Configuration method and device of satellite soil moisture data reconstruction model

Abstract

The invention provides a configuration method and device of a satellite soil moisture data reconstruction model, which are used for introducing a generative adversarial network model to build the soil moisture data reconstruction model under the condition of reconstructing soil moisture data observed by a satellite product based on a deep learning technology. The method mainly starts with a loss function involved in the training process to further enhance the reconstruction performance of the model on soil moisture data, so that the prediction effect of the model can be kept continuous in space, and dynamic change can be accurately reflected in time, thereby meeting the high-quality soil moisture observation requirement.

Description

Technical Field

[0001] The present application relates to the field of soil moisture detection, and in particular to a configuration method and device for a satellite soil moisture data reconstruction model. Background Art

[0002] Soil moisture, as a key factor in the land surface water cycle, plays an important role in energy conversion and water cycle. Satellite remote sensing technology, with its advantages of large-scale and continuous monitoring, provides a new means for soil moisture monitoring. At present, the main remote sensing bands used for remote sensing monitoring of soil moisture are visible light, near infrared, thermal infrared and microwave. Among them, passive microwave remote sensing is less affected by the shape and structure of the ground objects, and has the advantages of short revisit period, long time series and wide coverage. It is the main means of large-scale monitoring of soil moisture. Most of the existing soil moisture remote sensing products are based on passive microwave remote sensing inversion products.

[0003] However, due to the limitations of satellite orbital clearance and payload detection capabilities, satellite soil moisture products have missing information on the daily scale. This high proportion of data missing not only poses a challenge to the continuity and integrity of soil moisture inversion, but also greatly limits the subsequent use of soil moisture products. The reconstruction of missing values ​​of satellite soil moisture data has become a task that needs to be urgently solved. Currently, commonly used methods include interpolation, machine learning, and deep learning.

[0004] However, the inventors of the present application have discovered that, based on satellite remote sensing technology, the existing soil moisture data reconstruction scheme based on deep learning methods still has limitations in actual performance in terms of reconstruction accuracy, and it is difficult to meet the needs of high-quality soil moisture observation. Summary of the invention

[0005] The present application provides a configuration method and device for a satellite soil moisture data reconstruction model, which is used to reconstruct the soil moisture data observed by satellite products based on deep learning technology. A generative adversarial network model is introduced to create a soil moisture data reconstruction model, and the loss function involved in the training process is focused on to further enhance its reconstruction performance for soil moisture data, so that the prediction effect of the model can maintain continuity in space and accurately reflect dynamic changes in time, thereby meeting the needs of high-quality soil moisture observation.

[0006] In a first aspect, the present application provides a method for configuring a satellite soil moisture data reconstruction model, the method comprising: Acquire first sample satellite soil moisture observation data, wherein the sample satellite soil moisture observation data is soil moisture observation data obtained from satellite observation products; The first sample satellite soil moisture observation data is configured with a corresponding mask and annotation to obtain the second sample satellite soil moisture observation data, wherein the mask is used to shield part of the data content to form the missing content to be reconstructed, and the annotation is the soil moisture observation data obtained by observing the ground observation station; Based on the second sample satellite soil moisture observation data, a soil moisture data reconstruction model is trained, wherein the soil moisture data reconstruction model is used to reconstruct the satellite soil moisture observation data input into the model. The soil moisture data reconstruction model is specifically a three-dimensional generative adversarial network model, which includes a generator and a discriminator. During the model training process, the loss function corresponding to the generator is specifically configured based on the sum of the absolute value of the error of each pixel point, and the loss function corresponding to the discriminator is specifically configured based on the discriminant loss and the gradient penalty term.

[0007] In a second aspect, the present application provides a configuration device for a satellite soil moisture data reconstruction model, the device comprising: An acquisition unit is used to acquire first sample satellite soil moisture observation data, wherein the sample satellite soil moisture observation data is soil moisture observation data obtained from satellite observation products; A configuration unit is used to configure a corresponding mask and annotation for the first sample satellite soil moisture observation data to obtain second sample satellite soil moisture observation data, wherein the mask is used to shield part of the data content to form the missing content to be reconstructed, and the annotation is the soil moisture observation data obtained by observing the ground observation station; A training unit is used to train a soil moisture data reconstruction model based on the second sample satellite soil moisture observation data, wherein the soil moisture data reconstruction model is used to reconstruct the satellite soil moisture observation data input into the model, and the soil moisture data reconstruction model is specifically a three-dimensional generative adversarial network model. The three-dimensional generative adversarial network model includes a generator and a discriminator. During the model training process, the loss function corresponding to the generator is specifically configured based on the sum of the absolute value of the error of each pixel point, and the loss function corresponding to the discriminator is specifically configured based on the discriminant loss and the gradient penalty term.

