A multi-feature cooperative high-resolution image cultivated crop type classification method

Abstract

The application discloses a kind of multi-feature synergic high-resolution image cultivated land crop type classification method, comprising: three kinds of shallow feature extraction methods including mathematical morphology, extended attribute profile and grey level co-occurrence matrix are applied to each band of remote sensing image, respectively obtain the morphological shallow feature, attribute shallow feature and texture shallow feature of remote sensing image;High-resolution image cultivated land crop classification model is constructed, the morphological shallow feature, attribute shallow feature and texture shallow feature of remote sensing image are input into high-resolution image cultivated land crop classification model and are trained until maximum iteration number is reached, the training of high-resolution image cultivated land crop classification model is completed;Remote sensing image unlabeled region is input into the high-resolution image cultivated land crop classification model trained, and the probability of the type of cultivated land crop to which the cultivated land crop belongs is output, and the type with the maximum probability is used as the cultivated land crop type.Compared with the traditional deep learning model based on single-class feature input, the classification accuracy of the method is high.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing intelligent information extraction technology, specifically, it relates to a method for classifying crop types in high-resolution images of cultivated land using multi-feature synergy. Background Technology

[0002] Arable land refers to land used for growing crops, including: cultivated land, newly developed, reclaimed, and prepared land; fallow land (including rotational fallow land and fallow land); land primarily used for growing crops (including vegetables), interspersed with scattered fruit trees, mulberry trees, or other trees; and reclaimed tidal flats and coastal areas that can guarantee one harvest per year on average. Protecting the red lines of arable land and basic farmland is the foundation and lifeline of agricultural development and modernization, and the cornerstone of national food security. In recent years, due to the continuous improvement of socio-economic levels, vigorous development in the industrial sector, and rapid urbanization, large areas of arable land have been occupied by man-made structures. Therefore, obtaining timely, wide-coverage, and high-precision arable land data is of great significance for arable land protection. With the continuous development of remote sensing technology, the spatial and spectral resolution of remote sensing images is constantly improving, and the obtained ground feature information is becoming richer and more detailed, providing a data foundation for the accurate identification of arable land. Currently, the rapid development of deep learning technologies, mainly based on convolutional neural networks and attention mechanisms, provides new solutions for the classification of crop types in arable land from remote sensing images. Existing research has demonstrated that deep learning methods can meet the needs of ground feature recognition tasks to a certain extent. However, the reliability of deep learning methods heavily relies on a large and accurately labeled training sample library. Furthermore, for remote sensing images with diverse ground feature features and complex spatial relationships, these methods face far greater challenges than those used in natural images.

[0003] Therefore, for the classification of crop types in remote sensing images, existing deep learning-based classification methods have the following limitations: (1) The model suffers from severe overfitting: Due to the time-consuming and laborious sample labeling work of remote sensing images, the quantity and quality of labeled samples are difficult to meet the training requirements of conventional deep models. (2) Land feature features and their spatial relationships are not effectively represented: During the feature extraction process using deep neural networks, due to the lack of prior knowledge, the extracted features cannot accurately express the complex spatial information of land features in the image. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a high-resolution image-based method for classifying crop types in cultivated land using a multi-feature collaborative approach.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a multi-feature collaborative high-resolution image crop type classification method for cultivated land, specifically including the following steps:

[0006] Step 1: Collect remote sensing images and manually label the crop types of cultivated land in a certain area. Apply three shallow feature extraction methods, including mathematical morphology, extended attribute profile, and gray-level co-occurrence matrix, to each band of the remote sensing images to obtain the morphological shallow features, attribute shallow features, and texture shallow features of the remote sensing images respectively.

[0007] Step 2: Construct a high-resolution image crop classification model, which includes a first linear mapping layer, a multi-scale feature extraction module, a channel attention module, a spatial attention module, a second linear mapping layer, a deep feature fusion module, and a multilayer perceptron connected in sequence.

[0008] Step 3: Input the shallow morphological features, shallow attribute features, and shallow texture features of the acquired remote sensing images into the high-resolution image farmland crop classification model for training until the maximum number of iterations is reached, thus completing the training of the high-resolution image farmland crop classification model.

[0009] Step 4: Input the unlabeled areas of the remote sensing image into the trained high-resolution image farmland crop classification model, output the probability of the type of farmland crop, and take the type with the highest probability as the farmland crop type.

