone network, and downsampling module 2 is applied to the input of branch 3 in the HRNet backbone network.

[0027] Both downsampling module 1 and downsampling module 2 are convolutions with a 3×3 kernel and a stride of 2, stacked with a batch normalization (BN) layer, and then stacked with an activation layer ReLU.

[0028] S2.5. The constructed fused convolution is a superposition of a convolutional layer with a 1×1 kernel, a batch normalization layer, and a ReLU layer. The number of input channels is the sum of the three branches of HRNet, and the output is 512.

[0029] The constructed low-level feature extraction module includes a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation layer;

[0030] S2.6. The three branches of the constructed feature extraction backbone network are as follows:

[0031] Branch 1 is a high-resolution branch, consisting of two layers of Basic Blocks; Branch 2 is a medium-resolution branch, consisting of two layers of stacked Basic Blocks; Branch 3 is a low-resolution branch, consisting of two layers of stacked Basic Blocks.

[0032] After the initial feature extraction module Stem, the crop samples of step S1 are directly input into branch 1, directly input into downsampling block 1 and then input into branch 2, and directly input into downsampling block 1, then input into downsampling block 2 and then input into branch 3 respectively.

[0033] Branches 2 and 3 are upsampled to the size of the feature map of branch 1 and then concatenated with the output of branch 1 to form a fused feature. The fused feature is then passed through a fused convolutional layer to obtain the high-level semantic features extracted by the backbone network.

[0034] Branch 1 directly outputs the low-level detail features extracted by the backbone network.

[0035] Furthermore, the specific implementation method of step S3 includes the following steps:

[0036] S3.1. Construct an edge enhancement module EE_Block, equipped with edge detection kernels in the x and y directions. The edge detection kernels have sizes of 3×3 and 5×5. Based on the four initialized edge detection kernels, four edge features are obtained. Then, different edge weights are applied, and the corresponding multiplications are summed to normalize the features. Then, the edge enhancement features are obtained through a detail extraction module containing a convolutional layer with a 1×1 convolutional kernel, a batch normalization layer, and an activation layer.

[0037] S3.2. Construct the high-resolution preservation module HRK_Block, which concatenates the input features after passing them through four branches and then passes them through a 1×1 convolution, batch normalization layer and activation layer to obtain high-resolution preserved features. Branch 1 includes a 1×1 convolution layer, batch normalization layer and activation layer; Branch 2 includes a 3×3 convolution layer, batch normalization layer and activation layer; Branch 3 includes a 5×5 convolution layer, batch normalization layer and activation layer; Branch 4 includes a 3×3 dilated convolution layer, batch normalization layer and activation layer.

[0038] S3.3. Construct a multi-scale feature fusion module STD_ASPP. Taking the high-level semantic features extracted from the backbone network as input, it first passes through a projection layer module consisting of a 1×1 convolutional layer, a batch normalization layer, and an activation layer. The output is the projection layer output features. ;

[0039] First, input feature layer 1, which includes one HRK_Block, followed by a 3×3 convolution kernel, a dilated convolution with an inflation rate of 1, a batch normalization layer, and an activation layer. The output of feature layer 1 is F1.

[0040] F1 is then input into feature layer 2, which includes a 3×3 convolution kernel, a dilated convolution with an inflation rate of 2, a batch normalization layer, an activation layer, and an EE_Block, resulting in the output of feature layer 2 as F2.

[0041] F2 is then input into feature layer 3, which includes a 3×3 convolution kernel, a dilated convolution with an inflation rate of 4, a batch normalization layer, an activation layer, and an EE_Block, resulting in the output of feature layer 3 as F3;

[0042] F3 is then input into feature layer 4, which includes a 3×3 convolution with an inflation rate of 6, a batch normalization layer, and an activation layer, resulting in the output of feature layer 4 as F4.

[0043] After F1, F2, F3, and F4 are simultaneously input into a fusion layer consisting of a 1×1 convolution, a batch normalization layer, and an activation layer, they are concatenated with the outputs of the residual connection and the HRK_Block single connection part, and then passed through another fusion layer consisting of a 1×1 convolution, a batch normalization layer, and an activation layer to obtain the STD_ASPP fusion feature.

[0044] Furthermore, the specific implementation method of step S4 includes the following steps:

[0045] S4.1. Construct the segmentation decoder SEG_Decoder, which consists of a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.5, a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.1, and a 1×1 convolutional layer stacked in sequence.

[0046] S4.2. Construct a super-resolution decoder SR_Decoder, which consists of a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.5, a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.1, and a 1×1 convolutional layer stacked in sequence. The difference between SR_Decoder and SEG_Decoder is the difference in the number of input channels and the number of output channels.

[0047] S4.3. The low-level detail features extracted by the backbone network obtained in step S2 are processed by an input preprocessing module consisting of a 1×1 convolutional layer, a batch normalization layer, and an activation layer, and then concatenated with the STD_ASPP fusion features obtained in step S3 after interpolation to obtain the input features of the semantic segmentation head and the super-resolution head.

