Remote sensing image recognition deep neural network training and recognition method

Abstract

The embodiment of the invention provides a remote sensing image recognition deep neural network training and recognition method, and the method comprises the steps: obtaining a remote sensing image which comprises land types and the coverage area data of each land type; the remote sensing image is divided into a plurality of sampling units with fixed sizes, the land types of the sampling units and the coverage area of each land type are obtained, the plurality of sampling units form a sampling area, and all the sampling areas cover the remote sensing image; according to the land type of the sampling unit and the coverage area of each land type, and according to a predetermined conditional entropy calculation rule, obtaining a conditional entropy of the sampling unit; according to the conditional entropy values of the sampling units, selecting the sampling units as small and weak samples from the sampling areas according to a preset small and weak sample selection rule; rGB spectral information of the small and weak samples is acquired, and the remote sensing image recognition deep neural network is trained according to the small and weak samples. Therefore, the embodiment of the invention can improve the training efficiency of the deep neural network for the remote sensing image land type recognition task.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of remote sensing image recognition. Background Art

[0002] Land type recognition technology refers to the technology of taking images of the target area and identifying the target land type in the target area based on the image. For example, the distribution of forests in the target area can be identified based on the image. Rapid and accurate detection of land type distribution in the target area is crucial for urban and regional planning and resource surveys.

[0003] Existing land type recognition technologies include low- and medium-space drone or aircraft aerial images, and high-precision remote sensing images taken by high-space satellites. Among them, the land type recognition task of high-space remote sensing images usually requires a lot of time to extract features and select samples from images to train remote sensing image land recognition models, resulting in low efficiency of the entire land type recognition task.

[0004] Therefore, how to improve the training efficiency of deep neural networks used for land type recognition tasks in remote sensing images has become an urgent problem that technicians in this field need to solve. Summary of the invention

[0005] To solve the above problems, an embodiment of the present invention provides a training method for a remote sensing image recognition deep neural network, comprising:

[0006] Acquire a remote sensing image, wherein the remote sensing image includes land types and coverage area data of each land type;

[0007] Dividing the remote sensing image into a plurality of sampling units and obtaining the land types of the sampling units and the coverage area of ​​each land type, wherein the sampling unit has a fixed size, a plurality of sampling units form a sampling area, and all sampling areas cover the remote sensing image;

[0008] According to the land type of the sampling unit and the coverage area of ​​each land type, the conditional entropy value of the sampling unit is obtained according to a predetermined conditional entropy value calculation rule;

[0009] According to the conditional entropy value of the sampling unit and the predetermined small weak sample selection rule, the sampling unit is selected from each of the sampling areas as a small weak sample;

[0010] RGB spectrum information of small weak samples is obtained, and a remote sensing image recognition deep neural network is trained according to the small weak samples.

[0011] Optionally, the obtaining of the conditional entropy value of the sampling unit according to the land type of the sampling unit and the coverage area of ​​each land type according to the conditional entropy value calculation rule includes: calculating the conditional entropy value of each sampling unit by a conditional entropy value calculation formula, and the conditional entropy value calculation formula is:

[0012]

[0013] For land categories whose area percentage in remote sensing images is lower than the first threshold, p(i) is calculated by For land categories with an area percentage higher than the second threshold, p(i) is calculated by Calculation, when the area ratio is between the first threshold and the second threshold, p(i) is calculated by calculate;

[0014] Where J represents the sampling unit; n represents the number of land categories considered; i represents the specific land category; P(i) represents the probability that land category i is selected; Ji is the area of ​​land feature category i in sampling unit J; Maxi is the maximum area of ​​category i in the sampling unit within the sampling area where sampling unit J is located; Min i It is the minimum area of ​​category i in the sampling unit within the sampling area where sampling unit J is located; Avgi = Si / m, where m represents the number of sampling units contained in each sampling area, and Si represents the total area of ​​land category i in the remote sensing image.

[0015] Optionally, the first threshold comprises 1.2 of the total number of land categories, and the second threshold comprises 1.5 of the total number of land categories.

[0016] Optionally, selecting sampling units as small weak samples from each of the sampling areas according to the conditional entropy value of the sampling unit and a predetermined small weak sample selection rule includes:

[0017] The sampling unit with the largest conditional entropy is selected from each sampling area.