[0008] From the above content, it can be concluded that the present application has the following beneficial effects: Aiming at the goal of reconstructing soil moisture data, this application introduces a generative adversarial network model to create a soil moisture data reconstruction model based on deep learning technology to reconstruct the soil moisture data observed by satellite products, and focuses on the loss function involved in the training process to further enhance its reconstruction performance for soil moisture data, so that the prediction effect of the model can maintain continuity in space and accurately reflect dynamic changes in time, thereby meeting the needs of high-quality soil moisture observation. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1 A flowchart of a configuration method for a satellite soil moisture data reconstruction model of this application; Figure 2 A structural schematic diagram of a configuration device for the satellite soil moisture data reconstruction model of the present application. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0012] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices. The naming or numbering of steps in this application does not mean that the steps in the method flow must be executed in the time / logical sequence indicated by the naming or numbering. The process steps that have been named or numbered can change the execution order according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved.

[0013] The division of modules in this application is a logical division. There may be other division methods when it is implemented in actual applications. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between modules can be electrical or other similar forms, which are not limited in this application. In addition, the modules or submodules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present application.

[0014] First, see Figure 1 , Figure 1 A flow chart of a configuration method of a satellite soil moisture data reconstruction model of the present application is shown. The configuration method of a satellite soil moisture data reconstruction model provided by the present application may specifically include the following steps S101 to S103: Step S101, obtaining first sample satellite soil moisture observation data, wherein the sample satellite soil moisture observation data is soil moisture observation data obtained from satellite observation products; It can be understood that corresponding training samples need to be configured to meet the training requirements of the soil moisture data reconstruction model to meet the model training requirements.

[0015] In this regard, the present application can obtain the initial sample satellite soil moisture observation data. For the convenience of explanation, the present application will record the sample satellite soil moisture observation data obtained here as the first sample satellite soil moisture observation data, and the same applies to the subsequent second sample satellite soil moisture observation data and target satellite soil moisture observation data.

[0016] It should be noted that this application is specifically aimed at satellite soil moisture observation data, which is soil moisture observation data obtained by satellite observation products. Due to the limitations of satellite orbital gaps and payload detection capabilities, it has limitations in time resolution, and thus there is a corresponding need for soil moisture data reconstruction.

[0017] As an example, the sample satellite soil moisture observation data can be obtained from the Soil Moisture Active Passive (SMAP) satellite product. The SMAP satellite is equipped with an L-band radiometer that can cover the entire world at a frequency of three to four days and provide surface soil moisture information in the 0-5 cm soil layer with a spatial resolution of 36 km.

[0018] Of course, it is understandable that, in actual situations, this application does not specifically limit the type of satellite observation product used to obtain the first sample satellite soil moisture observation data, and it can be adjusted according to actual conditions.

[0019] Step S102, configuring corresponding masks and annotations for the first sample satellite soil moisture observation data to obtain second sample satellite soil moisture observation data, wherein the mask is used to shield part of the data content to form the missing content to be reconstructed, and the annotation is soil moisture observation data obtained by ground observation stations; It can be understood that after obtaining the first sample satellite soil moisture observation data, it is still necessary to process the data to form training samples that can be used for specific model training work.

[0020] Specifically, on the one hand, it is necessary to configure a corresponding mask for the first sample satellite soil moisture observation data. Mask is a conventional concept in the field of deep learning technology. In layman's terms, for a piece of data, the mask is used to block / shield part of the data content, so that sample data of different states can be formed. The present application can be used to process the first sample satellite soil moisture observation data into a data part with missing content, and the data part with missing content can be used as the reconstruction target of the subsequent model. For example, 20%-30% of the data content in the original data can be removed and a corresponding mask can be created (that is, the corresponding data content can be deleted behind the mask setting in the specific operation); on the other hand, in the process of model training, it is also necessary to combine the annotation, that is, the real / theoretical satellite soil moisture observation data to guide the model training. It should be noted that the data content processed by the annotation made by the present application here does not need to have a specific correlation with the data content processed by the mask. For the annotation of soil moisture data at different locations, the resolution is easy to understand, which is higher than the resolution of the first sample satellite soil moisture observation data. In this way, the resolution that can be obtained by the reconstruction processing of the model completed by subsequent training can be better than the satellite soil moisture observation data.

