Land use classification method and system based on data enhancement

Abstract

The application discloses a land use classification method and system based on data enhancement, and relates to the field of land classification.The method comprises the following steps: acquiring a target ground object image; inputting the target ground object image into a land use classification model to obtain the ground object type of the target ground object image; the land use classification model is obtained by training a semantic segmentation network by using a data enhanced ground object set; the data enhanced ground object set comprises a GID data set, a ground object composite image and a ground object type label image corresponding to the ground object composite image; the ground object composite image is obtained by inputting the GID data set into a GauGAN network; and the semantic segmentation network is a DeepLabV3+ network fused with a Dropout layer.The application can improve the precision of land use classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of land classification, and particularly relates to a land use classification method and system based on data enhancement. BACKGROUND

[0002] At present, there are problems of data imbalance and relatively small sample quantity in land use classification, and if it is used for classification, the ground object classification result is poor, which affects the improvement of land use classification accuracy to a certain extent. SUMMARY

[0003] Therefore, the embodiment of the present application provides a land use classification method and system based on data enhancement to improve the accuracy of land use classification.

[0004] To achieve the above object, the present application provides the following scheme:

[0005] A land use classification method based on data enhancement comprises:

[0006] obtaining a target ground object image;

[0007] inputting the target ground object image into a land use classification model to obtain a ground object type of the target ground object image;

[0008] The land use classification model is obtained by training a semantic segmentation network using a data enhanced ground object set.

[0009] The data enhanced ground object set comprises a GID data set, a ground object synthesis image and a ground object type label image corresponding to the ground object synthesis image; the ground object synthesis image is obtained by inputting the GID data set into a GauGAN network; and the semantic segmentation network is a DeepLab V3+ network fused with a Dropout layer.

[0010] Optionally, the GauGAN network comprises an image encoding module, a generator and a discriminator connected in sequence.

[0011] The image encoding module is configured to encode a real image in the GID data set and a ground object type label image corresponding to the real image according to the input to obtain an encoding result.

[0012] The generator is a full convolutional network composed of a spatial adaptive normalization residual block; and the generator is configured to synthesize a ground object synthesis image according to the input of the encoding result and a target ground object type label image.

[0013] The discriminator is configured to determine whether the ground object type of the ground object synthesis image is consistent with the target ground object type label image.

[0014] Optionally, the semantic segmentation network comprises an encoder and a decoder connected in sequence.

[0015] The encoder is configured to perform feature extraction on the input image, regularize multi-scale features obtained by the feature extraction using a Dropout layer, and obtain encoded features.

[0016] The decoder is configured to perform up-sampling on the encoded features, regularize the up-sampling result using a Dropout layer, obtain decoded features, and predict the feature type of the input image using the decoded features.

[0017] Optionally, the encoder comprises a deep convolutional neural network, a dilated spatial pyramid pooling module, a first feature fusion layer, a first Dropout layer and a first convolutional layer connected in sequence.

[0018] The deep convolutional neural network is configured to perform feature extraction on the input image to obtain first features and second features, wherein the size of the first features is smaller than the size of the second features.

[0019] The dilated spatial pyramid pooling module comprises a dilated convolution module and a pyramid pooling module connected in parallel, wherein the dilated convolution module is configured to perform dilated convolution on the second features at different dilation rates to output local features of different scales, and the pyramid pooling module is configured to perform global average pooling on the second features to output global features.

[0020] The first feature fusion layer is configured to fuse the local features of different scales and the global features to obtain multi-scale features.

[0021] The first Dropout layer is configured to regularize the multi-scale features.

[0022] The first convolutional layer is configured to adjust the dimension of the features regularized by the first Dropout layer to obtain encoded features.

[0023] Optionally, the dilated convolution module comprises a second convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a third dilated convolutional layer and a fourth dilated convolutional layer connected in parallel.

[0024] The first dilated convolutional layer has a dilation rate of 3, the second dilated convolutional layer has a dilation rate of 6, the third dilated convolutional layer has a dilation rate of 12, and the fourth dilated convolutional layer has a dilation rate of 18.