[0018] Optionally, the acquiring RGB spectrum information of small weak samples and training a remote sensing image recognition deep neural network according to the small weak samples includes:

[0019] Obtaining RGB spectrum information of the small weak sample;

[0020] Label the land type of the small weak sample and the coverage area of ​​each land type, wherein the label includes a target land type label and a non-target land type label;

[0021] A remote sensing image recognition deep neural network is trained based on the RGB spectral information and labels of the small weak samples.

[0022] Optionally, the training of a remote sensing image recognition deep neural network according to the RGB spectrum information and labels of the small weak samples includes:

[0023] Divide small weak samples into training set and validation set;

[0024] The batch size of the deep learning model is set to 16, the epoch is set to 100, the descent strategy is stochastic gradient descent, the learning rate is set to 0.0001 for the first 50 epochs, and the learning rate is set to 0.00001 for the next 50 epochs. Before each epoch training, the training data is randomly shuffled and then input, and the ResNet-50 module pre-trained on ImageNet is configured as the residual module of the model;

[0025] Train the deep learning model using the training set;

[0026] The IOU accuracy of the deep learning model for the validation set is calculated and recorded. The deep learning model with the highest IOU accuracy for the validation set is used as the final trained deep neural network for remote sensing image recognition.

[0027] Optionally, labeling according to the land type of the small weak sample and the coverage area of ​​the land type includes:

[0028] Experts label the land types and coverage areas of some small and weak samples, including target land type labels and non-target land type labels, and obtain texture features of some small and weak samples;

[0029] Using the partial small weak samples, texture features and labels of the small weak samples, a random forest classification model is trained;

[0030] The labels of the remaining small and weak samples are obtained through the trained random forest classification model; and the remote sensing image recognition deep neural network is trained according to the RGB spectral information and labels of the small and weak samples.

[0031] Optionally, the training of a remote sensing image recognition deep neural network according to the texture features, RGB spectrum information and labels of the small weak samples includes:

[0032] Preprocessing the small weak sample;

[0033] The deep neural network for remote sensing image recognition is trained based on the texture features, RGB spectral information and labels of small and weak samples after preprocessing.

[0034] Optionally, the acquiring of remote sensing images, wherein the remote sensing images include land types and coverage area data of each land type, includes: acquiring Globalland30 land cover classification data;

[0035] The obtaining of RGB spectrum information of the small weak sample includes: obtaining Google satellite images.

[0036] The present application also provides a method for identifying target land type based on a remote sensing image recognition deep neural network, including:

[0037] Acquire a remote sensing image, wherein the remote sensing image includes RGB spectral information;

[0038] According to the remote sensing image, identifying the target land type through a pre-trained remote sensing image recognition deep neural network;

[0039] The training method of the pre-trained remote sensing image recognition deep neural network includes:

[0040] Acquire a remote sensing image, wherein the remote sensing image includes land types and coverage area data of each land type;

[0041] Dividing the remote sensing image into a plurality of sampling units and obtaining the land types of the sampling units and the coverage area of ​​each land type, wherein the sampling unit has a fixed size, a plurality of sampling units form a sampling area, and all sampling areas cover the remote sensing image;

[0042] According to the land type of the sampling unit and the coverage area of ​​each land type, the conditional entropy value of the sampling unit is obtained according to a predetermined conditional entropy value calculation rule;

[0043] According to the conditional entropy value of the sampling unit and the predetermined small weak sample selection rule, the sampling unit is selected from each of the sampling areas as a small weak sample;

[0044] RGB spectrum information of small weak samples is obtained, and a remote sensing image recognition deep neural network is trained according to the small weak samples.

[0045] It can be seen that the training method and recognition method of the remote sensing image recognition deep neural network provided by the embodiment of the present invention, after acquiring the remote sensing image, divide the remote sensing image into a plurality of sampling units and a sampling area composed of a plurality of sampling units, select sampling units from each of the sampling areas as small weak samples according to the land type of the sampling unit and the coverage area of ​​each land type, according to the predetermined conditional entropy value calculation rule and the predetermined sample selection rule, and train the remote sensing image recognition deep neural network according to the small weak samples. In this way, the sampling units are selected by the predetermined conditional entropy value calculation rule, the more land types in the sampling unit, the greater the conditional entropy value, the more rare land types or the fewer common land types, the greater the conditional entropy value, so that the sampling units containing rich information can be selected from each sampling area as small weak samples, thereby ensuring the comprehensiveness and information richness of the small weak samples while reducing the number of samples, thereby shortening the sample preparation and processing time for model training, and improving the training efficiency of the deep neural network for the remote sensing image land type recognition task. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0047] Figure 1 It is a flow chart of a training method for a deep neural network for remote sensing image recognition provided by an embodiment of the present invention;