[0021] Specifically, the annotation processing done here can be done manually or through automated annotation tools in actual applications. The automated annotation tools need to be pre-configured with the corresponding automated annotation logic. In addition, other existing soil moisture observation data with higher resolution can be directly used. This will be more convenient in terms of operating costs. For example, soil moisture observation data obtained from ground observation stations can be used as the annotations involved here.

[0022] As an example, the annotation here, i.e., the soil moisture observation data obtained by ground observation stations, can specifically be the soil moisture observation data obtained by the International Soil Moisture Network (ISMN) product.

[0023] In this way, after completing the configuration of masks and annotations, the configuration of training samples is completed, and the subsequent specific model training work can be carried out.

[0024] Step S103: training a soil moisture data reconstruction model based on the second sample satellite soil moisture observation data, wherein the soil moisture data reconstruction model is used to reconstruct the satellite soil moisture observation data input into the model, and the soil moisture data reconstruction model is specifically a three-dimensional generative adversarial network model, which includes a generator and a discriminator. During the model training process, the loss function corresponding to the generator is specifically configured based on the sum of the absolute value of the error of each pixel point, and the loss function corresponding to the discriminator is specifically configured based on the discriminant loss and the gradient penalty term.

[0025] It can be understood that the soil moisture data reconstruction model to be constructed in this application is specifically constructed based on a three-dimensional generative adversarial network model (3D Generative Adversarial Networks, 3DGAN). The model itself is an existing model, which includes a generator (Generator, G) with three convolutional layers and a discriminator (Discriminator, D) with three convolutional layers. Its working mode can be briefly understood as: by alternately training the generator and the discriminator, the generator is continuously improved in the adversarial process until the data it generates is difficult to distinguish from the real data.

[0026] On this basis, it can be noted that this application specifically focuses on the loss function involved in the training process of the three-dimensional generative adversarial network model. The training of the deep learning model usually includes the following: A second sample of soil water data is input into the model, so that the model can carry out the corresponding soil moisture data reconstruction processing and realize forward propagation. Then, based on the soil moisture data reconstruction results output by the model, the loss function is calculated in combination with the annotations, and the model parameters are optimized according to the loss function calculation results to realize reverse propagation. In this way, when the model training requirements such as training time, training times or prediction accuracy are met, the model training can be completed and a soil moisture data reconstruction model that can be put into practical use can be obtained.

[0027] It can be concluded that if an adaptive loss function can be configured, it will help achieve more efficient and high-precision model training results.

[0028] In this regard, when creating the soil moisture data reconstruction model that the present application hopes to obtain based on the three-dimensional generative adversarial network model, the present application configures adapted specific loss functions for the generator and discriminator in the model, respectively. In this way, in the soil moisture data reconstruction scenario involved in the present application, the model training can be completed efficiently and with warning. The trained soil moisture data reconstruction model has excellent reconstruction effect for the satellite soil moisture data input into the model, and can reconstruct higher resolution soil moisture data.

[0029] Specifically, on the one hand, for the loss function corresponding to the generator, the present application can be configured based on the sum of the absolute values ​​of the errors of each pixel point, and on the other hand, for the loss function corresponding to the discriminator, it can be configured based on the discriminant loss and the gradient penalty term.

[0030] It can be understood that configuring the loss function based on these two major directions can effectively assist the model in analyzing and predicting soil moisture data. In addition, it also lays a good foundation for the specific loss function quantification formula that can be configured in the subsequent solution content of this application.

[0031] In addition, it should be noted that the reconstruction effect of soil moisture data in this application, based on the higher model prediction performance, has been significantly improved not only in terms of temporal resolution, but also in terms of spatial resolution, thus forming a high-resolution (such as sub-daily / daily scale, that is, 24-hour resolution) global soil moisture data reconstruction effect.

[0032] As an example, in practical applications, the soil moisture data reconstruction model can be specifically configured to carry out daily and global reconstruction processing of the satellite soil moisture observation data input into the model.

[0033] Of course, in terms of time resolution, it can also be twice a day (usually once a day by default) or other fine-grained resolutions.

[0034] After completing the training of the model, it can obviously be put into practical use. Correspondingly, the present application scheme can also involve subsequent model application links.

