A semi-supervised satellite image semantic segmentation network construction method and device and electronic equipment

Abstract

The application relates to a semi-supervised satellite image semantic segmentation network construction method and device and electronic equipment, and the method comprises the following steps: collecting optical remote sensing image images; the optical remote sensing image images are preprocessed to obtain training data; a full convolution image segmentation network is constructed; the full convolution image segmentation network is trained using the training data to obtain an optimized full convolution image segmentation network; a semi-supervised knowledge distillation network is constructed based on the optimized full convolution image segmentation network; and the semi-supervised knowledge distillation network is trained using the training data. The semi-supervised satellite image semantic segmentation network construction method, device and electronic equipment provided by the application extract label sample knowledge through a deep learning teacher network, and transfer the knowledge to a student network, so that the label sample category reaches a high prediction accuracy; the accuracy of the target category is improved, and the model generalization capability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical remote sensing image, in particular to a semi-supervised satellite image semantic segmentation network construction method and device and electronic equipment. BACKGROUND

[0002] Image segmentation technology is to divide an image into different types of uniform regions according to the internal characteristics of the image, and the segmentation edges between regions must be accurately defined, and the internal characteristics of the segmented objects have consistency or similarity, each region belongs to the same class, and different regions belong to different classes. The optical remote sensing image segmentation technology aims to classify the remote sensing image at the pixel level according to the actual semantic information, and divide it into a series of regions with road, farmland, village, industrial area and other land object class labels. In recent years, with the progress of remote sensing technology, image segmentation of massive observation image data obtained by remote sensing satellites (such as LandSat, sentinel series and MODIS series) is the processing basis for application researches such as city planning, disaster monitoring and target identification. However, the rapid growth of remote sensing data also brings many challenges to optical remote sensing image segmentation, such as higher land object complexity brought by improved spatial resolution, same object different spectrum and different object same spectrum phenomenon brought by rapidly changing spectral information, and processing of massive data and multi-scale feature extraction.

[0003] Early image segmentation is based on mathematical statistics knowledge, such as segmentation based on Markov random field, conditional random field, graph theory and wavelet transform. In recent years, with the new artificial intelligence boom of big data, new solutions have been brought to the field of image segmentation. The model based on convolutional neural network in deep learning, a typical example of which is the fully convolutional network, is considered to be a key architecture that breaks through the long-standing bottleneck of image segmentation technology. The strong robustness is a typical advantage of deep neural network, and the deep features extracted have high-level semantic information, and the deeper the network layer, the more abstract and global information can be extracted, and the stronger the feature representation ability. Later, the segmentation method based on deep features was gradually transferred to optical remote sensing images, and was improved according to the characteristics of remote sensing images, such as multi-spectrum, high land object complexity, high inter-class similarity and intra-class difference, which are different from natural images. Among them, the segmentation model based on convolutional neural network is a prominent representative, and most of the segmentation methods based on deep features are developed on the basis of this network.

[0004] Compared with the early traditional optical remote sensing image segmentation, the segmentation method based on deep features can use neural network to implicitly establish the mapping relationship from pixels to semantics, and the network automatically learns to extract target features to complete the entire segmentation process without introducing artificial feature intervention, which can better learn the characteristics of remote sensing images, improve the recognition accuracy and reduce time consumption. However, the performance of most neural networks is extremely dependent on the amount of training sample data, which belongs to the strong supervised semantic segmentation method, but the labeling of training samples is a task with high time and economic cost, and manual labeling also has subjectivity and uncertainty. When encountering insufficient original sample of remote sensing image, such as hyperspectral remote sensing data, the performance of strong supervised semantic segmentation will be restricted. SUMMARY

[0005] In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a semi-supervised satellite image semantic segmentation network construction method and device and electronic equipment.