[0048] Figure 2 It is a schematic diagram of sampling units and sampling area division of a remote sensing image recognition deep neural network provided by an embodiment of the present invention;

[0049] Figure 3 It is a schematic diagram of a sampling unit selection of a deep neural network for remote sensing image recognition provided by an embodiment of the present invention;

[0050] Figure 4 A remote sensing image of a target land type recognition method based on a remote sensing image recognition deep neural network provided by an embodiment of the present invention;

[0051] Figure 5 This is a recognition effect diagram of a target land type recognition method based on a remote sensing image recognition deep neural network provided by an embodiment of the present invention;

[0052] Figure 6 It is a schematic diagram of a training method and a recognition method of a deep neural network for remote sensing image recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Please refer to Figure 1 , Figure 1 FIG. 1 is a flow chart of a training method for a deep neural network for remote sensing image recognition provided by an embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a training method for a remote sensing image recognition deep neural network, comprising:

[0055] In step S1, a remote sensing image is acquired, wherein the remote sensing image includes land types and coverage area data of each land type.

[0056] Please refer to the remote sensing images obtained Figure 2 , Figure 2 Schematic diagram of a remote sensing image in a remote sensing image recognition deep neural network training method provided by an embodiment of the present invention. As shown in the figure, the remote sensing image is a high-altitude high-resolution remote-controlled image, from which small weak samples for training the remote sensing image recognition deep neural network are extracted.

[0057] In step S2, the remote sensing image is divided into a plurality of sampling units and the land types of the sampling units and the coverage area of ​​each land type are obtained. The sampling unit has a fixed size, and a plurality of sampling units form a sampling area. All sampling areas cover the remote sensing image.

[0058] Please combine Figure 1 refer to Figure 2 , Figure 2 The present invention provides a schematic diagram of a sampling unit and sampling area division of a remote sensing image recognition deep neural network. As shown in the figure, the remote sensing image is divided into multiple areas, each area includes multiple sampling units of the same size. In some embodiments, each area contains the same number of sampling units.

[0059] In some embodiments, the division may be performed based on latitude and longitude. Specifically, a square surrounded by longitude and latitude of a fixed span is a sampling unit.

[0060] In step S3, the conditional entropy value of the sampling unit is obtained according to the land type of the sampling unit and the coverage area of ​​each land type and in accordance with a predetermined conditional entropy value calculation rule.

[0061] Land types include cultivated land, forest land, grassland, shrubland, wetland, water bodies, tundra, artificial surface, bare land, glacier and permanent snow.

[0062] The conditional entropy theory shows that the more chaotic the system is, the greater the conditional entropy value is, and the conditional entropy is lowest when something is easier or less likely to happen. Specifically in this application, the predetermined conditional entropy value calculation rule means that the more land types there are in the sampling unit, the greater the conditional entropy value is, and the more rare land types or fewer common land types are contained, the greater the conditional entropy value is. In this way, it can be ensured that the sampling unit with a large conditional entropy value contains more information, thereby reducing the number of small weak samples required for training the remote sensing image recognition deep neural network.

[0063] In order to further improve the training efficiency of the deep neural network for remote sensing image recognition, in some embodiments, only land types with an area higher than a predetermined threshold in the remote sensing image are considered, specifically, land types with an area higher than one thousandth. In this way, land types that do not have enough small weak samples for the deep learning model to learn can be directly removed in the early stage, reducing the workload in the early stage and further improving the training efficiency.