[0035] Specifically, after step S103 trains the soil moisture data reconstruction model based on the second sample satellite soil moisture observation data, the method of the present application may further include: Obtain target satellite soil moisture observation data; Inputting the target satellite soil moisture observation data into the soil moisture data reconstruction model; Extract the soil moisture data reconstruction results output by the soil moisture data reconstruction model.

[0036] It can be understood that the target satellite soil moisture observation data here can be soil moisture observation data obtained by the same satellite observation product as the first satellite soil moisture observation data, or soil moisture observation data obtained by other satellite observation products.

[0037] In this way, after obtaining the target satellite soil moisture observation data that currently needs to be reconstructed to improve the resolution, the soil moisture data reconstruction model configured in the present application can be used to reconstruct it. After the soil moisture data reconstruction model completes the processing, the corresponding soil moisture data reconstruction result can be output, and the soil moisture data reconstruction result can be extracted.

[0038] At this time, the soil moisture data reconstruction results can be stored locally, stored remotely, forwarded, output as prompts for the completion of the reconstruction process, display content, or perform further analysis and processing. It can be understood that the specific data application content that may be involved later can be adjusted according to the real-time configuration or pre-configured data application strategy / rules, and this application does not make any specific limitations.

[0039] from Figure 1 It can be seen from the illustrated embodiments that, for the purpose of soil moisture data reconstruction, the present application introduces a generative adversarial network model to create a soil moisture data reconstruction model based on deep learning technology to reconstruct the soil moisture data observed by satellite products, and focuses on the loss function involved in the training process to further enhance its reconstruction performance for soil moisture data, so that the prediction effect of the model can maintain continuity in space and accurately reflect dynamic changes in time, thereby meeting the needs of high-quality soil moisture observation.

[0040] Next, we will further explain in detail the loss function specially designed for the three-dimensional generative adversarial network model mentioned above.

[0041] (1) Generator As an exemplary embodiment, the loss function corresponding to the generator can be specifically expressed as follows: , , , in, Indicates the total loss corresponding to the generator, that is, the unified loss value, represents the first balance factor, represents the total loss (value) within the effective area, represents the second balance factor, It represents the sum of the loss (value) of the holes in the data missing area. Mask corresponds to the mask, which is used to distinguish the validity and invalidity of the data. The mask of the valid area is 1, and the mask of the invalid area is 0. represents the predicted value of soil moisture, Indicates the true value of soil moisture.

[0042] It can be understood that in the embodiment here, the present application is a loss function that the generator can specifically involve, and a composite loss function is designed from the perspective of a quantization formula. This specific implementation scheme has better practical significance.

[0043] Furthermore, for the soil moisture data reconstruction model, the processing logic followed by the present application during the training process can be specifically designed to use 9 consecutive days of data as input units and predict soil moisture data with a time resolution of 1 day, i.e. 24 hours.

[0044] Specifically, the present application may use the training samples obtained by processing the soil moisture data of the target date T and the 4 days before and after it (ie, T-4 to T+4) as a training sample to carry out model training.

[0045] Correspondingly, as another exemplary embodiment, the time resolution of the soil moisture data reconstruction model may be specifically 1 day, and step S103 of training the soil moisture data reconstruction model based on the second sample satellite soil moisture observation data may include: Using 9 consecutive days of data as the training unit, the soil moisture data reconstruction model was trained based on the second sample satellite soil moisture observation data.

[0046] It can be understood that the present application believes that training with data of a time span of 9 consecutive days as units has good time / series continuity for the reconstruction of soil moisture data, which can bring about a smooth and accurate reconstruction effect.

[0047] At the same time, this setting also corresponds to the further configuration of the loss function involved in the training of the generator in this application.

[0048] Specifically, as another exemplary embodiment, the loss function corresponding to the generator constrains the smoothing loss of the predicted value, and the loss function corresponding to the generator can also be specifically expressed as follows: , , in, represents the third balance factor, It represents smoothing loss. The corresponding time of the data of 9 consecutive days is recorded as T-4, T-3, T-2, T-1, T, T1, T2, T3, T4, respectively. It represents the predicted soil moisture value from T1 to T4 (that is, the last 4 days of T). It represents the predicted soil moisture value from T-4 to T-1 (that is, the 4 days before T). Other formula characters that have appeared and been explained before will not be repeated here.

