Crop remote sensing image classification method and system based on improved U-Net

Through the multi-scale feature extraction and feature fusion device of the improved U-Net network, the problems of heterospectral and heterospectral of heterospectral in remote sensing image classification are solved, and the accuracy and deep feature extraction capabilities of crop classification are improved, so as to avoid information loss and calculation Increase in quantity.

CN114550002BActive Publication Date: 2024-12-06SHANDONG FENGSHI INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210032839.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-12
Publication Date
2024-12-06
Estimated Expiration
2042-01-12

AI Technical Summary

Technical Problem

The existing remote sensing image classification method is difficult to effectively distinguish different crop categories, and there are problems with heterospectral and heterospectral problems. Traditional machine learning and U-Net models perform poorly in deep feature extraction, and the splicing process of codecs leads to Information is lost.

Method used

Adopting an improved U-Net network, multi-scale feature extraction and new feature fusion device, combined with hollow convolution and parallel sampling technology, enhance the feature extraction capability of the network, and rationally configure each layer of features through feature fusion device to avoid excessive extraction, Increased computational volume and decreased performance.

Benefits of technology

It realizes efficient classification of remote sensing images, avoids interference in non-crop planting areas, deeply explores data information, and improves the accuracy of crop classification and multi-scale feature extraction capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114550002B_ABST
    Figure CN114550002B_ABST
Patent Text Reader

Abstract

The present invention discloses a crop remote sensing image classification method and system based on improved U-Net, wherein the method comprises: obtaining a crop remote sensing image to be classified; preprocessing the crop remote sensing image to be classified; using a trained improved U-Net network to classify the preprocessed image to obtain a classification result of the crop. The present invention has a more powerful multi-scale feature extraction capability and can deeply mine data information; the present invention considers the feature abstraction level processed by each layer of the network, reasonably configures the feature fusion part, and avoids the situation of a substantial increase in the amount of calculation or a decrease in performance due to excessive feature extraction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of remote sensing classification technology, and in particular to a crop remote sensing image classification method and system based on an improved U-Net. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] Agriculture has always been a major livelihood issue that the country has focused on. The planting area of ​​crops has a great impact on food production, so crop area estimation is an important part of agricultural detection. Among them, the classification of crops is the core issue of crop area estimation. Classifying crops through remote sensing is more efficient than manual operation and is of great significance.

[0004] Common research on crop classification through remote sensing images mainly uses two methods: element-oriented and object-oriented. The element-oriented method has the problem of different spectra for the same object and the same spectra for different objects, while the object-oriented classification has the contradiction of inconsistency between the image segmentation objects and geographic entities, and it is difficult to avoid interference from non-crop planting areas.

[0005] Traditional machine learning methods, such as SVM and decision trees, are also difficult to achieve good results when used for remote sensing image classification. These methods are shallow classification and it is difficult to extract deep features hidden in the image.

[0006] In addition, in the ordinary U-Net model, the splicing process between the codec is very simple, which only involves cropping the encoder input to a suitable size and then splicing it into the decoder, resulting in a large amount of loss of original image information. Summary of the invention

[0007] In order to overcome the deficiencies of the prior art, the present invention provides a crop remote sensing image classification method and system based on an improved U-Net, thereby improving the accuracy of crop classification.

[0008] In a first aspect, the present invention provides a crop remote sensing image classification method based on an improved U-Net;

[0009] The crop remote sensing image classification method based on improved U-Net includes:

[0010] Acquire remote sensing images of crops to be classified;

[0011] Preprocessing the crop remote sensing images to be classified;

[0012] The trained improved U-Net network is used to classify the preprocessed images and obtain the classification results of crops.

[0013] In a second aspect, the present invention provides a crop remote sensing image classification system based on an improved U-Net;

[0014] The crop remote sensing image classification system based on improved U-Net includes:

[0015] An acquisition module is configured to: acquire a remote sensing image of crops to be classified;

[0016] A preprocessing module is configured to: preprocess the crop remote sensing image to be classified;

[0017] The classification module is configured to: use the trained improved U-Net network to classify the preprocessed images to obtain the classification results of crops.