[0025] Optionally, the decoder comprises:

[0026] a first up-sampling layer, a third convolutional layer, a second feature fusion layer, a second Dropout layer, a fourth convolutional layer, a third Dropout layer and a second up-sampling layer;

[0027] The first up-sampling layer is configured to up-sample the encoded features output by the first convolutional layer;

[0028] The third convolutional layer is configured to adjust the dimensions of the first features output by the deep convolutional neural network;

[0029] The second feature fusion layer is configured to fuse the features output by the first up-sampling layer with the features output by the third convolutional layer to obtain deep fusion features;

[0030] The second Dropout layer is configured to perform regularization processing on the deep fusion features;

[0031] The fourth convolutional layer is configured to adjust the dimensions of the features after the regularization processing of the second Dropout layer;

[0032] The third Dropout layer is configured to perform regularization processing on the features output by the fourth convolutional layer;

[0033] The second up-sampling layer is configured to up-sample the features after the regularization processing of the third Dropout layer to obtain decoded features, and predict the feature type of the input image using the decoded features.

[0034] Optionally, the first up-sampling layer and the second up-sampling layer are both four times up-sampling layers.

[0035] Optionally, the sizes of the first convolutional layer, the second convolutional layer and the third convolutional layer are all 1x1, and the size of the fourth convolutional layer is 3x3.

[0036] The application further provides a land use classification system based on data enhancement, comprising:

[0037] an image acquisition module configured to acquire a target feature image;

[0038] a land use classification module configured to input the target feature image into a land use classification model to obtain the feature type of the target feature image;

[0039] wherein the land use classification model is obtained by training a semantic segmentation network using a data enhanced feature set;

[0040] The data enhanced ground object set comprises a GID data set, a ground object composite image and a ground object type label image corresponding to the ground object composite image; the ground object composite image is obtained by inputting the GID data set into a GauGAN network; and the semantic segmentation network is a DeepLab V3+ network fused with a Dropout layer.

[0041] Compared with the prior art, the present application has the following advantages:

[0042] The embodiment of the present application provides a land use classification method and system based on data enhancement, generates a ground object composite image with rich semantics and realistic effect based on a GauGAN network, and uses a data enhanced ground object set composed of a GID data set and the ground object composite image as training data, thereby increasing the diversity of the training data and maintaining the balance between different ground classes of the training data, so that the land use classification model obtained by training the semantic segmentation network based on the data enhanced ground object set is used for classification, and the precision of land use classification is improved; and the semantic segmentation network adopts a DeepLab V3+ network fused with a Dropout layer, so that the generalization ability of the land use classification model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0044] Figure 1 A flowchart of the land use classification method based on data enhancement provided by the embodiment of the present application is shown in the figure.

[0045] Figure 2 A schematic diagram of the GID data set with a small sample is shown in the figure.

[0046] Figure 3 A schematic diagram of the ground class sample proportion of the GID data set is shown in the figure.

[0047] Figure 4 A schematic diagram of the pix2pixhd network is shown in the figure.

[0048] Figure 5 A schematic diagram of the GauGAN network is shown in the figure.

[0049] Figure 6 A schematic diagram of the SPADE structure is shown in the figure.

[0050] Figure 7 A schematic diagram of the structure of the semantic segmentation network is shown in the figure.

[0051] Figure 8A structural schematic diagram of a hollow space pyramid pooling module;

[0052] Figure 9 A schematic diagram of a specific implementation process of a land use classification method. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0054] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0055] Figure 1 A flowchart of a land use classification method based on data enhancement provided by the embodiments of the present application.

[0056] Referring to Figure 1 , the land use classification method based on data enhancement comprises:

[0057] Step 101: acquiring a target ground object image.