[0048] S4.4. Construct a semantic segmentation detection head. Input the input features of the semantic segmentation head and the super-resolution head obtained in step S4.3 into SEG_Decoder. After extracting the features, interpolate and resample to output the semantic segmentation result.

[0049] S4.5. Construct a super-resolution detection head. Input the input features of the semantic segmentation head and the super-resolution head obtained in step S4.3 into SR_Decoder. Then, input them into the combined module consisting of the deconvolution module and the EDSR module three times in sequence. Finally, perform a 1×1 convolution to output the super-resolution result.

[0050] The EDSR module consists of a main branch that passes through a 3×3 convolutional layer, a ReLU activation layer, and another 3×3 convolutional layer in sequence, and a residual branch that passes through a 1×1 convolutional layer and is then connected to the main branch.

[0051] S4.6. The semantic segmentation result obtained in step S4.4 is input into CE Loss and Dice Loss to calculate the semantic loss. The super-resolution result obtained in step S4.5 is input into MSE Loss to calculate the super-resolution loss. Then, the semantic loss and the super-resolution loss are input into FA Loss to obtain the feature affinity loss, and the final crop classification result is obtained.

[0052] According to claim 5, a crop classification method based on semantic super-resolution edge three-drive is characterized in that step S5 outputs a raster image with classification results, wherein the value of each pixel in the image represents a preset crop category code.

[0053] The beneficial effects of this invention are:

[0054] This invention presents a crop classification method based on semantic super-resolution edge tri-drive, which, through an improved backbone network architecture, enables true multi-resolution feature collaborative extraction. Traditional backbone networks such as ResNet and VGG suffer from significant spatial information attenuation during feature extraction. As the number of network layers increases, the resolution of the feature map decreases sharply, leading to severe loss of boundary information and detailed features of small-target crops.

[0055] This invention discloses a crop classification method based on semantic super-resolution edge tri-drive, employing a parallel multi-branch design. The backbone network, through the collaborative work of high-resolution, medium-resolution, and low-resolution branches, simultaneously captures crop features at different scales. Meanwhile, the improved ASPP module also acquires multi-scale features through multiple branches. The backbone network fuses feature maps from multiple branches using channel splicing, forming richer feature representations with more channels, effectively integrating local details and global semantic information. Simultaneously, the 3×3 Sobel operator and the 5×5 Scharr-like operator are used to capture sharp edge and smooth edge features, respectively. Each scale includes edge detection in the x and y directions, ensuring complete capture of edge information in all directions. Through learnable edge weight parameters, the network automatically optimizes the contribution of different edge features. The super-resolution decoder is organically combined with the semantic segmentation decoder. The semantic segmentation-super-resolution branch helps learn higher-resolution dense label prediction representations. The single-image super-resolution branch can recover detailed structural information and combines four losses: CE_Loss, Dice_Loss, MSE_Loss, and FA_Loss (feature affinity loss), which enhance the high-resolution representation of semantic segmentation through detailed structural information.

[0056] This invention presents a crop classification method based on semantic super-resolution edge tri-drive, which maximizes the preservation of spatial information of crop boundaries and small targets. Through multi-level technological innovation, it achieves breakthroughs in backbone network architecture, feature enhancement mechanisms, multi-scale contextual understanding, and training strategies, providing a complete, efficient, and accurate technical solution for intelligent remote sensing monitoring of crops. This technology not only significantly improves the accuracy and boundary quality of crop segmentation but also demonstrates significant advantages in engineering practicality, possessing broad application prospects and important industrial value. Attached Figure Description

[0057] Figure 1 This is a flowchart of a crop classification method based on semantic super-resolution edge three-drive as described in this invention;

[0058] Figure 2 This is a model architecture diagram of a crop classification method based on semantic super-resolution edge three-drive as described in this invention;

[0059] Figure 3 This is a diagram showing several types of constituent modules of the backbone network described in this invention;

[0060] Figure 4 This is a diagram of the backbone network architecture described in this invention;

[0061] Figure 5 This is an architecture diagram of the edge enhancement module in the STD_ASPP module described in this invention;

[0062] Figure 6 This is a diagram of the high-resolution preservation module architecture in the STD_ASPP module described in this invention;

[0063] Figure 7 This is an architecture diagram of the STD_ASPP module described in this invention;

[0064] Figure 8 This is an architecture diagram of the decoders SEG_Decoder and SR_Decoder described in this invention;

[0065] Figure 9 This is a diagram illustrating the architecture of the detection head described in this invention.

[0066] Figure 10 This is a schematic diagram illustrating the effect that the edge enhancement module described in this invention can achieve at each step in a certain embodiment. (a) is the edge feature extracted by a 3×3 kernel size Sobel kernel in the x-direction, (b) is the edge feature extracted by a 3×3 kernel size Sobel kernel in the y-direction, (c) is the edge feature extracted by a 5×5 kernel size Scharr-like edge detection kernel in the x-direction, (d) is the edge feature extracted by a 5×5 kernel size Scharr-like edge detection kernel in the x-direction, (e) is the weighted fused edge, (f) is the normalized edge, (g) is the feature before enhancement, and (h) is the feature after enhancement, which is the final output.