[0018] In a third aspect, the present invention further provides an electronic device, comprising:

[0019] a memory for non-transitory storage of computer-readable instructions; and

[0020] a processor for executing the computer readable instructions,

[0021] When the computer-readable instructions are executed by the processor, the method described in the first aspect is executed.

[0022] In a fourth aspect, the present invention further provides a storage medium that non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, the instructions of the method described in the first aspect are executed.

[0023] In a fifth aspect, the present invention further provides a computer program product, comprising a computer program, wherein the computer program is used to implement the method described in the first aspect when running on one or more processors.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] Remote sensing images are divided into different plots through deep networks to distinguish different types of crops. There is no problem of different spectra for the same object or the same spectrum for different objects, and interference from non-crop planting areas can also be avoided. Compared with traditional machine learning and ordinary U-Net networks, the present invention has more powerful multi-scale feature extraction capabilities and can deeply mine data information. The present invention takes into account the degree of feature abstraction processed by each layer of the network, and reasonably configures the feature fusion part, avoiding a significant increase in computing power or performance degradation due to excessive feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0027] Figure 1 is a flow chart of the method of embodiment 1;

[0028] Figure 2(a) to Figure 2(d) Schematic diagram of the internal structure of four fusion devices in Example 1. DETAILED DESCRIPTION

[0029] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0030] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0032] In this embodiment, all data is obtained in compliance with laws and regulations and based on the user's consent, and is used legally.

[0033] The present invention aims to solve the problem of how to obtain the features that can reflect the crop category in remote sensing images, and achieves this goal through multi-scale feature extraction and the feature fusion device that the present invention focuses on improving. The present invention constructs a new model with a U-shaped convolutional network as the basic framework, such as Figure 1 shown.

[0034] Embodiment 1

[0035] This embodiment provides a crop remote sensing image classification method based on improved U-Net;

[0036] The crop remote sensing image classification method based on improved U-Net includes:

[0037] S101: Acquire a remote sensing image of crops to be classified;

[0038] S102: Preprocessing the crop remote sensing image to be classified;

[0039] S103: Using the trained improved U-Net network, classify the preprocessed image to obtain the classification result of the crops.

[0040] Furthermore, the S102: preprocessing the crop remote sensing images to be classified specifically includes:

[0041] The crop remote sensing image to be classified is randomly cut to obtain several segmented images.

[0042] For example, the original remote sensing image generally has a high resolution, and a certain preprocessing is required before the image is input into the network. In the present invention, the original image is randomly cut to obtain a small image with a resolution of 400*400.

[0043] Furthermore, if Figure 1 As shown, the improved U-Net network has a network structure including:

[0044] The encoder, connector and decoder are connected in sequence;

[0045] The encoder comprises: a first-level sub-encoder, a second-level sub-encoder, a third-level sub-encoder and a fourth-level sub-encoder connected in sequence; the fourth-level sub-encoder is connected to the input end of the connector;

[0046] The decoder comprises: a first-level sub-decoder, a second-level sub-decoder, a third-level sub-decoder and a fourth-level sub-decoder connected in sequence; the fourth-level sub-decoder is connected to the output end of the connector;

[0047] The output end of the first-level sub-encoder is connected to the input end of the fourth-level sub-decoder through a first fusion device;

[0048] The output end of the second-stage sub-encoder is connected to the input end of the third-stage sub-decoder via a second fusion device;

[0049] The output end of the third-level sub-encoder is connected to the input end of the second-level sub-decoder through a third fuser;

[0050] The output end of the fourth-level sub-encoder is connected to the input end of the first-level sub-decoder through a fourth fuser.

[0051] Furthermore, as shown in Figure 2(a), the first fuser includes: four parallel-connected dilated convolutional layers, the input ends of the four dilated convolutional layers are all connected to the output ends of the first-level sub-encoder; the output ends of the four dilated convolutional layers are all connected to the input ends of the fourth-level sub-decoder.