[0006] In the first aspect, the present application provides a semi-supervised satellite image semantic segmentation network construction method, which comprises the following steps:

[0007] Collecting optical remote sensing image;

[0008] Preprocessing the optical remote sensing image to obtain training data;

[0009] Constructing a full convolution image segmentation network;

[0010] Training the full convolution image segmentation network using the training data to obtain an optimized full convolution image segmentation network;

[0011] Constructing a semi-supervised knowledge distillation network based on the optimized full convolution image segmentation network;

[0012] Training the semi-supervised knowledge distillation network using the training data.

[0013] Preferably, the collecting of optical remote sensing image comprises the following steps:

[0014] Collecting optical remote sensing image of different regions and different months;

[0015] Judging whether the cloud content of the optical remote sensing image is greater than a preset threshold;

[0016] If yes, the optical remote sensing image is removed;

[0017] If no, the optical remote sensing image is retained;

[0018] Obtaining all the remaining optical remote sensing images.

[0019] Preferably, the preprocessing of the optical remote sensing image to obtain the training data comprises the steps of:

[0020] atmospheric correction of the optical remote sensing image;

[0021] geometric correction of the optical remote sensing image after atmospheric correction;

[0022] size cropping of the optical remote sensing image after geometric correction to obtain small-size optical patch data;

[0023] local tagging of the small-size optical patch data.

[0024] Preferably, the local tagging of the small-size optical patch data comprises the steps of:

[0025] selecting a label category;

[0026] randomly dividing the small-size optical patch data into first patch data and second patch data;

[0027] tagging the first patch data using the label category;

[0028] using the first patch data after tagging as first input data;

[0029] using the second patch data as second input data.

[0030] Preferably, the constructing of the fully convolutional image segmentation network comprises the steps of:

[0031] constructing an encoder using two-dimensional convolutional layers and pooling layers;

[0032] constructing a transition layer using two-dimensional convolutional layers;

[0033] constructing a decoder using deconvolution and skip-connection structure;

[0034] constructing a first loss function.

[0035] Preferably, the training of the fully convolutional image segmentation network using the training data to obtain an optimized fully convolutional image segmentation network comprises the steps of:

[0036] obtaining first input data with labels in the training data;

[0037] inputting the first input data into the encoder and the transition layer of the fully convolutional image segmentation network in sequence to obtain learning features;

[0038] inputting the learning features into the decoder of the fully convolutional image segmentation network to obtain multi-dimensional features;

[0039] aligning the multi-dimensional features with label data to obtain optimization data;

[0040] optimizing the fully convolutional image segmentation network using the optimization data to obtain a fully convolutional image segmentation model.

[0041] Preferably, the constructing a semi-supervised knowledge distillation network based on the optimized fully convolutional image segmentation network comprises steps of:

[0042] obtaining a fully convolutional image segmentation model in the optimized fully convolutional image segmentation network;

[0043] using the fully convolutional image segmentation model as a teacher network;

[0044] constructing a student network using the fully convolutional image segmentation network;

[0045] constructing a second loss function.

[0046] Preferably, the training the semi-supervised knowledge distillation network using the training data comprises steps of:

[0047] obtaining second input data without label in the training data;

[0048] inputting the second input data into the teacher network of the semi-supervised knowledge distillation network to make a prediction segmentation;

[0049] inputting the second input data into the student network of the semi-supervised knowledge distillation network to extract multi-level features;

[0050] segmenting the second input data into (K+M) classes;

[0051] aligning a first output result of the teacher network and a second output result of the student network to calculate a target class loss function;

[0052] performing spectral clustering-normal cut criterion on the second input data according to the second output result to calculate a soft label loss function;

[0053] combining the target class loss function and the soft label loss function to perform feedback optimization on the student network and output a student network segmentation model.