[0058] Step 102: inputting the target ground object image into a land use classification model to obtain a ground object type of the target ground object image; wherein the land use classification model is obtained by training a semantic segmentation network using a data enhanced ground object set; the data enhanced ground object set comprises a GID data set, a ground object synthetic image and a ground object type label image corresponding to the ground object synthetic image; the ground object synthetic image is obtained by inputting the GID data set into a GauGAN network; and the semantic segmentation network is a DeepLab V3+ network fused with a Dropout layer.

[0059] First, the data enhanced ground object set is introduced:

[0060] The original GID data set is a high-resolution land use data set based on high-resolution remote sensing satellite images of Gaofen 2, and the collection time is from December 2014 to October 2016, the collection range is more than 60 different cities in China, and the geographical range is more than 50000km 2 , containing more than 150 remote sensing images, the size of a single image is 6800*7200 pixels, the spatial resolution is 1m, and there are RGB format and NIR+RGB format.

[0061] The embodiment data enhances the GID dataset in the ground object collection with a small sample. The dataset is obtained by cropping the original GID dataset to a size of 224*224, a total of 1391, as shown in Figure 2 , Figure 2 The first row in the table represents the cropped real image, and the second row represents the ground object type label image corresponding to the real image. The ground object type label image represents the land use category, which is based on the Chinese Land Use Classification Criteria (GB / T 21010-2017) and contains five types of ground objects: built-up, farmland, meadow, water, and forest. The RGB composition is (255, 0, 0), (0, 255, 0), (0, 255, 255), (255, 255, 0), and (0, 0, 255), respectively, representing five different colors: red, green, cyan, yellow, and blue. As shown in Figure 3 , the GID dataset has a significant data imbalance problem, and meadow and forest ground objects belong to the few-sample ground object type.

[0062] Therefore, to perform data augmentation, a semantic-rich and realistic ground object synthetic image is generated based on the GID dataset. The image data augmentation ground object collection is formed by the GID dataset, the ground object synthetic image, and the ground object type label corresponding to the ground object synthetic image, which is used for subsequent model training to improve classification accuracy.

[0063] The GauGAN network for generating ground object synthetic images is introduced as follows:

[0064] The GauGAN network is a generative adversarial network modified based on the pix2pixhd network, which can generate high-fidelity images given semantic labels.

[0065] The pix2pixhd network is shown in Figure 4 , which is based on the CGAN framework and is mainly used for image-to-image translation to generate high-resolution images. The pix2pixhd network includes a generator and a discriminator. The generator is composed of a residual network, which maps the label onto the real image to synthesize a new image. The discriminator is multi-scale, mainly used to distinguish between real images and synthetic images. The loss function consists of GAN loss, feature matching loss, and content loss. Figure 4 In the table, G1 and G2 are the generators of the pix2pixhd network; D1, D2, and D3 together form a multi-scale discriminator.

[0066] GauGAN network as Figure 5 shown, the GauGAN network comprises an image encoding module (ImageEncoder) 1, a generator (Generator) 2 and a discriminator (Discriminator) 3 connected in sequence, as shown in Figure 5 part (a). The Generator is a fully convolutional network composed of Spatially-Adaptive Normalization Residual Blocks (SPADE ResBlk), as shown in Figure 5 part (b), the input is a normalized feature map, then the up-sampling is completed by calculating in the SPADE ResBlk, and finally the image is generated, and the semantic information is enhanced by using the labeled data in the generation process.

[0067] The SPADE normalization method is the main difference between GauGAN and pix2pixhd. Pix2pixhd uses an unconditional normalization method, which will cause semantic information loss. SPADE modifies batch normalization by setting different parameters for all pixels in the feature map to achieve better normalization, rather than setting parameters in the channel of the feature map for normalization.

[0068] The SPADE structure is shown in Figure 6 SPADE is not directly normalized to the feature map, but first adjusts it through a spatially adaptive mechanism and only normalizes the previous layer. The purpose is to learn the semantic information at different positions on the feature map, better protect the semantic information from the influence of common normalization layers, and enhance the discrimination of GauGAN for different semantic information. Referring to Figure 6 In SPADE, the image label is first projected into the embedding space, and then convolved to produce modulation parameters γ and β. Unlike previous conditional normalization methods, γ and β are not vectors, but tensors with spatial dimensions. The resulting γ and β are multiplied and added to the normalized activation elements.