[0052] Furthermore, the dilated convolution layer refers to a convolution method that expands the receptive field by adding holes on the basis of ordinary convolution kernels, and introduces a new parameter of "expansion rate" to the convolution layer to define the spacing between the values ​​of the convolution kernel. For example, the 3*3 convolution in the present invention has a dilation rate of 1. If the dilation rate is set to 2, it is equivalent to adding an interval between the values, and its receptive field will become 5*5.

[0053] Furthermore, the working principle of the first fusion device is:

[0054] By setting four dilation convolutions with different expansion rates, the feature maps output by the first-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding fourth-level sub-decoder to achieve feature fusion between the corresponding codecs.

[0055] Furthermore, as shown in Figure 2(b), the second fuser includes: three parallel-connected dilated convolutional layers, the input ends of the three dilated convolutional layers are all connected to the output ends of the second-level sub-encoder; the output ends of the three dilated convolutional layers are all connected to the input ends of the third-level sub-decoder.

[0056] Furthermore, the working principle of the second fusion device is:

[0057] By setting three dilation convolutions with different expansion rates, the feature maps output by the second-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding third-level sub-decoder to achieve feature fusion between the corresponding codecs.

[0058] Furthermore, as shown in Figure 2(c), the third fuser includes: two parallel-connected dilated convolutional layers, the input ends of the two dilated convolutional layers are both connected to the output ends of the third-level sub-encoder; the output ends of the two dilated convolutional layers are both connected to the input ends of the second-level sub-decoder.

[0059] Furthermore, the working principle of the third fusion device is:

[0060] By setting two dilation convolutions with different expansion rates, the feature maps output by the third-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding second-level sub-decoder to achieve feature fusion between the corresponding codecs.

[0061] Furthermore, as shown in Figure 2(d), the fourth fuser includes: a dilated convolution layer a; the input end of the dilated convolution layer a is connected to the output end of the fourth-level sub-encoder; and the output ends of the dilated convolution layer a are connected to the input ends of the first-level sub-decoder.

[0062] Furthermore, the working principle of the fourth fusion device is:

[0063] By setting a dilation rate hole convolution, the feature map output by the fourth-level sub-encoder is sampled, and the output result is spliced ​​to the input of the corresponding first-level sub-decoder to achieve feature fusion between the corresponding codecs.

[0064] Furthermore, the working principle of the improved U-Net network is:

[0065] Input a single-channel remote sensing image with a resolution of 400*400, and after two 3*3 convolutions and one maximum pooling downsampling, input it into the next level encoder; the network encoder has a total of 4 levels, and finally outputs a 20*20 512-channel highly abstract feature map;

[0066] The decoder part corresponds one-to-one with the encoder, and also has 4 levels. The 20*20 feature map finally output by the encoder is upsampled by deconvolution to double its size. After passing through the middle part fusion device, the input of the corresponding encoder and the output of the current decoder are merged to obtain a new feature map. Then, after two 3*3 convolutions, it is input into the next level decoder, and finally a 196*196 resolution feature map is obtained.

[0067] The convolution operations in the encoder / decoder all use Relu as the activation function, which is expressed as: f(x)=max(0,x).

[0068] Further, the internal structures of the first-level sub-encoder, the second-level sub-encoder, the third-level sub-encoder and the fourth-level sub-encoder are consistent;

[0069] The first-level sub-encoder includes: a convolutional layer c1, a convolutional layer c2 and a maximum pooling layer connected in sequence.

[0070] Further, the internal structures of the first-level sub-decoder, the second-level sub-decoder, the third-level sub-decoder and the fourth-level sub-decoder are consistent;

[0071] The first-level sub-decoder includes: a deconvolution layer, a convolution layer d1 and a convolution layer d2 connected in sequence.

[0072] Furthermore, the trained improved U-Net network is obtained by training the improved U-Net network using crop remote sensing images marked with classification labels.

[0073] Furthermore, for the trained improved U-Net network, the principle of stopping training is to stop training when the loss function value stops decreasing.

[0074] Furthermore, the loss function is:

[0075]

[0076] Among them, p l(x) (x) is the softmax loss function, l:Ω→{1,…,K} is the pixel label value, ω:Ω∈R is the pixel weight, giving higher weights to pixels at the boundary points in the image.