[0067] Figure 11 The crop identification result image described in this invention is compared with the original image, where (a) is the label, (b) is the identification result of the method described in this invention, (c) is the identification result of Deeplabv3+, and (d) is the original image. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0069] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0070] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 11 Detailed explanation is as follows:

[0071] Example 1:

[0072] A crop classification method based on semantic super-resolution edge tri-drive includes the following steps:

[0073] S1. Sample making: Select remote sensing images of crops with low cloud cover and good atmospheric transmittance during the key stages of sowing, growth, maturity or excavation, and make samples of various crops.

[0074] Furthermore, the specific implementation method for step S1 of creating multiple crop samples includes the following steps:

[0075] S1.1. Use annotation tools to draw vector-format polygonal crop surfaces on the acquired remote sensing images of crops;

[0076] S1.2. Convert the vector format to raster labels. At the same resolution, the raster labels are mapped one-to-one to the pixels of the crop remote sensing image.

[0077] S1.3. Combine the collected crop remote sensing image samples and raster labels into a unified format to obtain the dataset;

[0078] S1.4. Apply data augmentation techniques such as rotation, mirroring, scaling, sharpening, contrast adjustment, perspective adjustment, and gamma correction to the dataset obtained in step S1.3 to obtain multiple crop samples.

[0079] Furthermore, in this embodiment, the key phenological period for rapeseed and wheat, February and March, is selected as the main identification time.

[0080] Furthermore, in step S1 of this embodiment, the label assignment is as follows: background-0, wheat-1, rapeseed-2, other-3, and the slice size is 512×512.

[0081] S2. Feature Extraction: Combining the advantages of HRNet and Deeplabv3+ type backbone networks, a feature extraction backbone network Deep_HR is built to separate high-level semantic features and low-level detail features from the multi-class crop samples in step S1, thereby obtaining the high-level semantic features extracted by the backbone network and the low-level detail features extracted by the backbone network.

[0082] Furthermore, the specific implementation method of step S2 includes the following steps:

[0083] S2.1. Construct the feature extraction backbone network Deep_HR, including the initial feature extraction module Stem, the basic module Basic Block, the downsampling module, the fusion convolution module, and the low-level feature extraction module;

[0084] The feature extraction backbone network Deep_HR consists of three branches: Branch 1 is Branch1, Branch 2 is Branch2, and Branch 3 is Branch3.

[0085] S2.2. The initial feature extraction module Stem is constructed by sequentially connecting a 3×3 convolution with a stride of 2, a batch normalization (BN) layer, a ReLU activation layer, another 3×3 convolution with a stride of 2, a batch normalization (BN) layer, and a ReLU activation layer. The expression is as follows:

[0086] (BN( ( ))))))

[0087] in, These are the features output by the initial feature extraction module. It is the activation layer. It is a batch of normalized layers. It is a 3×3 convolutional layer with a stride of 2. These are the input features for the initial feature extraction module.

[0088] This method achieves preliminary feature extraction and spatial dimension compression of input data, maps multi-channel multispectral remote sensing data to a unified feature space, and achieves 4x downsampling through two stride 2 convolutions, balancing computational efficiency and feature preservation.

[0089] S2.3. The basic module Basic Block consists of a main branch and a stride-determining residual branch. The main branch is composed of a 3×3 convolutional layer with a given stride, a BN layer, a ReLU layer, and a 3×3 convolutional layer with a stride of 1 connected in sequence.

[0090] The condition for determining the residual branch by step size is that the step size is not 1 or the input and output channels are not equal. If the condition is not met, the residual branch remains unchanged. If the condition is met, it goes through a 3×3 convolutional layer and a BN layer with a step size of 1 as the output of the step size residual branch.

[0091] The outputs of the main branch and the residual branch are passed through a ReLU layer and used as the output of the basic module.

[0092] This basic module is used to build the foundational units of deep networks, solve the gradient vanishing problem, enable feature reuse through residual connections, enhance the network's expressive power, maintain the spatial resolution of the feature map, and focus on feature transformation.

[0093]

[0094]

[0095]

[0096] in As input features, Represents a 3×3 convolutional layer. Representatives of the normalization layer, Represents the activation layer. This represents a 3×3 convolutional layer with a stride of 1. Represents the output characteristics of the main branch. Represents the characteristics of the residual branch output. Represents the number of input channels. Represents the number of output channels. Otherwise, This represents the output of the basic module.

[0097] S2.4. The constructed downsampling module is divided into downsampling module 1 and downsampling module 2. Downsampling module 1 is applied to the input of branch 2 in the HRNet backbone network, and downsampling module 2 is applied to the input of branch 3 in the HRNet backbone network.

[0098] Downsampling module 1 and downsampling module 2 are both convolutional modules with a 3×3 kernel and a stride of 2, stacked with a batch normalization (BN) layer, and then stacked with a ReLU activation layer. The expression is as follows:

[0099]

[0100] in, This represents the downsampled output features. Represents the activation layer. Representatives of the normalization layer, This represents a 3×3 convolutional layer with a stride of 2. This represents the input characteristics. Downsampling module 1 and downsampling module 2 share this formula, differing only in the input and output channels.