[0010] Furthermore, when acquiring the shallow morphological features, the spatial window W is set to 7×7; when extracting the neighborhood features, the area threshold λ is... a Set to 150, diagonal length threshold λ d Set to 150, standard deviation threshold λ s Set to 20; when extracting the texture features, the spatial window W is set to 7×7, and the offset distance (a,b) is set to (2,2).

[0011] Furthermore, both the first linear mapping layer and the second linear mapping layer are 2D convolutional layers with a kernel size of 1×1 and a stride of 1.

[0012] Furthermore, the multi-scale feature extraction module consists of three parallel spatial feature extraction layers, each of which is composed of a 2D convolutional layer, GroupNorm normalization, and ReLU activation function; wherein the kernel sizes of the three 2D convolutional layers are 3×3, 5×5, and 7×7, respectively, and the stride is 1.

[0013] Furthermore, the deep feature fusion module includes: a first feature layer, a second feature layer, a third feature layer, a fourth feature layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer. The first feature layer is connected to the first downsampling layer and the second feature layer, and the first downsampling layer is connected to the third feature layer. The second feature layer is connected to the second downsampling layer, and the second downsampling layer is connected to the fourth feature layer. The third feature layer is connected to the third downsampling layer, and the fourth feature layer is connected to the fourth downsampling layer. The third downsampling layer and the fourth downsampling layer are connected.

[0014] Furthermore, the first, second, third, and fourth feature layers are all composed of 2D convolutional layers, GroupNorm normalization, and ReLU activation functions. The first and second downsampling layers are both Avg-Pooling layers, and the third and fourth downsampling layers are both Max-Pooling layers. The 2D convolutional layers of the first and fourth feature layers have a kernel size of 3×3 and a stride of 1. The 2D convolutional layers of the second and third feature layers have a kernel size of 5×5 and a stride of 1. The pooling kernels of the first, second, third, and fourth downsampling layers have a kernel size of 2×2 and a stride of 2.

[0015] Furthermore, the multilayer perceptron module consists of multiple fully connected layers and ReLU activation functions. The number of neurons in the last fully connected layer is equal to the number of crop categories in the cultivated land. The output of the last layer is passed through a Softmax function, and the output satisfies a probability distribution with a sum of 1.

[0016] Furthermore, step 3 includes the following sub-steps:

[0017] Step 3.1: Linearly map the shallow morphological features, shallow attribute features, and shallow texture features of the image obtained in Step 1 to the same channel dimension through a linear mapping layer;

[0018] Step 3.2: Extract multi-scale spatial features from each type of shallow feature after linear mapping in Step 3.1 using the multi-scale feature extraction module, and then superimpose the extracted multi-scale spatial features with each type of shallow feature after linear mapping according to the channel dimension to obtain high-dimensional features.

[0019] Step 3.3: First, the high-dimensional features are weighted in the channel dimension by the channel attention module to capture the channels of key information, and then weighted in the spatial dimension by the spatial attention module to capture the spatial location of key information, thus obtaining optimized features;

[0020] Step 3.4: Input the optimized features into the second linear mapping layer for linear mapping, mapping the optimized features to a low-dimensional space to obtain low-dimensional features;

[0021] Step 3.5: Input the low-dimensional features into the deep feature fusion module for fusion to obtain semantic features;

[0022] Step 3.6: Expand the semantic features into a one-dimensional vector, input it into the multilayer perceptron, and obtain the type probability of the cultivated land crop. Take the type with the highest probability as the cultivated land crop type.

[0023] Step 3.7: Repeat steps 3.1-3.6 until the maximum number of iterations is reached to complete the training of the high-resolution image crop classification model.

[0024] Furthermore, the parameters set during the training process of the high-resolution image-based crop classification model include: the optimizer is set to Adam, the batch size is set to 32, the maximum number of iterations is set to 100, and the learning rate (lr) is set to 10. -4 The loss function is CrossEntropy Loss.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] (1) This invention proposes a multi-feature collaborative high-resolution image crop type classification method. This method comprehensively utilizes the potential morphological features, neighborhood features, and texture features in the image to effectively represent different land cover types, thereby reducing the difficulty of feature extraction within the network model. In addition, this method only requires a small number of pixel-level training samples to achieve good classification accuracy, which greatly reduces the workload of remote sensing image sample labeling;

[0027] (2) The present invention designs a multi-scale feature extraction module and a deep feature fusion module. The multi-scale feature extraction module extracts multi-scale spatial information from shallow features in a parallel manner to deal with the spatial relationship of land features at different depths. The deep feature fusion module performs deep fusion on the obtained multi-scale features to obtain semantic features with more discriminative significance.