[0077]

[0078] where ω c :Ω∈R is the weight for balancing the category ratio, d1:Ω∈R is the distance from the pixel to the category closest to it, and d2:Ω∈R is the distance from the pixel to the second closest category.

[0079] The present invention proposes a new feature fuser for U-Net. For each set of corresponding codecs, the encoder input must pass through a larger-sized parallel sampling hole convolution group and set a suitable step size according to the resolution calculation to obtain the same resolution as the decoder output. Since there are 4 sets of codecs in the network, the present invention sets 4 sets of 3*3 parallel sampling hole convolutions. Figure 2(a) to Figure 2(d) The first group of encoders / decoders are numbered as follows. The first group of encoders / decoders is the outermost layer, and the feature graphs it processes have the lowest level of abstraction. Four parallel sampling hole convolutions are required to extract deep information. As the level of abstraction of the feature graph increases, the number of parallel sampling hole convolutions required will decrease accordingly. In order, the second, third, and fourth groups of encoders / decoders need to set 3, 2, and 1 parallel sampling hole convolutions, respectively. The outputs of each group are added and spliced ​​into the corresponding encoder input. Figure 2(a) to Figure 2(d) Schematic diagram of feature fusion devices at each layer.

[0080] Figure 2(a) shows that the weight intervals (rates) of the four parallel sampling dilated convolutions in the first group are 98, 40, 16, and 4 respectively;

[0081] Figure 2(b) shows that the weight intervals of the three parallel sampling dilated convolutions in the second group are 40, 16, and 4 respectively;

[0082] Figure 2(c) shows that the weight intervals of the two parallel sampling dilated convolutions in the third group are 16 and 4 respectively;

[0083] Figure 2(d) shows that the weight interval of 1 parallel sampling dilated convolution in the fourth group is 4;

[0084] The above parameters are determined by the image resolution of the input network.

[0085] The network finally outputs the segmented image through a layer of 1*1 convolution, and the loss function is a loss function with boundary weights:

[0086]

[0087] where p l(x) (x) is the softmax loss function, l:Ω→{1,…,K} is the pixel label value, ω:Ω∈R is the pixel weight, giving higher weights to pixels at the boundary points in the image.

[0088]

[0089] where ω c :Ω∈R is the weight for balancing the category ratio, d1:Ω∈R is the distance from the pixel to the category closest to it, and d2:Ω∈R is the distance from the pixel to the second closest category. In the experiment, ω0=10,σ≈5.

[0090] In order to ensure that a sufficient number of training sets are obtained, the present invention performs data augmentation operations on the input images in the experiment, namely:

[0091] (1) The original image and label are rotated 90°, 180°, and 270°;

[0092] (2) The original image and label are mirrored along the y-axis;

[0093] (3) Blur and adjust the lighting of the original image;

[0094] (4) Noise the original image.

[0095] 85% of the obtained data is set as the training set, and 15% is set as the validation set. Set the learning rate to 0.000001, repeat the training for 100 epochs to obtain a trained model, and then input the remote sensing image of the crop to be predicted for classification prediction. The database images used in the experiment of the present invention include corn, cotton, sesame, broad-leaved soybeans, narrow-leaved soybeans and rice, which are distinguished by digital labels 1-6. The output of the network is the digital label of the crop category at each pixel position in the input image. It should be noted that the image to be predicted also needs to be divided into small images before entering the network.

[0096] According to the output prediction results, the planting area of ​​each crop can also be counted. Different remote sensing images have their own spatial resolution. The planting area of ​​each crop can be estimated by multiplying the number of pixels contained in each category of crops in the statistical prediction results by the spatial resolution.

[0097] Embodiment 2

[0098] This embodiment provides a crop remote sensing image classification system based on improved U-Net;

[0099] The crop remote sensing image classification system based on improved U-Net includes:

[0100] An acquisition module is configured to: acquire a remote sensing image of crops to be classified;

[0101] A preprocessing module is configured to: preprocess the crop remote sensing image to be classified;

[0102] The classification module is configured to: use the trained improved U-Net network to classify the preprocessed images to obtain the classification results of crops.