[0101] S2.5. The constructed fused convolution is a superposition of a convolutional layer with a 1×1 kernel, a batch normalization layer, and a ReLU layer. The number of input channels is the sum of the three branches of HRNet, and the output is 512.

[0102] The constructed low-level feature extraction module includes a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation layer;

[0103] S2.6. The three branches of the constructed feature extraction backbone network are as follows:

[0104] Branch 1 is a high-resolution branch, consisting of two layers of Basic Blocks; Branch 2 is a medium-resolution branch, consisting of two layers of stacked Basic Blocks; Branch 3 is a low-resolution branch, consisting of two layers of stacked Basic Blocks.

[0105] After the initial feature extraction module Stem passes through the multi-type crop samples in step S1, they are directly input into branch 1, directly input into downsampling block 1 and then input into branch 2, and directly input into downsampling block 1, then input into downsampling block 2 and then input into branch 3 respectively.

[0106] Branches 2 and 3 are upsampled to the size of the feature map of branch 1 and then concatenated with the output of branch 1 to form a fused feature. The fused feature is then passed through a fused convolutional layer to obtain the high-level semantic features extracted by the backbone network.

[0107] Branch 1 directly outputs the low-level detail features extracted by the backbone network.

[0108] The expressions for the three branches of the feature extraction backbone network are as follows:

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115] in, This represents the characteristics that have passed through the Stem module. Represents input features, Represents the Stem module. Represents the basic module. The output feature representing the nth branch, where n∈[1,2,3] This represents the downsampling block m, where m∈[1,2]. Represents upsampling, This represents normal downsampling. Represents fused convolution, Represents splicing, Represents the low-level feature extraction module. Represents high-level characteristics, It represents a low-level feature.

[0116] The Deep_HR feature extraction backbone network combines HRNet and Deeplabv3+. The HRNet feature extraction part maintains high-resolution feature maps, suitable for dense, small targets. The Deeplabv3+ feature extraction part captures multi-scale contextual information. It also references the residual connection design of ResNet. Furthermore, multi-resolution feature extraction mainly follows the HRNet concept, maintaining high-resolution feature maps suitable for dense, small targets, processing features at different resolutions in parallel and sequential multi-branch processing, and enhancing expressive power through feature fusion. For crop segmentation optimization: dense connections are suitable for narrow farmland targets; efficient feature fusion, cross-resolution feature fusion, residual connections avoid gradient vanishing, and a lightweight design ensures efficiency.

[0117] S3. Module Design: Based on the characteristics of small crop samples, the multi-scale feature fusion module STD_ASPP is designed by combining the edge enhancement module EE_Block and the high resolution preservation module HRK_Block. The high-level semantic features extracted by the backbone network obtained in step S2 are input into the multi-scale feature fusion module STD_ASPP, and the STD_ASPP fused features are output.

[0118] Furthermore, the specific implementation method of step S3 includes the following steps:

[0119] S3.1. Construct an edge enhancement module EE_Block, equipped with edge detection kernels in both the x and y directions. The kernel sizes are 3×3 and 5×5. Four edge features are obtained based on four initialized edge detection kernels. Then, different edge weights are applied, multiplied together, and summed for feature normalization. Finally, a detail extraction module containing a convolutional layer with a 1×1 kernel, a batch normalization layer, and an activation layer is used to obtain the edge enhancement features, expressed as:

[0120]

[0121]

[0122]

[0123]

[0124] in, This represents the initial 3×3 edge detection convolution kernel in the x-direction. This represents the initial 3×3 edge detection convolution kernel in the y-direction. This represents the initial 5×5 edge detection convolution kernel in the x-direction. This represents the initial 5×5 edge detection convolution kernel in the y-direction.

[0125] S3.2. Construct the high-resolution preservation module HRK_Block, which concatenates the input features after passing them through four branches and then passes them through a 1×1 convolution, batch normalization layer, and activation layer to obtain high-resolution preserved features. Branch 1 includes a 1×1 convolutional layer, batch normalization layer, and activation layer; branch 2 includes a 3×3 convolutional layer, batch normalization layer, and activation layer; branch 3 includes a 5×5 convolutional layer, batch normalization layer, and activation layer; and branch 4 includes a 3×3 dilated convolutional layer, batch normalization layer, and activation layer. The expression is as follows:

[0126]

[0127]

[0128]

[0129]

[0130]

[0131] in, Let x represent the output feature of the x-th branch, where x∈[1,2,3,4]. The characteristics represented by the output of HRK_Block Represents input features, This represents an n×n convolutional layer, where n∈[1,3,5]. Represents a 3×3 dilated convolution. Representatives of the normalization layer, Represents the activation layer. It represents splicing.

[0132] S3.3. Construct a multi-scale feature fusion module STD_ASPP. Taking the high-level semantic features extracted from the backbone network as input, it first passes through a projection layer module consisting of a 1×1 convolutional layer, a batch normalization layer, and an activation layer. The output is the projection layer output features. ;

[0133] First, input feature layer 1, which includes one HRK_Block, followed by a 3×3 convolution kernel, a dilated convolution with an inflation rate of 1, a batch normalization layer, and an activation layer. The output of feature layer 1 is F1.