[0049] It can be understood that in the embodiment here, the present application continues to introduce smoothing loss on the basis of the previous loss function quantization formula, so that the loss function performance can be further improved in details, which can assist the auxiliary model training to converge faster and promote a more efficient and accurate generator training effect.

[0050] In addition, for the above three balance factors, namely , and , can be fixed to 1. Of course, the specific value can be adjusted according to actual needs in actual situations.

[0051] (2) Discriminator As another exemplary embodiment, the loss function corresponding to the discriminator can be specifically expressed as follows: , , , in, represents the total loss corresponding to the discriminator, Denotes the discriminant loss, penalty denotes the gradient penalty term, and penalty ensures that the gradient of the discriminator remains smooth in the input space and avoids instability during training by forcing the discriminator to satisfy the 1-Lipschitz constraint. represents the sum of the losses of the soil moisture data predicted by the generator after being judged by the discriminator, It represents the sum of the losses of the real soil moisture data after being judged by the discriminator. The discriminant loss is the absolute value of the difference between the predicted soil moisture value and the true value to measure the performance of the discriminator in distinguishing true and false soil moisture data. represents the gradient of the discriminator with respect to the interpolated data ϵ, represents the L2 norm, and N is the number of interpolation samples.

[0052] It can be seen that for the discriminator, similar to the previous generator, this application specifically designs a composite loss function, which consists of two parts: the discriminant loss and the gradient penalty term, providing a stable and effective framework for the training of the discriminator. This specific implementation scheme has better practical significance.

[0053] In this way, based on a series of specific loss functions designed above, specific adversarial training of the 3D generative adversarial network model can be carried out. When the data generated by the generator and the original data visually present a smooth transition effect, it can be considered that the model has reached the convergence standard and can be put into practical use.

[0054] In addition, for the model architecture of the 3D generative adversarial network model itself, this application also provides a specific implementation plan in terms of parameters.

[0055] Specifically, as another exemplary embodiment, in a three-dimensional generative adversarial network model, there may be: Both the generator and the discriminator include three convolutional layers. The three convolutional layers include three-dimensional generative adversarial network units and rectified linear units. The convolution kernel size of the first convolutional layer is 5×5×5, the convolution kernel size of the second convolutional layer is 5×5×5, and the convolution kernel size of the third convolutional layer is 3×3×3.

[0056] Among them, it should be noted that the fact that both the generator and the discriminator include three convolutional layers does not mean that the three convolutional layers included in the generator and the discriminator are the same. The embodiment here only points out that both are composed of three convolutional layers, and each convolutional layer is specifically composed of a three-dimensional generative adversarial network unit and a rectified linear unit (ReLU). The only difference between the two is the size of the convolution kernel.

[0057] In addition, it can be understood that for the three-dimensional generative adversarial network model, the coordination relationship between the generator and the discriminator involved in the working process of the model is not the focus of the present application, or it can follow the existing content, so no further explanation will be given here.

[0058] At the same time, for the above solution content, in order to further understand the excellent results that can be achieved in the reconstruction of satellite soil moisture data, the specific performance obtained in the verification phase of this application solution can also help to explain it more vividly: 1. Able to smoothly transition when filling data, avoiding unnatural boundaries; 2. The number of effective observation days has increased significantly; 3. Globally, the seasonal amplitude of the reconstructed data shows a high degree of consistency with the original data, indicating that it performs well in capturing seasonal changes; 4. Whether in the dry season or the rainy season, the changing trend of soil moisture data can be reproduced, showing a high degree of consistency; 5. Demonstrated strong ability in maintaining spatial and temporal continuity of data; 6. Not only can the spatial and temporal three-dimensional data of soil moisture be used more deeply, but also the complex relationship of nonlinear data can be better handled to obtain better filling effect.

[0059] The above is an introduction to the configuration method of the satellite soil moisture data reconstruction model provided by the present application. In order to facilitate better implementation of the configuration method of the satellite soil moisture data reconstruction model provided by the present application, the present application also provides a configuration device for the soil moisture data reconstruction model from the perspective of functional modules.