[0028] This invention has achieved good classification results in large-scale farmland image datasets. Compared with traditional deep learning models based on single-class feature input, the method of this invention has high overall accuracy, low requirement for labeled samples, and high running efficiency, providing a new solution to the problem of farmland crop classification based on remote sensing images. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the overall process of the high-resolution image crop type classification method for multi-feature synergy according to the present invention.

[0030] Figure 2 This is a schematic diagram of the multi-scale feature extraction module in this invention;

[0031] Figure 3 This is a schematic diagram of the deep feature fusion module in this invention;

[0032] Figure 4 This is a classification map of crop types for cultivated land using the method of the present invention. Figure 4 (a) in the image is a composite image of cultivated land. Figure 4 (b) in the text is the truth label. Figure 4 (c) in the table represents the classification result. Detailed Implementation

[0033] The technical solution of the present invention will be further explained and described below with reference to the accompanying drawings and embodiments.

[0034] like Figure 1 This is a flowchart illustrating the overall process of the high-resolution image crop type classification method based on multi-feature collaboration of the present invention. The method specifically includes the following steps:

[0035] Step 1: Collect remote sensing images and manually label the crop types of cultivated land in a certain area. Apply three shallow feature extraction methods—mathematical morphology, extended attribute profile, and gray-level co-occurrence matrix—to each band of the remote sensing images to obtain morphological shallow features, attribute shallow features, and texture shallow features. By effectively representing different land cover types, the difficulty of feature extraction within the network model is reduced. In this invention, the morphological shallow features extracted by mathematical morphology include: opening, closing, top-hat opening, top-hat closing, reconstruction opening, and reconstruction closing. The attribute shallow features extracted by extended attribute profile include: area of ​​connected regions, diagonal length of the bounding rectangle of connected regions, and standard deviation of pixel gray values ​​within connected regions. The texture shallow features extracted by gray-level co-occurrence matrix include: mean, entropy, variance, second moment of angle, and contrast. When acquiring morphological shallow features, the spatial window W is set to 7×7; when extracting neighborhood features, the area threshold λ is... a Set to 150, diagonal length threshold λ d Set to 150, standard deviation threshold λ s Set to 20; when extracting the texture features, the spatial window W is set to 7×7, and the offset distance (a,b) is set to (2,2).

[0036] Step 2: Construct a high-resolution image-based crop classification model. This model comprises a first linear mapping layer, a multi-scale feature extraction module, a channel attention module, a spatial attention module, a second linear mapping layer, a deep feature fusion module, and a multilayer perceptron, all connected sequentially. This high-resolution image-based crop classification model can fully utilize the synergy between features of different types and scales to improve the performance of deep neural networks in high-resolution remote sensing image classification. In this invention, the first and second linear mapping layers are both 2D convolutional layers with a kernel size of 1×1 and a stride of 1; for example... Figure 2 The multi-scale feature extraction module consists of three parallel spatial feature extraction layers. It extracts multi-scale spatial information from shallow features in a parallel manner to address spatial relationships of ground features at different depths. Each spatial feature extraction layer corresponds to a different spatial scale and consists of a 2D convolutional layer, GroupNorm normalization, and ReLU activation function. The kernel sizes of the three 2D convolutional layers are 3×3, 5×5, and 7×7, with a stride of 1 and zero padding, ensuring that the output features have the same height and width and the same number of channels as the input features. Figure 3 The deep feature fusion module includes: a first feature layer, a second feature layer, a third feature layer, a fourth feature layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer. The first feature layer is connected to the first downsampling layer and the second feature layer, and the first downsampling layer is connected to the third feature layer; the second feature layer is connected to the second downsampling layer, and the second downsampling layer is connected to the fourth feature layer; the third feature layer is connected to the third downsampling layer, and the fourth feature layer is connected to the fourth downsampling layer; the third downsampling layer and the fourth downsampling layer are also connected. The first, second, third, and fourth feature layers are all composed of 2D convolutional layers, GroupNorm normalization, and ReLU activation functions. The first downsampling layer, the second… All downsampling layers are Avg-Pooling layers, while the third and fourth downsampling layers are Max-Pooling layers. The 2D convolutional layers of the second and third feature layers have 5×5 kernels with a stride of 1, while the 2D convolutional layers of the first and fourth feature layers have 3×3 kernels with a stride of 1, padded with zeros, and the number of channels doubles after passing through the feature layers. The pooling kernels of the first, second, third, and fourth downsampling layers are 2×2 with a stride of 2, meaning that the height and width are halved after passing through the downsampling layers. A deep feature fusion module performs deep fusion on the obtained multi-scale features to obtain more discriminative semantic features. The multilayer perceptron module consists of multiple fully connected layers and ReLU activation functions. The number of neurons in the last fully connected layer equals the number of crop categories, and the output of the last layer is passed through a Softmax function, with the output satisfying a probability distribution that sums to 1.