[0054] In a second aspect, the present application provides a semi-supervised satellite image semantic segmentation network construction device, comprising:

[0055] an optical remote sensing image collection module, configured to collect optical remote sensing images;

[0056] an optical remote sensing image preprocessing module, configured to preprocess the optical remote sensing images to obtain training data;

[0057] a full convolution image segmentation network construction module, configured to construct a full convolution image segmentation network;

[0058] an optimized full convolution image segmentation network generation module, configured to train the full convolution image segmentation network using the training data to obtain an optimized full convolution image segmentation network;

[0059] a network construction semi-supervised knowledge distillation network segmentation module, configured to construct a semi-supervised knowledge distillation network based on the optimized full convolution image segmentation network;

[0060] a semi-supervised knowledge distillation network training module, configured to train the semi-supervised knowledge distillation network using the training data.

[0061] In a third aspect, an electronic device is provided, and the electronic device comprises:

[0062] at least one processor; and

[0063] a memory in communication with the at least one processor; wherein

[0064] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the semi-supervised satellite image semantic segmentation network construction method according to any one of the preceding claims 1-7.

[0065] Compared with the prior art, the above technical solutions provided by the embodiments of the present application have the following advantages:

[0066] The semi-supervised satellite image semantic segmentation network construction method, device and electronic device provided by the present application extract label sample knowledge through a deep learning teacher network, and transfer the knowledge to a student network, so that the prediction accuracy of the labeled sample categories is high, the accuracy of the target categories is improved, and the model generalization ability is improved. BRIEF DESCRIPTION OF DRAWINGS

[0067] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0069] Figure 1A flowchart of a semi-supervised satellite image semantic segmentation network construction method provided in an embodiment of the present application is shown in the figure.

[0070] Figure 2 A structural diagram of a semi-supervised satellite image semantic segmentation network construction device provided in an embodiment of the present application is shown in the figure.

[0071] Figure 3 A structural diagram of an electronic device provided in the present application is shown in the figure. DETAILED DESCRIPTION

[0072] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0073] Figure 1 A flowchart of a semi-supervised satellite image semantic segmentation network construction method provided in an embodiment of the present application is shown in the figure.

[0074] The present application provides a semi-supervised satellite image semantic segmentation network construction method, which comprises the following steps:

[0075] S1: Collect optical remote sensing image images;

[0076] In the embodiments of the present application, the collecting of the optical remote sensing image images comprises the following steps:

[0077] Collect optical remote sensing image images of different regions and different months;

[0078] Determine whether the cloud content of the optical remote sensing image images is greater than a preset threshold value;

[0079] If yes, discard the optical remote sensing image images;

[0080] If no, retain the optical remote sensing image images;

[0081] Obtain all the remaining optical remote sensing image images.

[0082] Specifically, optical remote sensing image images of different regions and different months are collected using a sentinel2 satellite, and in order to accurately analyze the subsequent results, images with a cloud content lower than a preset threshold value are selected from all the optical remote sensing image images as subsequent analysis objects to reduce the interference of clouds on the images. In the present application, 10 sentinel2 images with a cloud content of less than 5% are selected from a large number of optical remote sensing image images.

[0083] S2: pre-process the optical remote sensing image to obtain training data;

[0084] In the embodiments of the present application, the pre-processing of the optical remote sensing image to obtain training data includes the steps of:

[0085] atmospheric correction is performed on the optical remote sensing image;

[0086] geometric correction is performed on the optical remote sensing image after atmospheric correction;

[0087] size cropping is performed on the optical remote sensing image after geometric correction to obtain small-size optical spot data;

[0088] local marking is performed on the small-size optical spot data.

[0089] Specifically, the collected sentinel2 image is pre-processed by atmospheric correction, geometric correction, etc. to obtain pre-processed optical image pre-processing data. Due to the large size of the optical image pre-processing data, the subsequent analysis data volume will be increased. In order to reduce the analysis data volume, the large-size optical image pre-processing data needs to be cropped into small-size optical spot data, and the small-size optical spot data needs to be locally marked.

[0090] In the embodiments of the present application, the local marking of the small-size optical spot data includes the steps of:

[0091] selecting a label category;

[0092] randomly dividing the small-size optical spot data into first spot data and second spot data;

[0093] marking the first spot data using the label category;

[0094] the first spot data after marking is used as first input data;

[0095] the second spot data is used as second input data.