[0134] F1 is then input into feature layer 2, which includes a 3×3 convolution kernel, a dilated convolution with an inflation rate of 2, a batch normalization layer, an activation layer, and an EE_Block, resulting in the output of feature layer 2 as F2.

[0135] F2 is then input into feature layer 3, which includes a 3×3 convolution kernel, a dilated convolution with an inflation rate of 4, a batch normalization layer, an activation layer, and an EE_Block, resulting in the output of feature layer 3 as F3;

[0136] F3 is then input into feature layer 4, which includes a 3×3 convolution with an inflation rate of 6, a batch normalization layer, and an activation layer, resulting in the output of feature layer 4 as F4.

[0137] F1, F2, F3, and F4 are simultaneously input into a fusion layer consisting of a 1×1 convolution, batch normalization layer, and activation layer. After concatenation with the outputs of the residual connections and the HRK_Block single-connection portion, the result is passed through another fusion layer consisting of a 1×1 convolution, batch normalization layer, and activation layer to obtain the STD_ASPP fusion feature, expressed as:

[0138]

[0139]

[0140]

[0141]

[0142]

[0143]

[0144]

[0145]

[0146] in, Represents the output features of the projection layer. High-level features extracted from the backbone network, Represents a 1×1 convolutional layer. Representatives of the normalization layer, Represents the activation layer. Represents a high-resolution holding module. Let i represent a 3×3 convolutional layer with an inflation rate of i, where i∈[1,2,4,6]. Represents the edge enhancement module. F1~4 represent the features output by Stages 1~4. Represents splicing, and It refers to the process characteristics produced by the process, and has no special meaning. This represents the characteristics output by the STD_ASPP module.

[0147] S4. Multi-task head construction: Establish a semantic segmentation head based on the segmentation decoder SEG_Decoder and a super-resolution head based on the super-resolution decoder SR_Decode. Input the low-level detail features extracted by the backbone network obtained in step S2 and the STD_ASPP fusion features obtained in step S3 into the semantic segmentation head and the super-resolution head at the same time. Combine the cross-entropy loss CE Loss, Dice coefficient loss Dice Loss, mean squared error loss MSE Loss and feature affinity loss FA Loss for training to obtain the trained crop classification model based on semantic super-resolution edge three-drive, and output the final crop classification result;

[0148] Furthermore, the specific implementation method of step S4 includes the following steps:

[0149] S4.1. Construct the segmentation decoder SEG_Decoder, which consists of a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.5, a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.1, and a 1×1 convolutional layer stacked in sequence.

[0150] S4.2. Construct a super-resolution decoder SR_Decoder, consisting of a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.5, another 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.1, and a 1×1 convolutional layer stacked sequentially. The difference between SR_Decoder and SEG_Decoder is the difference in the number of input and output channels. Its expression is:

[0151]

[0152]

[0153] in, Represents the intermediate vector of the splicing. Represents splicing, Represents the activation layer. Representatives of the normalization layer, Represents a 1×1 convolutional layer. Represents low-level characteristics. Represents interpolation. This represents the characteristics output by the STD_ASPP module. Representing the output of the Decoder, SEG_Decoder and SR_Decoder differ only in their input and output channels; the formulas are identical. This represents a random deactivation layer with a ratio of 0.5. This represents a 3×3 convolutional layer. The difference between the two decoders is that the input and output channels of the segmentation decoder SEG_Decoder are 304 and 256, 256 and 256, and 256 and the number of categories, respectively, while the input and output channels of the super-resolution decoder SR_Decoder are 304 and 256, 256 and 128, and 128 and 64, respectively. The segmentation decoder outputs the same number of channels as the number of categories, while the super-resolution decoder outputs 64 channels.

[0154] S4.3. The low-level detail features extracted by the backbone network obtained in step S2 are processed by an input preprocessing module consisting of a 1×1 convolutional layer, a batch normalization layer, and an activation layer, and then concatenated with the STD_ASPP fusion features obtained in step S3 after interpolation to obtain the input features of the semantic segmentation head and the super-resolution head.

[0155] S4.4. Construct a semantic segmentation detection head. Input the input features of the semantic segmentation head and the super-resolution head obtained in step S4.3 into SEG_Decoder. After extracting the features, interpolate and resample to output the semantic segmentation result.

[0156] S4.5. Construct a super-resolution detection head. Input the input features of the semantic segmentation head and the super-resolution head obtained in step S4.3 into SR_Decoder. Then, input them into the combined module consisting of the deconvolution module and the EDSR module three times in sequence. Finally, perform a 1×1 convolution to output the super-resolution result.

[0157] The EDSR module consists of a main branch that passes through a 3×3 convolutional layer, a ReLU activation layer, and another 3×3 convolutional layer in sequence, and a residual branch that passes through a 1×1 convolutional layer and is then connected to the main branch.

[0158] S4.6. The semantic segmentation result obtained in step S4.4 is input into CE Loss and Dice Loss to calculate the semantic loss. The super-resolution result obtained in step S4.5 is input into MSE Loss to calculate the super-resolution loss. Then, the semantic loss and the super-resolution loss are input into FA Loss to obtain the feature affinity loss, and the final crop classification result is obtained.