[0069] Specifically, the image encoding module is used to encode the real image in the input GID dataset and the real image corresponding to the ground object type label image to obtain an encoding result, and the encoding result is used to guide the generator to generate an image with a certain style. The generator is a fully convolutional network composed of spatially adaptive normalization residual blocks; the generator is used to synthesize a ground object synthesis image according to the input encoding result and target ground object type label image. The discriminator is used to determine whether the ground object type of the ground object synthesis image is consistent with the target ground object type label image.

[0070] The following is an introduction to semantic segmentation networks:

[0071] The semantic segmentation network is an improved DeepLab V3+ network, and the improvement is that a Dropout layer is fused into the DeepLab V3+ network.

[0072] Semantic segmentation networks such as Figure 7 As shown, the semantic segmentation network includes an encoder and a decoder connected in sequence. The encoder extracts features from the input image and uses a Dropout layer to regularize the multi-scale features obtained from the feature extraction to obtain encoded features. The decoder upsamples the encoded features and uses a Dropout layer to regularize the upsampled results to obtain decoded features. The decoded features are then used to predict the land cover type of the input image.

[0073] The encoder consists of a deep convolutional network with dilated convolutions and a spatial pyramid pooling module, which mainly learns land cover information at different scales, especially fine land use features. The decoder is a simple upsampling process with Dropout regularization added to deeply fuse low-level and high-level features, thereby improving edge segmentation accuracy.

[0074] The encoder primarily modifies the hole ratio in the Atrous Spatial Pyramid Pooling (ASPP) module and incorporates a Dropout layer to improve model generalization. The overall network structure is as follows: Figure 7 As shown.

[0075] ASPP is a feature extraction method, and the void spatial pyramid pooling module is like... Figure 8 As shown, this method learns feature maps at different scales to extract denser features, and uses dilated convolutions to further enhance the network's ability to extract dense features. The principle is to use dilated convolutions and pooling layers with different dilation rates, connected in parallel, and finally fused together to extract multi-scale information.

[0076] In the ASPP module, dilated convolutions with higher dilation rates are better for extracting larger-sized features, while those with lower dilation rates are better for extracting smaller features. The original ASPP module has dilation rates of 6, 12, and 18. To further improve the model's ability to learn small-sized land use features, this embodiment adds dilated convolutions with a dilation rate of 3 to the ASPP module. Figure 7As shown, finally, the different scale features above are fused (combined) in the channel dimension, then input into a Dropout layer for regularization and a 1x1 convolution layer for dimension reduction to obtain a new feature map, which is output to the decoder.

[0077] Specifically, the encoder comprises, in sequence, a deep convolutional neural network, a dilated spatial pyramid pooling module, a first feature fusion layer, a first Dropout layer, and a first convolutional layer. The first convolutional layer can have a size of 1x1.

[0078] The deep convolutional neural network is configured to extract features from an input image to obtain a first feature and a second feature, wherein the first feature has a smaller size than the second feature.

[0079] The dilated spatial pyramid pooling module comprises a parallel dilated convolution module and a pyramid pooling module. The dilated convolution module is configured to perform dilated convolution on the second feature at different dilated rates to output local features of different scales. The pyramid pooling module is configured to perform global average pooling on the second feature to output a global feature. The dilated convolution module comprises a parallel second convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a third dilated convolutional layer, and a fourth dilated convolutional layer. The second convolutional layer can have a size of 1x1. The first dilated convolutional layer has a dilated rate of 3. The second dilated convolutional layer has a dilated rate of 6. The third dilated convolutional layer has a dilated rate of 12. The fourth dilated convolutional layer has a dilated rate of 18.

[0080] The first feature fusion layer is configured to fuse the local features of different scales and the global feature to obtain multi-scale features.

[0081] The first Dropout layer is configured to perform regularization processing on the multi-scale features.