[0096] Specifically, the label categories cannot be too few to avoid affecting the subsequent deep learning feature extraction effect, and after comprehensive consideration, the label categories are divided into 12 categories, such as residential area, industrial area, forest, river, garden, farmland, bare land, and highway. After testing, 1 / 3 of the small size optical spot data is randomly selected for labeling, and 2 / 3 of the small size optical spot data does not need to be labeled. Among them, 1 / 3 of the small size optical spot data labeled is the first input data, which is used for subsequent input into the full convolution image segmentation network; 2 / 3 of the small size optical spot data without labeling is the second input data, which is used for subsequent input into the semi-supervised knowledge distillation network.

[0097] S3: constructing a full convolution image segmentation network;

[0098] In the embodiment of the application, the construction of the full convolution image segmentation network comprises the steps of:

[0099] using a two-dimensional convolution layer and a pooling layer to construct an encoder;

[0100] using a two-dimensional convolution layer to construct a transition layer;

[0101] using deconvolution and skip structure to construct a decoder;

[0102] constructing a first loss function.

[0103] Specifically, the encoder in the application is composed of 2 groups of encoder units 1 and 3 groups of encoder units 2, the encoder unit 1 is composed of two two-dimensional convolution layers and a pooling layer, and the encoder unit 2 is composed of three two-dimensional convolution layers and a pooling layer. The encoder outputs local feature 1, local feature 2, local feature 3, local feature 4, and local feature 5 in turn. The transition layer in the application is composed of three two-dimensional convolution layers. The decoder in the application is composed of two decoder units and a deconvolution, wherein the output result of the deconvolution in the first decoder unit is added to the local feature 4 (randomly cropped to be consistent with the size of the output result of the deconvolution) output by the encoder to form the skip structure skips in the first decoder unit, and the output result of the deconvolution in the second decoder unit is added to the local feature 3 (randomly cropped to be consistent with the size of the output result of the deconvolution) output by the encoder to form the skip structure skips in the second decoder unit. Increasing the skips structure combines the prediction of the last layer (with more global information) and the prediction of the shallower layer (with more local details), so that the network performs global prediction while performing local prediction. Since the training data of the full convolution image segmentation network, i.e., each optical spot data, is only partially labeled, the first loss function only calculates the pixels with labels, and its expression is:

[0104] ;

[0105] wherein i belongs to the pixels with labels.

[0106] S4: training the fully convolutional image segmentation network using the training data to obtain an optimized fully convolutional image segmentation network;

[0107] In the embodiments of the present application, the training of the fully convolutional image segmentation network using the training data to obtain an optimized fully convolutional image segmentation network comprises the steps of:

[0108] obtaining first input data with labels in the training data;

[0109] inputting the first input data into the encoder and the transition layer of the fully convolutional image segmentation network in sequence to obtain learning features;

[0110] inputting the learning features into the decoder of the fully convolutional image segmentation network to obtain multi-dimensional features;

[0111] aligning the multi-dimensional features with label data to obtain optimized data;

[0112] optimizing the fully convolutional image segmentation network using the optimized data to obtain a fully convolutional image segmentation model.

[0113] Specifically, the first input data with labels (optical spot data) are input into the encoder and the transition layer in sequence, and a plurality of levels of features (a relatively shallow convolutional layer perceives a relatively small field, and learns some local region features (local features); a relatively deep convolutional layer has a relatively large field, and can learn more abstract features (global features)) are learned by using a multi-layer convolutional structure; then the learned features are input into the decoder to output 12-dimensional features, the decoder combines the global features and the local features to prevent the loss of detail information and avoid the outline becoming blurred and unable to accurately extract the edges of objects; finally, the output result of the decoder is aligned with the label data, the data without labels is ignored, and an optimized output fully convolutional image segmentation model is fed back.