[0037] Step 3: Input the shallow morphological features, shallow attribute features, and shallow texture features of the acquired remote sensing images into the high-resolution image farmland crop classification model for training until the maximum number of iterations is reached, thus completing the training of the high-resolution image farmland crop classification model; specifically, this includes the following sub-steps:

[0038] Step 3.1: Linearly map the shallow morphological features, shallow attribute features, and shallow texture features of the image obtained in Step 1 to the same channel dimension through the first linear mapping layer;

[0039] Step 3.2: Extract multi-scale spatial features from each type of shallow feature after linear mapping in Step 3.1 using the multi-scale feature extraction module, and then superimpose the extracted multi-scale spatial features with each type of shallow feature after linear mapping according to the channel dimension to obtain high-dimensional features.

[0040] Step 3.3: First, the high-dimensional features are weighted in the channel dimension by the channel attention module to capture the channels of key information, and then weighted in the spatial dimension by the spatial attention module to capture the spatial location of key information, thereby suppressing the interference information in the high-dimensional features and obtaining optimized features.

[0041] Step 3.4: Input the optimized features into the second linear mapping layer for linear mapping, mapping the optimized features to a low-dimensional space to obtain low-dimensional features;

[0042] Step 3.5: Input the low-dimensional features into the deep feature fusion module for fusion. Different layers in the deep feature fusion module are associated through skip connections. Finally, the extracted features are fused through residual connections to obtain semantic features.

[0043] Step 3.6: Expand the semantic features into a one-dimensional vector, input it into the multilayer perceptron, and obtain the type probability of the cultivated land crop. Take the type with the highest probability as the cultivated land crop type.

[0044] Step 3.7: Repeat steps 3.1-3.6 until the maximum number of iterations is reached to complete the training of the high-resolution image crop classification model.

[0045] The parameters set during the training of the high-resolution image crop classification model included: optimizer set to Adam, batch size set to 32, maximum number of iterations set to 100, and learning rate set to 10. -4 The loss function is CrossEntropy Loss.

[0046] Step 4: Input the unlabeled areas of the remote sensing image into the trained high-resolution image farmland crop classification model, output the probability of the type of farmland crop, and take the type with the highest probability as the farmland crop type.

[0047] The multi-feature collaborative high-resolution image crop type classification method of this invention comprehensively considers the synergistic effect between multi-type shallow feature extraction, multi-scale spatial information superposition, high-dimensional feature optimization and deep feature fusion. It has good generalization and robustness, and the accuracy of crop type classification is high with a small number of labeled samples.

[0048] Example

[0049] The dataset used in this embodiment is an aerial remote sensing image dataset of Matiwan Village in Xiong'an New Area. Its spatial resolution is 0.5 meters, and the image size is 3750×1580, containing 20 different types of land cover, including rice paddies, vegetable fields, corn, and soybeans. The dataset itself is a hyperspectral remote sensing image containing 256 spectral bands. In the specific implementation process, only data from three bands (R:120, G:72, B:36, where the numbers represent band numbers) are used.