[0103] It should be noted that the acquisition module, preprocessing module and classification module described above correspond to steps S101 to S103 in Embodiment 1, and the examples and application scenarios implemented by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the modules described above as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0104] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0105] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0106] Embodiment 3

[0107] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory so that the electronic device executes the method described in the above embodiment one.

[0108] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0109] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0110] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software.

[0111] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0112] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0113] Embodiment 4

[0114] This embodiment further provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the method described in the first embodiment is completed.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. The crop remote sensing image classification method based on improved U-Net is characterized by: include: Obtaining remote sensing images of crops to be classified; Preprocessing of crop remote sensing images to be classified; The trained improved U-Net network is used to classify the preprocessed images and obtain the classification results of crops; The improved U-Net network has a network structure including: The encoder, connector and decoder are connected in sequence; The encoder comprises: a first-stage sub-encoder, a second-stage sub-encoder, a third-stage sub-encoder and a fourth-stage sub-encoder connected in sequence; the fourth-stage sub-encoder is connected to the input end of the connector; The decoder comprises: a first-level sub-decoder, a second-level sub-decoder, a third-level sub-decoder and a fourth-level sub-decoder connected in sequence; the fourth-level sub-decoder is connected to the output end of the connector; The output of the first-level sub-encoder is connected to the input of the fourth-level sub-decoder via a first fuser; the output of the second-level sub-encoder is connected to the input of the third-level sub-decoder via a second fuser; the output of the third-level sub-encoder is connected to the input of the second-level sub-decoder via a third fuser; and the output of the fourth-level sub-encoder is connected to the input of the first-level sub-decoder via a fourth fuser. The first fuser includes: four parallel-connected dilated convolutional layers, wherein the input ends of the four dilated convolutional layers are connected to the output ends of the first-stage sub-encoder; and the output ends of the four dilated convolutional layers are connected to the input ends of the fourth-stage sub-decoder; The working principle of the first fusion device is: By setting four dilated convolutions with different expansion rates, the feature maps output by the first-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding fourth-level sub-decoder to achieve feature fusion between the corresponding codecs. The second fuser includes: three parallel-connected dilated convolutional layers, wherein the input ends of the three dilated convolutional layers are connected to the output ends of the second-stage sub-encoder; and the output ends of the three dilated convolutional layers are connected to the input ends of the third-stage sub-decoder; The working principle of the second fusion device is: By setting three dilated convolutions with different expansion rates, the feature maps output by the second-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding third-level sub-decoder to achieve feature fusion between the corresponding codecs. The third fuser includes: two parallel-connected dilated convolutional layers, wherein the input ends of the two dilated convolutional layers are both connected to the output end of the third-stage sub-encoder; and the output ends of the two dilated convolutional layers are both connected to the input end of the second-stage sub-decoder; The working principle of the third fusion device is: By setting two dilated convolutions with different expansion rates, the feature maps output by the third-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding second-level sub-decoder to achieve feature fusion between the corresponding codecs. The fourth fuser includes: a dilated convolution layer a; the input end of the dilated convolution layer a is connected to the output end of the fourth-stage sub-encoder; the output ends of the dilated convolution layer a are both connected to the input end of the first-stage sub-decoder; The working principle of the fourth fusion device is: By setting a suitable dilation rate dilated convolution, the feature map output by the fourth-level sub-encoder is sampled, and the output result is spliced ​​to the input of the corresponding first-level sub-decoder to achieve feature fusion between the corresponding codecs; The improved U-Net network works as follows: A 400*400 resolution single-channel remote sensing image is input, and after two 3*3 convolutions and one maximum pooling downsampling, it is input to the next level encoder. The network encoder has a total of 4 levels, and finally outputs a 20*20 512-channel highly abstract feature map. The decoder part corresponds one-to-one with the encoder and also has 4 levels. The 20*20 feature map output by the encoder is upsampled by deconvolution to double its size. After passing through the intermediate fuser, the input of the corresponding encoder and the output of the current decoder are merged to obtain a new feature map. After two 3*3 convolutions, it is input into the next level decoder to finally obtain a feature map with a resolution of 196*196.