[0114] S5: constructing a semi-supervised knowledge distillation network based on the optimized fully convolutional image segmentation network;

[0115] In the embodiments of the present application, the construction of the semi-supervised knowledge distillation network based on the optimized fully convolutional image segmentation network comprises the steps of:

[0116] obtaining a fully convolutional image segmentation model in the optimized fully convolutional image segmentation network;

[0117] using the fully convolutional image segmentation model as a teacher network;

[0118] constructing a student network using the fully convolutional image segmentation network;

[0119] constructing a second loss function.

[0120] Specifically, the full convolution image segmentation model in the optimized full convolution image segmentation network is loaded as a teacher network of the knowledge distillation network to predict the input data. Since the remote sensing image belongs to a natural image and has the characteristics of complex ground objects, the 12 self-defined labels in the application cannot cover all object types. In the application, the student network adopts the form of adding new categories to prevent hard segmentation of the network. After testing, adding 8 categories is the best. The student network is composed of an encoder, a transition layer and a decoder in the optimized full convolution image segmentation network. Due to the addition of categories, the output feature latitude of the last layer of the decoder is 20. The specific construction steps of the second loss function are as follows:

[0121] (1) Construct the target class loss function L local : Since the knowledge distillation network-teacher network will perform global 12-class label hard segmentation on the input data, that is, each pixel of the optical image patch data has a predicted label, and the label is [0-12) class; the knowledge distillation network-student network will perform global 20-class label prediction on the input data, and the label is [0-20) class; the minimization of the target class loss function is to force the output label of the student network to align with the output label of the teacher network, that is, the knowledge transfer in knowledge distillation.

[0122] (2) Construct the soft label loss function : Since the remote sensing image belongs to a natural image and has the characteristics of complex ground objects, manual labeling cannot cover all ground object types, so the student network adds 8 categories to represent other ground object element categories. The spectral clustering-norm cut criterion is used as the loss function of the student network (20-class label) image segmentation, that is, the global 20-class label is maximized in class distance and minimized in class distance by learning the spatial proximity relationship and spectral similarity of the optical image.

[0123] (3) Construct the total loss function : The total loss function combines the target class loss function and the soft label loss function. The goal is to learn the spatial proximity relationship and spectral similarity of the optical image, learn the knowledge of the teacher network, perform 20-class segmentation on the optical image, improve the accuracy of the target 12 classes, and cluster the unlabeled pixels to improve the generalization ability of the model. The total loss function provides a label basis for subsequent new categories. The total loss function is defined as: .

[0124] S6: training the semi-supervised knowledge distillation network using the training data.

[0125] In the embodiments of the application, the training of the semi-supervised knowledge distillation network using the training data includes the following steps:

[0126] obtain second input data without a label in the training data;

[0127] input the second input data into a teacher network of the semi-supervised knowledge distillation network for prediction segmentation;

[0128] input the second input data into a student network of the semi-supervised knowledge distillation network for extracting multi-level features;

[0129] segment the second input data into (K+M) classes;

[0130] align a first output result of the teacher network and a second output result of the student network to calculate a target class loss function;

[0131] perform spectral clustering-norm cut criterion on the second input data according to the second output result to calculate a soft label loss function;

[0132] combine the target class loss function and the soft label loss function to perform feedback optimization on the student network and output a student network segmentation model.

[0133] Specifically, the second input data without a label (optical spot data) is input into the teacher network, and the optical spot data without a label is predicted and segmented into 12 classes by using the knowledge of the teacher network; then the second input data without a label (optical spot data) is input into the student network to extract multi-level features, and the optical spot data is segmented into (K+M) classes (such as 20 classes); the label output by the teacher network is aligned with the label output by the student network to calculate a target class loss function, and the spectral clustering-norm cut criterion is performed on the optical spot data according to the label output by the student network to calculate a soft label loss function, and the student network is optimized by combining the two loss functions and outputting a student network segmentation model.