[0050] The above remote sensing images are used to obtain the crop type classification method of cultivated land crop type in the high-resolution image of the present invention through multi-feature synergy. The specific process is as follows:

[0051] Step 1: Using the band clipping function of the remote sensing processing software ENVI, select three bands (R:120, G:72, B:36) corresponding to the remote sensing image and save them as separate TIFF files. Use Matlab programming to extract morphological shallow features, neighborhood shallow features, and texture shallow features. First, extract morphological shallow features using mathematical morphological operations including opening, closing, top-hat opening, top-hat closing, reconstruction opening, and reconstruction closing, obtaining a total of 18 features from the three bands. Next, extract neighborhood shallow features using attribute profiling methods including connected region area, diagonal length of the bounding rectangle of the connected region, and standard deviation of pixel gray values ​​within the connected region. Each attribute profiling method includes thin and thick operations, obtaining a total of 18 features from the three bands. Finally, texture shallow features calculated based on the gray-scale matrix method include mean, entropy, variance, second moment of angle, and contrast, obtaining a total of 15 features from the three bands. The parameter settings for the above three shallow feature acquisition methods are shown in Table 1:

[0052] Table 1: Parameter settings for three types of shallow feature acquisition

[0053] Feature Name Parameter settings Morphological features Window size W = 7×7, window shape is rectangular Shallow features of the neighborhood <![CDATA[Area threshold λ a = 150, diagonal threshold λ d = 150, standard deviation threshold λ s = 20]]> shallow texture features The window size W = 7×7, and the offset distance (a,b) is set to (2,2).

[0054] Step 2: Use the PyTorch deep learning framework to build a high-resolution image crop classification model for farmland; the parameter settings in the multi-scale feature extraction module are shown in Table 2:

[0055] Table 2: Parameter settings in the multi-scale feature extraction module

[0056] Multiscale nuclear size standardization Activation function Step length Fill value Parallelism Output Spatial Scale -1 16×3×3 GN ReLU 1 0 ×3 16×32×32 Spatial Scale -2 16×5×5 GN ReLU 1 0 ×3 16×32×32 Spatial Scale -3 16×7×7 GN ReLU 1 0 ×3 16×32×32

[0057] The parameter settings in the deep feature fusion module are shown in Table 3:

[0058] Table 3: Parameter Settings in the Deep Feature Fusion Module

[0059] Layer name nuclear size standardization Activation function Step length Fill value Way Output Feature layer 1 48×3×3 GN ReLU 1 0 \ 96×32×32 Downsampling 1 2×2 \ \ 2 \ Avg 96×16×16 Feature layer 2 48×5×5 GN ReLU 1 0 \ 96×32×32 Downsampling 2 2×2 \ \ 2 \ Avg 96×16×16 Feature layer 3 96×5×5 GN ReLU 1 0 \ 192×16×16 Downsampling 3 2×2 \ \ 2 \ Max 192×8×8 Feature layer 4 96×3×3 GN ReLU 1 0 \ 192×16×16 Downsampling 4 2×2 \ \ 2 \ Max 192×8×8

[0060] Step 3: Set the parameters for the high-resolution image crop classification model during training, including: optimizer set to Adam, batch size set to 32, maximum number of iterations set to 100, and learning rate (lr) set to 10. -4 The loss function is CrossEntropy Loss; the shallow morphological features, shallow attribute features, and shallow texture features of the acquired remote sensing images are input into the high-resolution image cultivated land crop classification model for training, specifically including the following sub-steps:

[0061] Step 3.1: Manually label the cultivated crops in a certain area of ​​the remote sensing image. Centered on the labeled sample, linearly map the shallow morphological features, shallow attribute features, and shallow texture features to the same channel dimension through a linear mapping layer according to the image size of 32×32. After linear mapping, the number of channels for the three types of shallow features is unified to 16.

[0062] Step 3.2: Extract multi-scale spatial features from each type of shallow feature after linear mapping in Step 3.1 using the multi-scale feature extraction module, and then superimpose the extracted multi-scale spatial features with each type of shallow feature after linear mapping according to the channel dimension to obtain a high-dimensional feature of size 192×32×32.