[0064] In one embodiment, the obtaining of the conditional entropy value of the sampling unit according to the land type of the sampling unit and the coverage area of ​​each land type according to the conditional entropy value calculation rule includes:

[0065] The conditional entropy value of each sampling unit is calculated by the conditional entropy value calculation formula, and the conditional entropy value calculation formula is:

[0066] For land categories whose area percentage in remote sensing images is lower than the first threshold, p(i) is calculated by For land categories with an area percentage higher than the second threshold, p(i) is calculated by Calculation, when the area ratio is between the first threshold and the second threshold, p(i) is calculated by calculate;

[0067] Where J represents the sampling unit; n represents the number of land types considered; i represents the specific land type; P(i) represents the probability of land type i being selected; Ji is the area of ​​land type i in sampling unit J; Max i is the maximum area of ​​category i in the sampling unit within the sampling area where sampling unit J is located; Min i It is the minimum area of ​​category i in the sampling unit within the sampling area where sampling unit J is located; Avgi = Si / m, where m represents the number of sampling units contained in each sampling area, and Si represents the total area of ​​land category i in the remote sensing image.

[0068] In this way, when the area proportion of land category i is lower than the first threshold, the closer the distribution area of ​​land category i in the sampling unit is to the maximum value, the greater its uncertainty is, and the larger the conditional entropy of the sampling unit is; when the area proportion of land category i is higher than the second threshold, the closer the distribution area of ​​land category i in the sampling unit is to the minimum value, the greater its uncertainty is, and the larger the conditional entropy of the sampling unit is; when the area proportion of land category i is between the two, the closer the distribution area of ​​land category i in the sampling unit is to the average value, the greater its uncertainty is, and the larger the conditional entropy of the sampling unit is. In some embodiments, the first threshold includes 1.2 of the total number of land categories, and the second threshold includes 1.5 of the total number of land categories, which can ensure the balance of small weak samples.

[0069] In step S4, sampling units are selected from each of the sampling areas as small weak samples according to the conditional entropy value of the sampling unit and a predetermined small weak sample selection rule.

[0070] Please combine Figure 1 refer to Figure 3 , Figure 3 It is a schematic diagram of sampling unit selection for a remote sensing image recognition deep neural network provided by an embodiment of the present invention. In some embodiments, the small weak sample selection rule may be to select sampling units with the top 20 percent of conditional entropy values ​​from each sampling area; in some embodiments, it may be to select sampling units with conditional entropy greater than a threshold from each sampling area; in other embodiments, as shown in the figure, it may be to select sampling units with the largest conditional entropy from each sampling area.

[0071] In step S5, RGB spectrum information of small weak samples is obtained, and a remote sensing image recognition deep neural network is trained based on the small weak samples.

[0072] The remote sensing image recognition deep neural network is a deep neural network used to identify the target land type in the remote sensing image, and the target land type refers to the remote sensing image recognition deep neural network. In some embodiments, the remote sensing image recognition deep neural network trained according to the small weak sample includes: U-net, U-net++, DeeplabV3, DeeplabV3+, PSPNet, FPN.

[0073] In order to further improve the training speed and efficiency of the deep neural network for remote sensing image recognition, it is necessary to set the hyperparameters of the deep learning model. In some embodiments, the batch size of the deep learning model can be set to 16, the epoch is set to 100, the descent strategy is stochastic gradient descent, the learning rate of the first 50 epochs is set to 0.0001, and the learning rate of the last 50 epochs is set to 0.00001. Before each epoch training, the training data is randomly shuffled and then input, and the ResNet-50 module based on ImageNet pre-training is configured as the residual module of the model. A deep learning model will be saved in each epoch. When the deep learning model is trained, the IOU accuracy of the deep learning model saved in each epoch for the validation set recognition is calculated and recorded, and the deep learning model with the highest IOU accuracy for the validation set recognition is selected as the final trained remote sensing image recognition deep neural network. Among them, Batch size determines the number of small weak samples processed simultaneously in each model training iteration. It is usually necessary to experiment and adjust to find the best Batch Size value to meet the needs of specific tasks. Here, 16 is the best after testing. Epoch refers to the process of all data passing through the neural network model through a complete forward propagation and back propagation. As the number of epochs increases, the number of weight update iterations increases, and the curve goes from the initial non-fitting state to the optimized fitting state, and finally to overfitting. The size of the epoch is related to the diversity of the data set. Generally, the greater the diversity, the larger the epoch. The model usually contains a large number of parameters, and the gradient descent algorithm can automatically adjust the values ​​of these parameters, so that the model can predict the data more accurately. Stochastic gradient descent refers to randomly selecting a small weak sample or a small batch of small weak samples from the training data set each time, calculating the gradient of the loss function with respect to the model parameters, and then updating the model parameters in the direction of the negative gradient, and repeating this process until the model parameters converge or the preset number of iterations is reached.