[0060] See also Figure 2 , Figure 2 This is a schematic diagram of a configuration device for a satellite soil moisture data reconstruction model of the present application. In the present application, the configuration device 200 for a satellite soil moisture data reconstruction model may specifically include the following structure: An acquisition unit 201 is used to acquire first sample satellite soil moisture observation data, wherein the sample satellite soil moisture observation data is soil moisture observation data obtained from satellite observation products; The configuration unit 202 is used to configure a corresponding mask and annotation for the first sample satellite soil moisture observation data to obtain the second sample satellite soil moisture observation data, wherein the mask is used to shield part of the data content to form the missing content to be reconstructed, and the annotation is the soil moisture observation data obtained by the ground observation station observation; The training unit 203 is used to train a soil moisture data reconstruction model based on the second sample satellite soil moisture observation data, wherein the soil moisture data reconstruction model is used to reconstruct the satellite soil moisture observation data input into the model, and the soil moisture data reconstruction model is specifically a three-dimensional generative adversarial network model. The three-dimensional generative adversarial network model includes a generator and a discriminator. During the model training process, the loss function corresponding to the generator is specifically configured based on the sum of the absolute value of the error of each pixel point, and the loss function corresponding to the discriminator is specifically configured based on the discriminant loss and the gradient penalty term.

[0061] In an exemplary embodiment, the loss function corresponding to the generator is specifically expressed as follows: , , , in, represents the total loss corresponding to the generator, represents the first balance factor, represents the total loss in the effective area, represents the second balance factor, It represents the total loss of holes in the data missing area. Mask corresponds to the mask. The mask of the valid area is 1, and the mask of the invalid area is 0. represents the predicted value of soil moisture, Indicates the true value of soil moisture.

[0062] In another exemplary embodiment, the time resolution of the soil moisture data reconstruction model is 1 day, and the training unit 203 is specifically used for: Using 9 consecutive days of data as the training unit, the soil moisture data reconstruction model was trained based on the second sample satellite soil moisture observation data.

[0063] In another exemplary embodiment, the loss function corresponding to the generator constrains the smoothing loss of the predicted value, and the loss function corresponding to the generator is specifically expressed as follows: , , in, represents the third balance factor, It represents smoothing loss. The corresponding time of the data of 9 consecutive days is recorded as T-4, T-3, T-2, T-1, T, T1, T2, T3, T4, respectively. represents the predicted soil moisture value from T1 to T4, Indicates the predicted soil moisture value from T-4 to T-1.

[0064] In yet another exemplary embodiment, , and Both are 1.

[0065] In another exemplary embodiment, the loss function corresponding to the discriminator is specifically expressed as follows: , , , in, represents the total loss corresponding to the discriminator, represents the discriminant loss, penalty represents the gradient penalty term, represents the sum of the losses of the soil moisture data predicted by the generator after being judged by the discriminator, It represents the total loss of the real soil moisture data after being judged by the discriminator. represents the gradient of the discriminator with respect to the interpolated data ϵ, represents the L2 norm, and N is the number of interpolation samples.

[0066] In another exemplary embodiment, in a three-dimensional generative adversarial network model, there are: Both the generator and the discriminator include three convolutional layers. The three convolutional layers include three-dimensional generative adversarial network units and rectified linear units. The convolution kernel size of the first convolutional layer is 5×5×5, the convolution kernel size of the second convolutional layer is 5×5×5, and the convolution kernel size of the third convolutional layer is 3×3×3.

[0067] In yet another exemplary embodiment, the soil moisture data reconstruction model is specifically used to carry out daily and global reconstruction processing on the satellite soil moisture observation data input into the model.

[0068] In yet another exemplary embodiment, the apparatus further includes an application unit 204, configured to: Obtain target satellite soil moisture observation data; Inputting the target satellite soil moisture observation data into the soil moisture data reconstruction model; Extract the soil moisture data reconstruction results output by the soil moisture data reconstruction model.

[0069] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the configuration device of the satellite soil moisture data reconstruction model and the specific working process of its corresponding units described above can refer to the following. Figure 1 The description of the configuration method of the satellite soil moisture data reconstruction model in the corresponding embodiment will not be repeated here.

[0070] The configuration method and device of the satellite soil moisture data reconstruction model provided by the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the core idea of ​​the present application; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A configuration method for a satellite soil moisture data reconstruction model, characterized in that: The method comprises: Acquire first sample satellite soil moisture observation data, wherein the sample satellite soil moisture observation data is soil moisture observation data obtained from a satellite observation product; Configuring corresponding masks and annotations for the first sample satellite soil moisture observation data to obtain second sample satellite soil moisture observation data, wherein the mask is used to shield part of the data content to form the missing content to be reconstructed, and the annotation is the soil moisture observation data obtained by observing the ground observation station; Based on the second sample satellite soil moisture observation data, a soil moisture data reconstruction model is trained, wherein the soil moisture data reconstruction model is used to reconstruct the satellite soil moisture observation data input into the model, and the soil moisture data reconstruction model is specifically a three-dimensional generative adversarial network model, and the three-dimensional generative adversarial network model includes a generator and a discriminator. During the model training process, the loss function corresponding to the generator is specifically configured based on the sum of the absolute value of the error of each pixel point, and the loss function corresponding to the discriminator is specifically configured based on the discriminant loss and the gradient penalty term.