[0082] The first convolutional layer is configured to perform dimension adjustment on the features after the regularization processing of the first Dropout layer to obtain encoded features.

[0083] The decoder comprises a first upsampling layer, a third convolutional layer, a second feature fusion layer, a second Dropout layer, a fourth convolutional layer, a third Dropout layer, and a second upsampling layer. The first upsampling layer and the second upsampling layer are both four times upsampling layers. The third convolutional layer can have a size of 1x1. The fourth convolutional layer can have a size of 3x3.

[0084] The first upsampling layer is configured to perform upsampling on the encoded features output by the first convolutional layer.

[0085] The third convolutional layer is configured to adjust the dimension of the first feature output by the deep convolutional neural network.

[0086] The second feature fusion layer is configured to fuse the feature output by the first up-sampling layer and the feature output by the third convolutional layer to obtain a deep fusion feature.

[0087] The second Dropout layer is configured to perform regularization processing on the deep fusion feature.

[0088] The fourth convolutional layer is configured to adjust the dimension of the feature after the regularization processing of the second Dropout layer.

[0089] The third Dropout layer is configured to perform regularization processing on the feature output by the fourth convolutional layer.

[0090] The second up-sampling layer is configured to perform up-sampling on the feature after the regularization processing of the third Dropout layer to obtain a decoding feature, and predict the feature type of the input image by using the decoding feature.

[0091] The regularization process of the Dropout layer is introduced as follows:

[0092] The DeepLab V3+ model has a large number of parameters and is prone to overfitting, therefore, the Dropout method is used to regularize it. Dropout is to randomly delete some neural units of the network model, forward propagate the input neural units in the deleted network, back propagate the error in the deleted network, and repeat the above process in different batches of training samples.

[0093] In this embodiment, the Dropout regularization module is integrated into the DeepLab V3+ model, and a suitable Dropout rate is set, so as to obtain a DeepLab V3+ network model with fewer nodes, smaller size and stronger generalization performance.

[0094] The precision of the land use classification model is also evaluated in this embodiment.

[0095] The precision evaluation of the embodiment is mainly from two aspects of qualitative and quantitative. In the qualitative aspect, the new data samples generated by pix2pixhd and GauGAN are compared, and the evaluation is made from the aspects of feature and clarity of ground objects; the land use classification results are evaluated according to subjective experience such as whether the ground objects are continuous and whether the edge contour is clear. In the quantitative aspect, the land use classification results are mainly analyzed by using the indicators such as recall, precision, F1-score, overall accuracy (OA) and MIoU. TP is the sample predicted as positive in the positive sample, FN is the sample predicted as negative in the positive sample, FP is the sample predicted as positive in the sample, and TN is the sample predicted as negative in the negative sample.

[0096] Recall is the proportion of the number of correct classification in the positive sample to the total number of positive samples, and the formula is as follows:

[0097]

[0098] Precision is the proportion of the number of correct classification in the positive sample to the total number of positive samples in the classification result, and the formula is as follows:

[0099]

[0100] F1score is an index based on Recall and Precision, which is an overall evaluation of the two indicators, and the formula is as follows:

[0101]

[0102] OA is the proportion of the number of correct classification to the total number of samples, and the formula is as follows:

[0103]

[0104] MIoU is the average of the ratio of intersection to union of true value and predicted value, which is the evaluation of semantic segmentation result in the global, and the specific definition is as follows:

[0105]

[0106] Wherein, i is the class number, representing the ith class, i starts from 0. K is the total number of classes. Therefore, MIoU is the average of k+1 classes of IoU, and IoU represents the ratio of intersection to union.

[0107] In actual application, a specific implementation process of the above land use classification method is shown in Figure 9 .

[0108] In order to improve the accuracy of land use classification, the embodiment solves the data imbalance problem based on the GauGAN data enhancement strategy, and proposes a land use classification method improved further for the DeepLab V3+ network. The method is trained and tested on the GID dataset, the GauGAN synthesized ground feature image has more rich semantic information and clear ground feature texture characteristics, the GauGAN-GID dataset (data enhanced ground feature set) has excellent precision performance, finally reduces the interference of the sample imbalance problem, and improves the land use classification precision.