[0159] S5. Result Inference: Collect untrained remote sensing images of crops and input them into the trained crop classification model based on semantic super-resolution edge three-drive, built in steps S2-S4, to perform crop inference prediction and output crop classification results.

[0160] Furthermore, step S5 outputs a raster image with the crop classification results. The value of each pixel in the image represents a preset crop category code, where 0 is the background, 1 is rice, 2 is corn, 3 is wheat, and 4 is soybean.

[0161] Furthermore, the commonly used evaluation indicators and formulas for accuracy evaluation are as follows:

[0162] The formula for the average crossover ratio:

[0163]

[0164] The formula for overall recognition accuracy:

[0165]

[0166] The formula for accuracy:

[0167]

[0168] The formula for recall rate:

[0169]

[0170] Formula for F1 score:

[0171]

[0172] Where k is the total number of pixel categories; true positive (TP) represents the number of pixels predicted as true and actually true; false positive (FP) represents the number of pixels predicted as true and actually false; false negative (FN) represents the number of pixels predicted as false and actually true; and true negative (TN) represents the number of pixels predicted as false and actually false.

[0173] The software and hardware parameters for model training in this embodiment are shown in Table 1:

[0174] Table 1

[0175]

[0176] To facilitate understanding of the technical effects of this embodiment, a comparison of the effects of the method of this embodiment with existing technologies is shown in Table 2. Table 2 provides a quantitative accuracy evaluation of the development model of the patent in this embodiment.

[0177] Table 2

[0178]

[0179] As shown in Table 2, the improved model proposed in this embodiment has high overall accuracy for pixels. In particular, the improved model proposed in this embodiment leads in all indicators, demonstrating superior overall performance.

[0180] A qualitative comparative analysis was conducted using the improved model algorithm and the optimal crop classification model of Deeplabv3+ to observe whether there were oversegmentation, undersegmentation, omissions, or missegmentation in the crop classification results of each model. The comparison figures are attached. Figure 11 From the appendix Figure 11 It can be intuitively observed that the improved model has significantly better performance. In terms of detail recognition, the classification is more refined, the edges are more rounded and smooth, and they are closer to the actual edges, with fewer of the aforementioned poor segmentation phenomena.

[0181] This embodiment has the following advantages compared to other crop classification methods:

[0182] 1. Innovative architecture design for multimodal feature fusion

[0183] This method breaks through the traditional single feature extraction mode and constructs a three-driven collaborative architecture of "super-resolution-semantic segmentation-edge enhancement". By using the Deep_HR backbone network to achieve parallel extraction and complementary fusion of multi-resolution features, it effectively solves the technical challenges commonly found in crop remote sensing images, such as large scale differences and blurred boundaries. Experimental verification shows that this architecture significantly improves the recognition accuracy of small-target crops while maintaining high-resolution details.

[0184] 2. Optimization of dedicated modules based on crop characteristics

[0185] To address the unique characteristics of crop images, this method employs a dedicated STD_ASPP module, innovatively integrating edge enhancement and high-resolution preservation mechanisms into the multi-scale context extraction process. Through the collaborative work of four progressive branches, it achieves comprehensive feature capture from local details to global semantics, making it particularly suitable for the morphological changes and spatial distribution characteristics during crop growth.

[0186] 3. Advantages of the dual-decoder collaborative training strategy

[0187] A dual-task learning mechanism, employing a segmentation decoder and a super-resolution decoder in parallel, is adopted. During the training phase, the network is forced to learn more discriminative feature representations through joint optimization of four loss functions, including feature affinity loss. This design not only improves the model's generalization ability but also effectively alleviates the overfitting problem through the regularization effect between multiple tasks.

[0188] In summary, the crop remote sensing classification method based on a three-driven mechanism of super-resolution, semantic segmentation, and edge enhancement described in this embodiment includes the following steps: First, a multi-category crop sample set is constructed by selecting remote sensing images with key phenological periods and low cloud coverage. Pixel-level fine annotation distinguishes crop types such as rice, corn, wheat, soybeans, and rapeseed. Second, a Deep_HR backbone network integrating the advantages of HRNet and DeepLabv3+ is designed to achieve multi-resolution feature extraction and semantic-detail feature separation, effectively preserving crop boundaries and suppressing the loss of small targets. Furthermore, an improved multi-scale feature fusion module STD_ASPP is proposed, integrating an edge enhancement module EE_Block and a high-resolution preservation module HRK_Block, enhancing local-to-global context awareness through a four-stage progressive structure. Subsequently, a dual-detector architecture is constructed, jointly using a segmentation decoder and a super-resolution decoder, introducing cross-entropy loss, Dice loss, mean squared error loss, and feature affinity loss for collaborative training. Finally, in the inference stage, untrained remote sensing images of the same resolution are used for crop classification, outputting a raster map with category encoding. The super-resolution branch only participates in training and is not used for inference output. This method balances detail preservation, multi-scale semantic understanding, and boundary accuracy in crop identification tasks, making it highly practical and worthy of wider application.