[0074] In some embodiments, the training of a remote sensing image recognition deep neural network based on the small weak samples includes:

[0075] In step S51, the RGB spectrum information of the small weak sample is obtained.

[0076] In step S52, labels are marked according to the land types of small weak samples and the coverage area of ​​each land type, wherein the labels include target land type labels and non-target land type labels.

[0077] Specifically, the target land type may be forest land.

[0078] In some embodiments, an expert may determine whether a small weak sample is a target land type by analyzing the land type of each small weak sample and the coverage area of ​​each land type; in other embodiments, to further improve the training efficiency of the model, the labeling according to the land type of the small weak sample and the coverage area of ​​the land type includes:

[0079] In step S521, experts label the land types and coverage areas of some small weak samples, including target land type labels and non-target land type labels, and obtain texture features of the small weak samples.

[0080] The texture features include mean, variance, conditional entropy, homogeneity, contrast, correlation, dissimilarity, second-order moment and other features. In some embodiments, the texture features can be calculated for the sampling unit using the gray level co-occurrence matrix method with a texture window size of 11*11. Specifically, the texture feature calculation tool of the envi software can be used. After the calculation is completed, the RGB spectrum information and texture information of the Google image are superimposed together to form multiple bands as the feature set input. In this way, by extracting the texture features of small weak samples, the accuracy of supervised classification can be improved.

[0081] In step S522, the random forest classification model is trained using the partial small weak samples, texture features of the small weak samples and labels.

[0082] In step S523, labels of the remaining small weak samples are obtained through the trained random forest classification model.

[0083] In this way, only some labels need to be annotated by experts, and the labels of the remaining small and weak samples can be obtained through the random forest classification model, which further reduces the manual workload and improves the efficiency of neural network training for land type recognition in remote sensing images.

[0084] In step S53, a remote sensing image recognition deep neural network is trained according to the RGB spectrum information and labels of the small weak samples.

[0085] In some embodiments, the training of a remote sensing image recognition deep neural network based on the RGB spectrum information and labels of the small weak samples includes:

[0086] In step S531, the small weak samples are preprocessed.

[0087] The preprocessing includes remote sensing image cropping, normalization and small weak sample enhancement. The remote sensing image cropping includes cropping according to 256*256 or 512*512 pixel size, and randomly dividing small weak samples into training sets and validation sets according to a certain ratio. Specifically, 70% of small weak samples can be divided into training sets and 30% of small weak samples can be divided into validation sets at a ratio of 7:3. The normalization refers to normalizing small weak samples, which can accelerate model convergence, improve model performance and stability, and avoid gradient disappearance or explosion. The small weak sample enhancement refers to performing random rotation, flipping, scaling and other operations on small weak samples to increase data diversity and enhance the generalization ability of the model.

[0088] In step S532, a remote sensing image recognition deep neural network is trained based on the RGB spectral information and labels of the preprocessed small weak samples.

[0089] It can be seen that the training method for remote sensing image recognition deep neural network provided by the embodiment of the present invention, after acquiring the remote sensing image, divides the remote sensing image into a plurality of sampling units and a sampling area composed of a plurality of sampling units, selects sampling units as small weak samples from each sampling area according to the land type of the sampling unit and the coverage area of ​​each land type, according to the predetermined conditional entropy value calculation rule and the predetermined small weak sample selection rule, and trains the remote sensing image recognition deep neural network according to the small weak samples. In this way, the sampling units are selected by the predetermined conditional entropy value calculation rule, the more land types there are in the sampling unit, the greater the conditional entropy value, the more rare land types there are or the fewer common land types there are, the greater the conditional entropy value, so that the sampling units containing rich information can be selected from each sampling area as small weak samples, thereby ensuring the comprehensiveness and information richness of the small weak samples while reducing the number of samples required, thereby shortening the sample preparation and processing time for model training, and improving the training efficiency of the deep neural network for the remote sensing image land type recognition task.

[0090] In some embodiments, the acquiring of remote sensing images, which include land types and coverage area data of each land type, includes: acquiring Globalland30 land cover classification data; and the acquiring of RGB spectral information of small weak samples includes: acquiring Google satellite images.

[0091] In this way, by utilizing the two major public data sets of Globalland30 land cover classification data and Google satellite images, it is possible to easily and quickly obtain the land types required for training and the coverage area data of each land type, as well as the RGB spectral information of small weak samples, thereby further improving the training efficiency of the remote sensing image recognition deep neural network provided in the embodiment of the present application.