[0109] In future research work, first, the model method needs to be improved, while reducing the number of parameters, the performance of model training is not reduced. Real ground feature data needs to be collected, and the expansion of the land use dataset is completed, which is helpful for the GAN network to generate land use data containing rich semantic information and realistic effect, so as to realize image enhancement and improve the land use classification precision.

[0110] The application also provides a land use classification system based on data enhancement, comprising:

[0111] An image acquisition module is configured to acquire a target ground feature image.

[0112] A land use classification module is configured to input the target ground feature image into a land use classification model to obtain a ground feature type of the target ground feature image.

[0113] The land use classification model is obtained by training a semantic segmentation network using a data enhanced ground feature set.

[0114] The data enhanced ground feature set comprises a GID dataset, a ground feature synthesized image and a ground feature type label image corresponding to the ground feature synthesized image; the ground feature synthesized image is obtained by inputting the GID dataset into a GauGAN network; and the semantic segmentation network is a DeepLab V3+ network fused with a Dropout layer.

[0115] The system of the embodiment selects the GauGAN network with the best result performance to generate land use sample data with rich semantics and realistic effect by comparing the ground feature data synthesized by different GAN networks, increases the diversity of data, and maintains the balance between different ground classes in the dataset. At the same time, the DeepLabV3+ network is improved to complete land use classification. The embodiment can provide a new idea for solving the data imbalance problem, and better improve the generalization learning ability and classification precision of the land use classification model.

[0116] The various embodiments described in this specification are presented for the purpose of illustrating the principles of the application and its best mode of operation. Each of the embodiments described in this specification has different advantages and disadvantages, and the embodiments disclosed in this specification are not intended to limit the scope of the application to the embodiments disclosed. Embodiments of the system disclosed in this specification are intended to be limited only by the appended claims.

[0117] The principles and implementations of the present application are described in this specification with specific examples. The above description of the embodiments is only to help understand the method of the present application and its core idea. For those skilled in the art, the specific implementation and application scope of the present application can be changed according to the idea of the present application. In summary, the content of this specification should not be understood as a limitation of the present application.

Claims

1. A land use classification method based on data augmentation, characterized in that, include: Acquire images of the target ground features; The target feature image is input into the land use classification model to obtain the feature type of the target feature image; The land use classification model is obtained by training a semantic segmentation network using a data-enhanced set of land features. The data-augmented feature set includes: a GID dataset, a composite feature map, and feature type label images corresponding to the composite feature map; the composite feature map is obtained by inputting the GID dataset into a GauGAN network; the semantic segmentation network is a DeepLab V3+ network with Dropout layers. The semantic segmentation network includes an encoder and a decoder connected in sequence; The encoder is used to extract features from the input image. The Dropout layer is used to regularize the multi-scale features obtained from the feature extraction to obtain the encoded features. The decoder is used to upsample the encoded features, and the upsampling result is regularized using a Dropout layer to obtain decoded features. The decoded features are then used to predict the land cover type of the input image. The encoder includes: a deep convolutional neural network, a dilated spatial pyramid pooling module, a first feature fusion layer, a first dropout layer, and a first convolutional layer connected in sequence; The deep convolutional neural network is used to extract features from the input image to obtain a first feature and a second feature; the size of the first feature is smaller than the size of the second feature. The dilated spatial pyramid pooling module includes: a parallel dilated convolution module and a pyramid pooling module; the dilated convolution module is used to perform dilated convolution on the second feature at different dilation rates to output local features at different scales; the pyramid pooling module is used to perform global flat pooling on the second feature to output global features. The first feature fusion layer is used to fuse local features at different scales with the global features to obtain multi-scale features; The first Dropout layer is used to perform regularization processing on the multi-scale features. The first convolutional layer is used to adjust the dimensions of the features after the first Dropout layer has been regularized, so as to obtain the encoded features.