[0134] As Figure 2 , the present application provides a semi-supervised satellite image semantic segmentation network construction device, comprising:

[0135] An optical remote sensing image collection module 10 is configured to collect optical remote sensing images.

[0136] An optical remote sensing image preprocessing module 20 is configured to preprocess the optical remote sensing images to obtain training data.

[0137] A full convolution image segmentation network construction module 30 is configured to construct a full convolution image segmentation network.

[0138] An optimized full convolution image segmentation network generation module 40 is configured to train the full convolution image segmentation network using the training data to obtain an optimized full convolution image segmentation network.

[0139] The network constructing semi-supervised knowledge distillation network segmentation module 50 is configured to construct a semi-supervised knowledge distillation network based on the optimized full convolutional image segmentation network.

[0140] The semi-supervised knowledge distillation network training module 60 is configured to train the semi-supervised knowledge distillation network using the training data.

[0141] The semi-supervised satellite image semantic segmentation network constructing device provided in the present application can be used in the semi-supervised satellite image semantic segmentation network constructing method provided in the present application.

[0142] Reference will be made to the following description Figure 3 , which shows a structural schematic diagram of an electronic device 100 suitable for implementing the embodiments of the present disclosure. The electronic device in the embodiments of the present disclosure can include, but is not limited to, a mobile terminal such as a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a PMP (Portable Multimedia Player), a vehicle terminal (such as a vehicle navigation terminal), and the like, and a fixed terminal such as a digital TV, a desktop computer, and the like. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functions and use range of the embodiments of the present disclosure.

[0143] As shown in Figure 3 , the electronic device 100 can include a processing device (such as a central processor, a graphics processor, etc.) 101, which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 102 or loaded into a random access memory (RAM) 103 from a storage device 108. In the RAM 103, various programs and data required for the operation of the electronic device 100 are also stored. The processing device 101, the ROM 102, and the RAM 103 are connected to each other through a bus 104. An input / output (I / O) interface 105 is also connected to the bus 104.

[0144] Generally, the following devices can be connected to the I / O interface 105: an input device 106 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, and the like; an output device 107 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, and the like; a storage device 108 including, for example, a magnetic tape, a hard disk, and the like; and a communication device 109. The communication device 109 can allow the electronic device 100 to communicate with other devices wirelessly or by wire to exchange data. Although the electronic device 100 having various devices is shown in the figure, it should be understood that it is not required to implement or have all the devices shown. More or fewer devices can be alternatively implemented or had.

[0145] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for performing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication device 109, or installed from the storage device 108, or installed from the ROM 102. When the computer program is executed by the processing device 101, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.

[0146] The semi-supervised satellite image semantic segmentation network construction method and device and electronic equipment provided by the application extract label sample knowledge through a deep learning teacher network, and transfer the knowledge to a student network, so that the prediction accuracy of the labeled sample category is high; the accuracy of the target category is improved, and the model generalization ability is improved.

[0147] It should be noted that, in this document, relational terms such as“first” and“second”, and the like, are used solely to distinguish one entity or action from another entity or action, without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms“comprises”,“comprising”, or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by“comprises a...” does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0148] The above description is merely that of specific embodiments of the present application, making it possible for those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the present application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a semi-supervised satellite image semantic segmentation network, characterized in that, The method includes the following steps: Collect optical remote sensing images; The optical remote sensing images are preprocessed to obtain training data; Construct a fully convolutional image segmentation network; The fully convolutional image segmentation network is trained using the training data to obtain an optimized fully convolutional image segmentation network. A semi-supervised knowledge distillation network is constructed based on the optimized fully convolutional image segmentation network. The semi-supervised knowledge distillation network is trained using the training data; Training the semi-supervised knowledge distillation network using the training data includes the following steps: Obtain the second input data that is not labeled from the training data; The second input data is input into the teacher network of the semi-supervised knowledge distillation network for predictive segmentation; The second input data is input into the student network of the semi-supervised knowledge distillation network to extract multi-level features; The second input data is divided into K+M classes; Align the first output of the teacher network with the second output of the student network to compute the target class loss function; Based on the second output result, the second input data is subjected to spectral clustering-normal cut criterion to calculate the soft label loss function; The student network is optimized by combining the target class loss function and the soft label loss function, and a student network segmentation model is output.