[0092] The present invention also provides a method for identifying target land type based on remote sensing image recognition deep neural network, please refer to Figure 4-Figure 6 , Figure 4 A remote sensing image is provided in an embodiment of the present invention in a method for identifying a target land type based on a remote sensing image recognition deep neural network. Figure 5 This is a recognition effect diagram of a target land type recognition method based on a remote sensing image recognition deep neural network provided by an embodiment of the present invention. Figure 6 It is a schematic diagram of a training method and a recognition method of a deep neural network for remote sensing image recognition provided by an embodiment of the present invention.

[0093] The target land type recognition method based on remote sensing image recognition deep neural network provided by the embodiment of the present invention includes:

[0094] Acquire a remote sensing image, wherein the remote sensing image includes RGB spectral information;

[0095] According to the remote sensing image, identifying the target land type through a pre-trained remote sensing image recognition deep neural network;

[0096] The training method of the pre-trained remote sensing image recognition deep neural network includes:

[0097] Acquire a remote sensing image, wherein the remote sensing image includes land types and coverage area data of each land type;

[0098] Dividing the remote sensing image into a plurality of sampling units and obtaining the land types of the sampling units and the coverage area of ​​each land type, wherein the sampling unit has a fixed size, a plurality of sampling units form a sampling area, and all sampling areas cover the remote sensing image;

[0099] According to the land type of the sampling unit and the coverage area of ​​each land type, the conditional entropy value of the sampling unit is obtained according to a predetermined conditional entropy value calculation rule;

[0100] According to the conditional entropy value of the sampling unit and the predetermined small weak sample selection rule, the sampling unit is selected from each of the sampling areas as a small weak sample;

[0101] RGB spectrum information of small weak samples is obtained, and a remote sensing image recognition deep neural network is trained according to the small weak samples.

[0102] It can be seen that the target land type identification method based on remote sensing image recognition deep neural network provided by the embodiment of the present invention can quickly identify the target land type.

[0103] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein. The above describes multiple embodiment schemes provided by the embodiments of the present invention, and the various optional methods introduced in the various embodiment schemes may be combined and cross-referenced with each other without conflict, thereby extending a variety of possible embodiment schemes, which may all be considered as the embodiment schemes disclosed and disclosed by the embodiments of the present invention.

[0104] Although the embodiments of the present invention are disclosed above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A training method for a deep neural network for remote sensing image recognition, characterized in that: include: Acquire a remote sensing image, wherein the remote sensing image includes land types and coverage area data of each land type; Dividing the remote sensing image into a plurality of sampling units and obtaining the land types of the sampling units and the coverage area of ​​each land type, wherein the sampling unit has a fixed size, a plurality of sampling units form a sampling area, and all sampling areas cover the remote sensing image; According to the land type of the sampling unit and the coverage area of ​​each land type, the conditional entropy value of the sampling unit is obtained according to a predetermined conditional entropy value calculation rule; According to the conditional entropy value of the sampling unit and the predetermined small weak sample selection rule, the sampling unit is selected from each of the sampling areas as a small weak sample; RGB spectrum information of small weak samples is obtained, and a remote sensing image recognition deep neural network is trained according to the small weak samples.

2. The method for training a deep neural network for remote sensing image recognition according to claim 1, characterized in that: The step of obtaining the conditional entropy value of the sampling unit according to the land type of the sampling unit and the coverage area of ​​each land type and according to the conditional entropy value calculation rule includes: calculating the conditional entropy value of each sampling unit by a conditional entropy value calculation formula, wherein the conditional entropy value calculation formula is: For land categories whose area percentage in remote sensing images is lower than the first threshold, p(i) is calculated by For land categories with an area percentage higher than the second threshold, p(i) is calculated by Calculation, when the area ratio is between the first threshold and the second threshold, p(i) is calculated by calculate; Where J represents the sampling unit; n represents the number of land types considered; i represents the specific land type; P(i) represents the probability of land type i being selected; Ji is the area of ​​land type i in sampling unit J; Max i is the maximum area of ​​category i in the sampling unit within the sampling area where sampling unit J is located; Min i It is the minimum area of ​​category i in the sampling unit within the sampling area where sampling unit J is located; Avgi = Si / m, where m represents the number of sampling units contained in each sampling area, and Si represents the total area of ​​land category i in the remote sensing image.