[0063] Step 3.3: The high-dimensional features are first weighted along the channel dimension by a channel attention module to capture key information channels. Specifically, global average pooling and global max pooling are used to compress the spatial dimension of the input high-dimensional features, outputting compressed features of size 2×192. The compressed features are then passed through two fully connected layers with shared parameters and ReLU activation, and then added together to output key features of size 1×192. The first fully connected layer has 192 input and 12 output neurons, and the second fully connected layer has 12 input and 192 output neurons. After passing through S... The igmoid activation function is used, and multiplication is performed with the multi-scale features according to the channel dimension. Then, the spatial attention module performs spatial weighting to capture the spatial location of key information and obtain optimized features. Specifically, global average pooling and global max pooling are used to compress the channel dimension of the input key features, and the output result is 2×32×32. The output result is passed through a 2D convolutional layer with a kernel size of 7×7 and a stride of 1, and the output result is 1×32×32. The sigmoid function is used for activation, and multiplication is performed with the multi-scale features according to the spatial dimension to obtain optimized features.

[0064] Step 3.4: Input the optimized features into the second linear mapping layer for linear mapping, and map the optimized features to a low-dimensional space to obtain a low-dimensional feature of size 48×32×32. The output channel number of the low-dimensional feature is 48.

[0065] Step 3.5: Input the low-dimensional features into the deep feature fusion module for fusion to obtain semantic features. Specifically, firstly, the results of the first feature layer are used as inputs to the second feature layer and the first downsampling layer, respectively. Secondly, the results of the second feature layer will pass through the second downsampling layer, the fourth feature layer, and the fourth downsampling layer in sequence, and the results of the first downsampling layer will pass through the third feature layer and the third downsampling layer in sequence. Finally, the output features of the third downsampling layer and the output features of the fourth downsampling layer are added together to output a semantic feature of size 192×8×8.

[0066] Step 3.6: Unfold the semantic features into a one-dimensional vector and input it into a multilayer perceptron to obtain the probability of crop types. The type with the highest probability is selected as the crop type. In this embodiment, the multilayer perceptron includes four fully connected layers. The output of each layer is activated using the ReLU function. The first three fully connected layers have 512, 256, and 128 output neurons, respectively. The drop rate of the last fully connected layer is set to 50%, and the number of output neurons corresponds to the number of categories. In this embodiment, the number of categories is 20. The output of the multilayer perceptron module is then processed by a Softmax function to ensure a probability distribution that sums to 1.

[0067] Step 3.7: Repeat steps 3.1-3.6 until the maximum number of iterations is reached to complete the training of the high-resolution image crop classification model.

[0068] Step 4: Input the unlabeled areas of the remote sensing image into the trained high-resolution image farmland crop classification model, output the probability of the type of farmland crop, and take the type with the highest probability as the farmland crop type.

[0069] like Figure 4 (a) in the image is a composite image of cultivated land. Figure 4 (b) in the text is the truth label. Figure 4 (c) in the figure represents the classification result. Compared with the true label, the accuracy of the classification result reaches 99.08%. Compared with the traditional deep learning model based on single-class feature input, the method of this invention has high overall accuracy, small requirement for labeled samples, and high running efficiency. It can be widely applied to the classification of cultivated land crops based on remote sensing images.

[0070] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A multi-feature collaborative high-resolution image crop type classification method for cultivated land, characterized in that, Specifically, the following steps are included: Step 1: Collect remote sensing images and manually label the crop types of cultivated land in a certain area. Apply three shallow feature extraction methods, including mathematical morphology, extended attribute profile, and gray-level co-occurrence matrix, to each band of the remote sensing images to obtain the morphological shallow features, attribute shallow features, and texture shallow features of the remote sensing images respectively. Step 2: Construct a high-resolution image crop classification model, which includes a first linear mapping layer, a multi-scale feature extraction module, a channel attention module, a spatial attention module, a second linear mapping layer, a deep feature fusion module, and a multilayer perceptron connected in sequence. Step 3: Input the shallow morphological features, shallow attribute features, and shallow texture features of the acquired remote sensing images into the high-resolution image farmland crop classification model for training until the maximum number of iterations is reached, thus completing the training of the high-resolution image farmland crop classification model. Step 4: Input the unlabeled areas of the remote sensing image into the trained high-resolution image farmland crop classification model, output the probability of the type of farmland crop, and take the type with the highest probability as the farmland crop type.