2. The crop remote sensing image classification system based on improved U-Net is characterized by: include: An acquisition module is configured to: acquire a remote sensing image of crops to be classified; A preprocessing module is configured to: preprocess the crop remote sensing image to be classified; The classification module is configured to: use the trained improved U-Net network to classify the preprocessed images to obtain the classification results of crops; The improved U-Net network has a network structure including: The encoder, connector and decoder are connected in sequence; The encoder comprises: a first-stage sub-encoder, a second-stage sub-encoder, a third-stage sub-encoder and a fourth-stage sub-encoder connected in sequence; the fourth-stage sub-encoder is connected to the input end of the connector; The decoder comprises: a first-level sub-decoder, a second-level sub-decoder, a third-level sub-decoder and a fourth-level sub-decoder connected in sequence; the fourth-level sub-decoder is connected to the output end of the connector; The output of the first-level sub-encoder is connected to the input of the fourth-level sub-decoder via a first fuser; the output of the second-level sub-encoder is connected to the input of the third-level sub-decoder via a second fuser; the output of the third-level sub-encoder is connected to the input of the second-level sub-decoder via a third fuser; and the output of the fourth-level sub-encoder is connected to the input of the first-level sub-decoder via a fourth fuser. The first fuser includes: four parallel-connected dilated convolutional layers, wherein the input ends of the four dilated convolutional layers are connected to the output ends of the first-stage sub-encoder; and the output ends of the four dilated convolutional layers are connected to the input ends of the fourth-stage sub-decoder; The working principle of the first fusion device is: By setting four dilated convolutions with different expansion rates, the feature maps output by the first-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding fourth-level sub-decoder to achieve feature fusion between the corresponding codecs. The second fuser includes: three parallel-connected dilated convolutional layers, wherein the input ends of the three dilated convolutional layers are connected to the output ends of the second-stage sub-encoder; and the output ends of the three dilated convolutional layers are connected to the input ends of the third-stage sub-decoder; The working principle of the second fusion device is: By setting three dilated convolutions with different expansion rates, the feature maps output by the second-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding third-level sub-decoder to achieve feature fusion between the corresponding codecs. The third fuser includes: two parallel-connected dilated convolutional layers, wherein the input ends of the two dilated convolutional layers are both connected to the output end of the third-stage sub-encoder; and the output ends of the two dilated convolutional layers are both connected to the input end of the second-stage sub-decoder; The working principle of the third fusion device is: By setting two dilated convolutions with different expansion rates, the feature maps output by the third-level sub-encoder are sampled in parallel at different sampling rates. The output results are added and spliced ​​to the input of the corresponding second-level sub-decoder to achieve feature fusion between the corresponding codecs. The fourth fuser includes: a dilated convolution layer a; the input end of the dilated convolution layer a is connected to the output end of the fourth-stage sub-encoder; the output ends of the dilated convolution layer a are both connected to the input end of the first-stage sub-decoder; The working principle of the fourth fusion device is: By setting a suitable dilation rate dilated convolution, the feature map output by the fourth-level sub-encoder is sampled, and the output result is spliced ​​to the input of the corresponding first-level sub-decoder to achieve feature fusion between the corresponding codecs; The improved U-Net network works as follows: A 400*400 resolution single-channel remote sensing image is input, and after two 3*3 convolutions and one maximum pooling downsampling, it is input to the next level encoder. The network encoder has a total of 4 levels, and finally outputs a 20*20 512-channel highly abstract feature map. The decoder part corresponds one-to-one with the encoder and also has 4 levels. The 20*20 feature map output by the encoder is upsampled by deconvolution to double its size. After passing through the intermediate fuser, the input of the corresponding encoder and the output of the current decoder are merged to obtain a new feature map. After two 3*3 convolutions, it is input into the next level decoder to finally obtain a feature map with a resolution of 196*196.

3. An electronic device, comprising: a memory for non-transitory storage of computer-readable instructions; as well as a processor for executing said computer-readable instructions, When the computer-readable instructions are executed by the processor, the method according to claim 1 is performed.

4. A storage medium, characterized in that: Computer-readable instructions are non-transitory stored, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method of claim 1 are executed.