2. The method for constructing a semi-supervised satellite image semantic segmentation network according to claim 1, characterized in that, The process of collecting optical remote sensing images includes the following steps: Collect optical remote sensing images from different regions and in different months; Determine whether the cloud cover in the optical remote sensing image is greater than a preset threshold; If so, discard the optical remote sensing image; If not, retain the optical remote sensing image; Acquire all remaining optical remote sensing images.

3. The method for constructing a semi-supervised satellite image semantic segmentation network according to claim 1, characterized in that, The preprocessing of the optical remote sensing image to obtain training data includes the following steps: Atmospheric correction is performed on the optical remote sensing image; Perform geometric correction on the atmospherically corrected optical remote sensing image; The geometrically corrected optical remote sensing image is cropped to obtain smaller optical patch data. The small-sized optical patch data is locally marked.

4. The method for constructing a semi-supervised satellite image semantic segmentation network according to claim 3, characterized in that, The process of locally marking the small-sized optical patch data includes the following steps: Select the label category; The small-sized optical patch data is randomly divided into first patch data and second patch data; The first patch data is labeled using the aforementioned label categories; Use the marked first patch data as the first input data; Use the second patch data as the second input data.

5. The method for constructing a semi-supervised satellite image semantic segmentation network according to claim 3, characterized in that, The construction of the fully convolutional image segmentation network includes the following steps: The encoder is constructed using two-dimensional convolutional and pooling layers; Use two-dimensional convolutional layers to construct transition layers; The decoder is constructed using deconvolution and skip-level structures; Construct the first loss function.

6. The method for constructing a semi-supervised satellite image semantic segmentation network according to claim 1, characterized in that, The step of training the fully convolutional image segmentation network using the training data to obtain an optimized fully convolutional image segmentation network includes the following steps: Obtain the first labeled input data from the training data; The first input data is sequentially input into the encoder and transition layer of the fully convolutional image segmentation network to obtain learned features; The learned features are input into the decoder of the fully convolutional image segmentation network to obtain multidimensional features; Align the multidimensional features with the label data to obtain optimized data; The optimized data is used to optimize the fully convolutional image segmentation network to obtain a fully convolutional image segmentation model.

7. The method for constructing a semi-supervised satellite image semantic segmentation network according to claim 1, characterized in that, The process of constructing a semi-supervised knowledge distillation network based on the optimized fully convolutional image segmentation network includes the following steps: Obtain the fully convolutional image segmentation model in the optimized fully convolutional image segmentation network; The fully convolutional image segmentation model is used as the teacher network; The student network was constructed using the fully convolutional image segmentation network described above. Construct a second loss function.

8. A semi-supervised satellite image semantic segmentation network construction apparatus, applicable to the method described in any one of claims 1 to 7, characterized in that, include: Optical remote sensing image acquisition module, used to collect optical remote sensing images; An optical remote sensing image preprocessing module is used to preprocess the optical remote sensing image to obtain training data; The fully convolutional image segmentation network building module is used to construct fully convolutional image segmentation networks; An optimized fully convolutional image segmentation network generation module is used to train the fully convolutional image segmentation network using the training data to obtain an optimized fully convolutional image segmentation network. A semi-supervised knowledge distillation network segmentation module is used to construct a semi-supervised knowledge distillation network based on the optimized fully convolutional image segmentation network. A semi-supervised knowledge distillation network training module is used to train the semi-supervised knowledge distillation network using the training data.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the semi-supervised satellite image semantic segmentation network construction method according to any one of claims 1 to 7.

Images