[0189] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0190] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A crop classification method based on semantic super-resolution edge three-driven approach, characterized in that, Includes the following steps: S1. Sample making: Select remote sensing images of crops with low cloud cover and good atmospheric transmittance during the key stages of sowing, growth, maturity or excavation, and make samples of various crops. S2. Feature Extraction: Combining the advantages of HRNet and Deeplabv3+ type backbone networks, a feature extraction backbone network Deep_HR is built to separate high-level semantic features and low-level detail features from the multi-class crop samples in step S1, thereby obtaining the high-level semantic features extracted by the backbone network and the low-level detail features extracted by the backbone network. The specific implementation method of step S2 includes the following steps: S2.

1. Construct the feature extraction backbone network Deep_HR, including the initial feature extraction module Stem, the basic module BasicBlock, the downsampling module, the fusion convolution module, and the low-level feature extraction module; The feature extraction backbone network Deep_HR consists of three branches: Branch 1 is Branch1, Branch 2 is Branch2, and Branch 3 is Branch3. S2.

2. The initial feature extraction module Stem is constructed by sequentially connecting a 3×3 convolution with a stride of 2, a batch normalization layer (BN), an activation layer (ReLU), a 3×3 convolution with a stride of 2, a batch normalization layer (BN), and an activation layer (ReLU). S2.

3. The basic module Basic Block consists of a main branch and a stride-determining residual branch. The main branch is composed of a 3×3 convolutional layer with a given stride, a BN layer, a ReLU layer, and a 3×3 convolutional layer with a stride of 1 connected in sequence. The condition for determining the residual branch by step size is that the step size is not 1 or the input and output channels are not equal. If the condition is not met, the residual branch remains unchanged. If the condition is met, it goes through a 3×3 convolutional layer and a BN layer with a step size of 1 as the output of the step size residual branch. The outputs of the main branch and the residual branch are passed through a ReLU layer and used as the output of the basic module. S2.

4. The constructed downsampling module is divided into downsampling module 1 and downsampling module 2. Downsampling module 1 is applied to the input of branch 2 in the HRNet backbone network, and downsampling module 2 is applied to the input of branch 3 in the HRNet backbone network. Both downsampling module 1 and downsampling module 2 are convolutions with a 3×3 kernel and a stride of 2, stacked with a batch normalization (BN) layer, and then stacked with an activation layer ReLU. S2.

5. The constructed fused convolution is a superposition of a convolutional layer with a 1×1 kernel, a batch normalization layer, and a ReLU layer. The number of input channels is the sum of the three branches of HRNet, and the output is 512. The constructed low-level feature extraction module includes a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation layer; S2.

6. The three branches of the constructed feature extraction backbone network are as follows: Branch 1 is a high-resolution branch, consisting of two layers of Basic Blocks; Branch 2 is a medium-resolution branch, consisting of two layers of stacked Basic Blocks; Branch 3 is a low-resolution branch, consisting of two layers of stacked Basic Blocks. After the initial feature extraction module Stem passes through the multi-type crop samples in step S1, they are directly input into branch 1, directly input into downsampling block 1 and then input into branch 2, and directly input into downsampling block 1, then input into downsampling block 2 and then input into branch 3 respectively. Branches 2 and 3 are upsampled to the size of the feature map of branch 1 and then concatenated with the output of branch 1 to form a fused feature. The fused feature is then passed through a fused convolutional layer to obtain the high-level semantic features extracted by the backbone network. Branch 1 directly outputs the low-level detail features extracted by the backbone network; S3. Module Design: Based on the characteristics of small crop samples, the multi-scale feature fusion module STD_ASPP is designed by combining the edge enhancement module EE_Block and the high resolution preservation module HRK_Block. The high-level semantic features extracted by the backbone network obtained in step S2 are input into the multi-scale feature fusion module STD_ASPP, and the STD_ASPP fused features are output. The specific implementation method of step S3 includes the following steps: S3.

1. Construct an edge enhancement module EE_Block, equipped with edge detection kernels in the x and y directions. The edge detection kernels have sizes of 3×3 and 5×5. Based on the four initialized edge detection kernels, four edge features are obtained. Then, different edge weights are applied, and the corresponding multiplications are summed to normalize the features. Then, the edge enhancement features are obtained through a detail extraction module containing a convolutional layer with a 1×1 convolutional kernel, a batch normalization layer, and an activation layer. S3.

2. Construct the high-resolution preservation module HRK_Block, which concatenates the input features after passing them through four branches and then passes them through a 1×1 convolution, batch normalization layer and activation layer to obtain high-resolution preserved features. Branch 1 includes a 1×1 convolution layer, batch normalization layer and activation layer; Branch 2 includes a 3×3 convolution layer, batch normalization layer and activation layer; Branch 3 includes a 5×5 convolution layer, batch normalization layer and activation layer; Branch 4 includes a 3×3 dilated convolution layer, batch normalization layer and activation layer. S3.