2. The land use classification method based on data augmentation according to claim 1, characterized in that, The GauGAN network includes: an image encoding module, a generator, and a discriminator connected in sequence; The image encoding module is used to encode the real images and the corresponding land cover type label images in the input GID dataset to obtain the encoding result; The generator is a fully convolutional network composed of spatially adaptive normalized residual blocks; the generator is used to synthesize a composite land cover map based on the input encoding result and the target land cover type label image. The discriminator is used to determine whether the land cover type in the composite land cover map is consistent with the target land cover type label image.

3. The land use classification method based on data augmentation according to claim 1, characterized in that, The dilated convolution module includes: a parallel second convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a third dilated convolutional layer, and a fourth dilated convolutional layer; The first dilated convolutional layer has a porosity of 3; the second dilated convolutional layer has a porosity of 6; the third dilated convolutional layer has a porosity of 12; and the fourth dilated convolutional layer has a porosity of 18.

4. The land use classification method based on data augmentation according to claim 3, characterized in that, The decoder includes: First upsampling layer, third convolutional layer, second feature fusion layer, second dropout layer, fourth convolutional layer, third dropout layer, and second upsampling layer; The first upsampling layer is used to upsample the encoded features output by the first convolutional layer; The third convolutional layer is used to adjust the dimensions of the first feature output by the deep convolutional neural network; The second feature fusion layer is used to fuse the features output by the first upsampling layer with the features output by the third convolutional layer to obtain deep fused features; The second Dropout layer is used to perform regularization processing on the deep fusion features; The fourth convolutional layer is used to adjust the dimensions of the features after the second Dropout layer has been regularized. The third Dropout layer is used to perform regularization processing on the features output by the fourth convolutional layer. The second upsampling layer is used to upsample the features after the regularization process of the third Dropout layer to obtain decoded features, and to use the decoded features to predict the land cover type of the input image.

5. A land use classification method based on data augmentation according to claim 4, characterized in that, Both the first upsampling layer and the second upsampling layer are quadruple upsampling layers.

6. A land use classification method based on data augmentation according to claim 4, characterized in that, The size of the first convolutional layer, the second convolutional layer, and the third convolutional layer is 1×1; the size of the fourth convolutional layer is 3×3.

7. A land use classification system based on data augmentation, characterized in that, include: The image acquisition module is used to acquire images of the target features. The land use classification module is used to input the target land feature image into the land use classification model to obtain the land feature type of the target land feature image; The land use classification model is obtained by training a semantic segmentation network using a data-enhanced set of land features. The data-augmented feature set includes: a GID dataset, a composite feature map, and feature type label images corresponding to the composite feature map; the composite feature map is obtained by inputting the GID dataset into a GauGAN network; the semantic segmentation network is a DeepLab V3+ network with Dropout layers. The semantic segmentation network includes an encoder and a decoder connected in sequence; The encoder is used to extract features from the input image. The Dropout layer is used to regularize the multi-scale features obtained from the feature extraction to obtain the encoded features. The decoder is used to upsample the encoded features, and the upsampling result is regularized using a Dropout layer to obtain decoded features. The decoded features are then used to predict the land cover type of the input image. The encoder includes: a deep convolutional neural network, a dilated spatial pyramid pooling module, a first feature fusion layer, a first dropout layer, and a first convolutional layer connected in sequence; The deep convolutional neural network is used to extract features from the input image to obtain a first feature and a second feature; the size of the first feature is smaller than the size of the second feature. The dilated spatial pyramid pooling module includes: a parallel dilated convolution module and a pyramid pooling module; the dilated convolution module is used to perform dilated convolution on the second feature at different dilation rates to output local features at different scales; the pyramid pooling module is used to perform global flat pooling on the second feature to output global features. The first feature fusion layer is used to fuse local features at different scales with the global features to obtain multi-scale features; The first Dropout layer is used to perform regularization processing on the multi-scale features. The first convolutional layer is used to adjust the dimensions of the features after the first Dropout layer has been regularized, so as to obtain the encoded features.

Images