2. The method for classifying crop types in high-resolution imagery using multiple features in accordance with claim 1, characterized in that, When acquiring the shallow morphological features, the spatial window... W Set to 7×7; when extracting shallow features of the attribute, the area threshold is... λ a Set to 150, diagonal length threshold λ d Set to 150, standard deviation threshold λ s Set to 20; during the extraction of shallow texture features, the spatial window... W Set to 7×7, offset distance ( a , b Set it to (2, 2).

3. The method for classifying crop types in high-resolution imagery using multiple features in accordance with claim 1, characterized in that, Both the first linear mapping layer and the second linear mapping layer are 2D convolutional layers with a kernel size of 1×1 and a stride of 1.

4. The method for classifying crop types in high-resolution images of cultivated land using multi-feature synergy as described in claim 1, characterized in that, The multi-scale feature extraction module consists of three parallel spatial feature extraction layers. Each spatial feature extraction layer consists of a 2D convolutional layer, GroupNorm normalization, and ReLU activation function. The kernel sizes of the three 2D convolutional layers are 3×3, 5×5, and 7×7, respectively, and the stride is 1.

5. The method for classifying crop types in high-resolution imagery using multiple features in accordance with claim 1, characterized in that, The deep feature fusion module includes: a first feature layer, a second feature layer, a third feature layer, a fourth feature layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer. The first feature layer is connected to the first downsampling layer and the second feature layer, and the first downsampling layer is connected to the third feature layer. The second feature layer is connected to the second downsampling layer, and the second downsampling layer is connected to the fourth feature layer. The third feature layer is connected to the third downsampling layer, and the fourth feature layer is connected to the fourth downsampling layer. The third downsampling layer and the fourth downsampling layer are connected.

6. The method for classifying crop types in high-resolution images of cultivated land using multi-feature synergy as described in claim 5, characterized in that, The first, second, third, and fourth feature layers are all composed of 2D convolutional layers, GroupNorm normalization, and ReLU activation functions. The first and second downsampling layers are both Avg-Pooling layers, and the third and fourth downsampling layers are both Max-Pooling layers. The 2D convolutional layers of the first and fourth feature layers have a kernel size of 3×3 and a stride of 1. The 2D convolutional layers of the second and third feature layers have a kernel size of 5×5 and a stride of 1. The pooling kernels of the first, second, third, and fourth downsampling layers have a kernel size of 2×2 and a stride of 2.

7. The method for classifying crop types in high-resolution images of cultivated land using multi-feature synergy as described in claim 1, characterized in that, The multilayer perceptron module consists of multiple fully connected layers and ReLU activation functions. The number of neurons in the last fully connected layer is equal to the number of crop categories. The output of the last layer is passed through the Softmax function, and the output satisfies a probability distribution that sums to 1.

8. The method for classifying crop types in high-resolution images of cultivated land using multi-feature synergy as described in claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Linearly map the shallow morphological features, shallow attribute features, and shallow texture features of the image obtained in Step 1 to the same channel dimension through the first linear mapping layer; Step 3.2: Extract multi-scale spatial features from each type of shallow feature after linear mapping in Step 3.1 using the multi-scale feature extraction module, and then superimpose the extracted multi-scale spatial features with each type of shallow feature after linear mapping according to the channel dimension to obtain high-dimensional features. Step 3.3: First, the high-dimensional features are weighted in the channel dimension by the channel attention module to capture the channels of key information, and then weighted in the spatial dimension by the spatial attention module to capture the spatial location of key information, thus obtaining optimized features; Step 3.4: Input the optimized features into the second linear mapping layer for linear mapping, mapping the optimized features to a low-dimensional space to obtain low-dimensional features; Step 3.5: Input the low-dimensional features into the deep feature fusion module for fusion to obtain semantic features; Step 3.6: Expand the semantic features into a one-dimensional vector, input it into the multilayer perceptron, and obtain the type probability of the cultivated land crop. Take the type with the highest probability as the cultivated land crop type. Step 3.7: Repeat steps 3.1-3.6 until the maximum number of iterations is reached to complete the training of the high-resolution image crop classification model.

9. A method for classifying crop types in high-resolution images using multi-feature synergy, as described in claim 1 or 8, characterized in that, The parameters set during the training process of the high-resolution image-based crop classification model include: optimizer set to Adam, batch size set to 32, maximum number of iterations set to 100, and learning rate set to 10. -4 The loss function is CrossEntropy Loss.

Images