3. Construct a multi-scale feature fusion module STD_ASPP. Taking the high-level semantic features extracted from the backbone network as input, it first passes through a projection layer module consisting of a 1×1 convolutional layer, a batch normalization layer, and an activation layer. The output is the projection layer output features. ; First, input feature layer 1, which includes one HRK_Block, followed by a 3×3 convolution kernel, a dilated convolution with an inflation rate of 1, a batch normalization layer, and an activation layer. The output of feature layer 1 is F1. F1 is then input into feature layer 2, which includes a 3×3 convolution kernel, a dilated convolution with an inflation rate of 2, a batch normalization layer, an activation layer, and an EE_Block, resulting in the output of feature layer 2 as F2. F2 is then input into feature layer 3, which includes a 3×3 convolution kernel, a dilated convolution with an inflation rate of 4, a batch normalization layer, an activation layer, and an EE_Block, resulting in the output of feature layer 3 as F3; F3 is then input into feature layer 4, which includes a 3×3 convolution with an inflation rate of 6, a batch normalization layer, and an activation layer, resulting in the output of feature layer 4 as F4. F1, F2, F3, and F4 are simultaneously input into a fusion layer consisting of a 1×1 convolution, a batch normalization layer, and an activation layer. After being concatenated with the output of the residual connection and the HRK_Block single connection, they are passed through a fusion layer consisting of a 1×1 convolution, a batch normalization layer, and an activation layer again to obtain the STD_ASPP fusion feature. S4. Multi-task head construction: Establish a semantic segmentation head based on the segmentation decoder SEG_Decoder and a super-resolution head based on the super-resolution decoder SR_Decoder. Input the low-level detail features extracted by the backbone network obtained in step S2 and the STD_ASPP fusion features obtained in step S3 into the semantic segmentation head and the super-resolution head at the same time. Combine the cross-entropy loss CELows, Dice coefficient loss Dice Loss, mean squared error loss MSE Loss and feature affinity loss FA Loss for training to obtain the trained crop classification model based on semantic super-resolution edge three-drive, and output the final crop classification result; S5. Result Inference: Collect untrained crop remote sensing images of the same resolution, input them into the trained crop classification model based on semantic super-resolution edge three-drive constructed in steps S2-S4, perform crop inference prediction, and output crop classification results.

2. The crop classification method based on semantic super-resolution edge three-driven method according to claim 1, characterized in that, The specific implementation method for creating multiple crop samples in step S1 includes the following steps: S1.

1. Use annotation tools to draw vector-format polygonal crop surfaces on the acquired remote sensing images of crops; S1.

2. Convert the vector format to raster labels. At the same resolution, the raster labels are mapped one-to-one to the pixels of the crop remote sensing image. S1.

3. Combine the collected crop remote sensing image samples and raster labels into a unified format to obtain the dataset; S1.

4. Apply data augmentation techniques such as rotation, mirroring, scaling, sharpening, contrast adjustment, perspective adjustment, and gamma correction to the dataset obtained in step S1.3 to obtain multiple crop samples.

3. The crop classification method based on semantic super-resolution edge three-driven method according to claim 2, characterized in that, The specific implementation method of step S4 includes the following steps: S4.

1. Construct the segmentation decoder SEG_Decoder, which consists of a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.5, a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.1, and a 1×1 convolutional layer stacked in sequence. S4.

2. Construct a super-resolution decoder SR_Decoder, which consists of a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.5, a 3×3 convolutional layer, a batch normalization layer, an activation layer, a Dropout layer with a dropout probability of 0.1, and a 1×1 convolutional layer stacked in sequence. The difference between SR_Decoder and SEG_Decoder is the difference in the number of input channels and the number of output channels. S4.

3. The low-level detail features extracted by the backbone network obtained in step S2 are processed by an input preprocessing module consisting of a 1×1 convolutional layer, a batch normalization layer, and an activation layer, and then concatenated with the STD_ASPP fusion features obtained in step S3 after interpolation to obtain the input features of the semantic segmentation head and the super-resolution head. S4.

4. Construct a semantic segmentation detection head. Input the input features of the semantic segmentation head and the super-resolution head obtained in step S4.3 into SEG_Decoder. After extracting the features, interpolate and resample to output the semantic segmentation result. S4.

5. Construct a super-resolution detection head. Input the input features of the semantic segmentation head and the super-resolution head obtained in step S4.3 into SR_Decoder. Then, input them into the combined module consisting of the deconvolution module and the EDSR module three times in sequence. Finally, perform a 1×1 convolution to output the super-resolution result. The EDSR module consists of a main branch that passes through a 3×3 convolutional layer, a ReLU activation layer, and another 3×3 convolutional layer in sequence, and a residual branch that passes through a 1×1 convolutional layer and is then connected to the main branch. S4.

6. The semantic segmentation result obtained in step S4.4 is input into CE Loss and Dice Loss to calculate the semantic loss. The super-resolution result obtained in step S4.5 is input into MSE Loss to calculate the super-resolution loss. Then, the semantic features output by SEG_Decoder and the super-resolution features output by SR_Decoder are input into FA Loss to obtain the feature affinity loss, and the final crop classification result is obtained.

4. The crop classification method based on semantic super-resolution edge three-driven method according to claim 3, characterized in that, Step S5 outputs a raster map with the crop classification results, where the value of each cell in the map represents a preset crop category code.

Images