2. The method according to claim 1, characterized in that The loss function corresponding to the generator is specifically expressed as follows: ， ， ， in, represents the total loss corresponding to the generator, represents the first balance factor, represents the total loss in the effective area, represents the second balance factor, represents the total loss of holes in the data missing area, mask corresponds to the mask, mask=1 in the valid area, and mask=0 in the invalid area. represents the predicted value of soil moisture, Indicates the true value of soil moisture.

3. The method according to claim 1 or 2, characterized in that: The time resolution of the soil moisture data reconstruction model is 1 day. The training of the soil moisture data reconstruction model based on the second sample satellite soil moisture observation data includes: The soil moisture data reconstruction model is trained based on the second sample satellite soil moisture observation data, using data from nine consecutive days as a training unit.

4. The method according to claim 3, characterized in that The loss function corresponding to the generator constrains the smoothing loss of the predicted value. The loss function corresponding to the generator is specifically expressed as follows: ， ， in, represents the third balance factor, represents smoothing loss, and the corresponding time of the data of the consecutive 9 days is recorded as T-4, T-3, T-2, T-1, T, T1, T2, T3, T4, respectively. represents the predicted soil moisture value from T1 to T4, Indicates the predicted soil moisture value from T-4 to T-1.

5. The method according to claim 4, characterized in that , and Both are 1.

6. The method according to claim 4, characterized in that The loss function corresponding to the discriminator is specifically expressed as follows: ， ， ， in, represents the total loss corresponding to the discriminator, represents the discriminant loss, penalty represents the gradient penalty term, represents the sum of the losses of the soil moisture data predicted by the generator after being judged by the discriminator, represents the total loss of the real soil moisture data after being judged by the discriminator, represents the gradient of the discriminator with respect to the interpolated data ϵ, represents the L2 norm, and N is the number of interpolation samples.

7. The method according to claim 1, characterized in that In the three-dimensional generative adversarial network model, there are: Both the generator and the discriminator include three convolutional layers, and the three convolutional layers include three-dimensional generative adversarial network units and rectified linear units. The convolution kernel size of the first convolutional layer is 5×5×5, the convolution kernel size of the second convolutional layer is 5×5×5, and the convolution kernel size of the third convolutional layer is 3×3×3.

8. The method according to claim 1, characterized in that The soil moisture data reconstruction model is specifically used to carry out daily and global reconstruction processing on the satellite soil moisture observation data input into the model.

9. The method according to claim 1, characterized in that: After training the soil moisture data reconstruction model based on the second sample satellite soil moisture observation data, the method further includes: Obtain target satellite soil moisture observation data; Inputting the target satellite soil moisture observation data into the soil moisture data reconstruction model; The soil moisture data reconstruction result output by the soil moisture data reconstruction model is extracted.

10. A configuration device for satellite soil moisture data reconstruction model, characterized in that: The device comprises: An acquisition unit, configured to acquire first sample satellite soil moisture observation data, wherein the sample satellite soil moisture observation data is soil moisture observation data obtained from a satellite observation product; a configuration unit, configured to configure a corresponding mask and annotation for the first sample satellite soil moisture observation data to obtain second sample satellite soil moisture observation data, wherein the mask is used to shield part of the data content to form the missing content to be reconstructed, and the annotation is the soil moisture observation data obtained by observing the ground observation station; A training unit is used to train a soil moisture data reconstruction model based on the second sample satellite soil moisture observation data, wherein the soil moisture data reconstruction model is used to reconstruct the satellite soil moisture observation data input into the model, and the soil moisture data reconstruction model is specifically a three-dimensional generative adversarial network model, and the three-dimensional generative adversarial network model includes a generator and a discriminator. During the model training process, the loss function corresponding to the generator is specifically configured based on the sum of the absolute value of the error of each pixel point, and the loss function corresponding to the discriminator is specifically configured based on the discriminant loss and the gradient penalty term.

Images