3. The training method for remote sensing image recognition deep neural network according to claim 2, characterized in that: The first threshold comprises 1.2 of the total number of land categories, and the second threshold comprises 1.5 of the total number of land categories.

4. The method for training a deep neural network for remote sensing image recognition according to claim 2, wherein: The step of selecting sampling units as small weak samples from each of the sampling areas according to the conditional entropy value of the sampling unit and a predetermined small weak sample selection rule includes: The sampling unit with the largest conditional entropy is selected from each sampling area.

5. The method for training a deep neural network for remote sensing image recognition according to claim 2, wherein: The step of obtaining RGB spectrum information of a small weak sample and training a remote sensing image recognition deep neural network according to the small weak sample includes: Obtaining RGB spectrum information of the small weak sample; Label the land type of the small weak sample and the coverage area of ​​each land type, wherein the label includes a target land type label and a non-target land type label; A remote sensing image recognition deep neural network is trained based on the RGB spectral information and labels of the small weak samples.

6. The method for training a deep neural network for remote sensing image recognition according to claim 5, characterized in that: The method of training a remote sensing image recognition deep neural network according to the RGB spectrum information and labels of the small weak samples includes: Divide small weak samples into training set and validation set; The batch size of the deep learning model is set to 16, the epoch is set to 100, the descent strategy is stochastic gradient descent, the learning rate is set to 0.0001 for the first 50 epochs, and the learning rate is set to 0.00001 for the next 50 epochs. Before each epoch training, the training data is randomly shuffled and then input, and the ResNet-50 module pre-trained on ImageNet is configured as the residual module of the model; Train the deep learning model using the training set; The IOU accuracy of the deep learning model for the validation set is calculated and recorded. The deep learning model with the highest IOU accuracy for the validation set is used as the final trained deep neural network for remote sensing image recognition.

7. The method for training a deep neural network for remote sensing image recognition according to claim 5, characterized in that: The labeling according to the land type and the coverage area of ​​the land type of the small weak sample includes: Experts label the land types and coverage areas of some small and weak samples, including target land type labels and non-target land type labels, and obtain texture features of some small and weak samples; Using the partial small weak samples, texture features and labels of the small weak samples, a random forest classification model is trained; The labels of the remaining small and weak samples are obtained through the trained random forest classification model; and the remote sensing image recognition deep neural network is trained according to the RGB spectral information and labels of the small and weak samples.

8. The method for training a deep neural network for remote sensing image recognition according to claim 7, wherein: The method of training a remote sensing image recognition deep neural network according to the texture features, RGB spectrum information and labels of the small weak samples includes: Preprocessing the small weak sample; The deep neural network for remote sensing image recognition is trained based on the texture features, RGB spectral information and labels of small and weak samples after preprocessing.

9. The method for training a deep neural network for remote sensing image recognition according to any one of claims 1 to 8, characterized in that: The acquiring of remote sensing images, wherein the remote sensing images include land types and coverage area data of each land type, includes: acquiring Globalland30 land cover classification data; The obtaining of RGB spectrum information of the small weak sample includes: obtaining Google satellite images.

10. A method for identifying target land types based on remote sensing image recognition deep neural network, characterized in that: include: Acquire a remote sensing image, wherein the remote sensing image includes RGB spectral information; According to the remote sensing image, identifying the target land type through a pre-trained remote sensing image recognition deep neural network; The training method of the pre-trained remote sensing image recognition deep neural network includes: Acquire a remote sensing image, wherein the remote sensing image includes land types and coverage area data of each land type; Dividing the remote sensing image into a plurality of sampling units and obtaining the land types of the sampling units and the coverage area of ​​each land type, wherein the sampling unit has a fixed size, a plurality of sampling units form a sampling area, and all sampling areas cover the remote sensing image; According to the land type of the sampling unit and the coverage area of ​​each land type, the conditional entropy value of the sampling unit is obtained according to a predetermined conditional entropy value calculation rule; According to the conditional entropy value of the sampling unit and the predetermined small weak sample selection rule, the sampling unit is selected from each of the sampling areas as a small weak sample; RGB spectrum information of small weak samples is obtained, and a remote sensing image recognition deep neural network is trained according to the small weak samples.

Images