A method for reconstructing vegetation optical thickness data based on partial convolutional neural network

Abstract

The present application provides a method for reconstructing vegetation optical thickness data based on a partial convolutional neural network, which is used for the first time to introduce a partial convolutional neural network into the data missing filling work of vegetation optical thickness, and to realize spatiotemporal reconstruction of the initial vegetation optical thickness data through the partial convolutional neural network, so as to obtain target vegetation optical thickness data with significantly improved spatiotemporal resolution, which can capture the fine changes in vegetation water content, help to have a deeper understanding of the physiological mechanism of vegetation response to drought, and provide strong data support for the application of vegetation optical thickness data in fine-scale vegetation monitoring research.

Description

Technical Field

[0001] The present application relates to the field of vegetation optical thickness data analysis and processing, and specifically to a vegetation optical thickness data reconstruction method based on a partial convolutional neural network. Background Art

[0002] Vegetation is a key factor in the balance of energy, water and carbon on the land surface, and plays an important role in regulating climate change and the water-carbon energy cycle. As one of the important parameters for assessing the growth status and health of vegetation, vegetation optical depth (VOD) has been widely used in the past few decades. Vegetation optical depth characterizes the attenuation degree of microwave radiation from the earth's surface by the vertical structure of the entire vegetation. This parameter is closely related to factors such as vegetation density, vegetation water content (VWC) and above ground biomass (AGB).

[0003] Since the late 1970s, satellite-borne microwave radiometers have been providing measurement data of Earth surface radiation, which can be used to invert vegetation optical depth. Compared with traditional optical remote sensing technology, microwave remote sensing is less affected by factors such as atmosphere, clouds and sunlight, and has the advantages of large-scale and long-term observation. Therefore, it has been increasingly widely used in the field of large-scale vegetation dynamic monitoring. At present, vegetation optical depth has been used in vegetation growth, carbon balance, phenological monitoring, biomass assessment, vegetation recovery after fire, and vegetation response to drought. In summary, by analyzing the changes in VOD, the dynamic changes of vegetation can be effectively monitored, providing important information for ecological research and environmental management.

[0004] The inventors of the present application have discovered that the acquisition of vegetation optical thickness data mainly relies on a multi-parameter collaborative inversion algorithm. However, due to the limitations of satellite orbit coverage and inversion algorithms in actual situations, the existing daily-scale vegetation optical thickness products cannot achieve global seamless temporal and spatial coverage and have an obvious missing rate problem (for example, the missing rate can reach 40%-50%), which limits the application of vegetation optical thickness in fine-scale vegetation monitoring research. Summary of the invention

[0005] The present application provides a method for reconstructing vegetation optical thickness data based on a partial convolutional neural network, which is used for the first time to introduce a partial convolutional neural network into the data missing filling work of vegetation optical thickness, and to realize spatiotemporal reconstruction of the initial vegetation optical thickness data through the partial convolutional neural network, so as to obtain target vegetation optical thickness data with significantly improved spatiotemporal resolution, which can capture the fine changes in vegetation water content, help to have a deeper understanding of the physiological mechanism of vegetation response to drought, and provide strong data support for the application of vegetation optical thickness data in fine-scale vegetation monitoring research.

[0006] In a first aspect, the present application provides a method for reconstructing vegetation optical thickness data based on a partial convolutional neural network, the method comprising:

[0007] Acquire initial vegetation optical thickness data to be reconstructed, wherein the initial vegetation optical thickness data is a data product obtained by inverting vegetation optical thickness based on observation data of a satellite-borne microwave radiometer;

[0008] Inputting the initial vegetation optical thickness data into a preset vegetation optical thickness data reconstruction model, so that the vegetation optical thickness data is subjected to corresponding vegetation optical thickness data reconstruction processing based on the initial vegetation optical thickness data, wherein the vegetation optical thickness data reconstruction model is obtained by training the initial model in advance with labeled sample vegetation optical thickness data, and the vegetation optical thickness data reconstruction model is used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input into the model, so as to improve the spatiotemporal resolution of the vegetation optical thickness data input into the model, and the vegetation optical thickness data reconstruction model specifically adopts a model architecture of a partial convolutional neural network;

[0009] The vegetation optical thickness data reconstruction result corresponding to the initial vegetation optical thickness data output by the vegetation optical thickness data reconstruction model is extracted and output as the target vegetation optical thickness data.

[0010] In a second aspect, the present application provides a vegetation optical thickness data reconstruction device based on a partial convolutional neural network, the device comprising:

[0011] An acquisition unit is used to acquire initial vegetation optical thickness data to be reconstructed, wherein the initial vegetation optical thickness data is a data product obtained by inverting the vegetation optical thickness based on observation data of a satellite-borne microwave radiometer;

[0012] A reconstruction unit, used for inputting the initial vegetation optical thickness data into a preset vegetation optical thickness data reconstruction model, so that the vegetation optical thickness data is subjected to corresponding vegetation optical thickness data reconstruction processing based on the initial vegetation optical thickness data, wherein the vegetation optical thickness data reconstruction model is obtained by training the initial model in advance with labeled sample vegetation optical thickness data, and the vegetation optical thickness data reconstruction model is used for conducting corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input by the model, so as to improve the spatiotemporal resolution of the vegetation optical thickness data input by the model, and the vegetation optical thickness data reconstruction model specifically adopts a model architecture of a partial convolutional neural network;

[0013] The extraction unit is used to extract the vegetation optical thickness data reconstruction result corresponding to the initial vegetation optical thickness data output by the vegetation optical thickness data reconstruction model, and output it as the target vegetation optical thickness data.

[0014] From the above content, it can be concluded that the present application has the following beneficial effects:

[0015] Aiming at the goal of filling missing data of vegetation optical thickness, this application introduces partial convolutional neural networks into the work of filling missing data of vegetation optical thickness for the first time. The partial convolutional neural networks are used to realize spatiotemporal reconstruction of the initial vegetation optical thickness data to obtain target vegetation optical thickness data with significantly improved spatiotemporal resolution. This can capture the subtle changes in vegetation water content, help to have a deeper understanding of the physiological mechanism of vegetation response to drought, and provide strong data support for the application of vegetation optical thickness data in fine-scale vegetation monitoring research. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] 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.

[0017] Figure 1 A schematic diagram of a process of a vegetation optical thickness data reconstruction method based on a partial convolutional neural network in this application;

[0018] Figure 2 This is a structural schematic diagram of a vegetation optical thickness data reconstruction device based on a partial convolutional neural network in the present application. DETAILED DESCRIPTION

[0019] 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.

[0020] 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.

[0021] 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.

[0022] See also Figure 1 , Figure 1 A schematic flow chart of a method for reconstructing vegetation optical thickness data based on a partial convolutional neural network of the present application is shown. The method for reconstructing vegetation optical thickness data based on a partial convolutional neural network provided by the present application may specifically include the following steps S101 to S103:

[0023] Step S101, obtaining initial vegetation optical thickness data to be reconstructed, wherein the initial vegetation optical thickness data is a data product obtained by inverting vegetation optical thickness based on observation data of a satellite-borne microwave radiometer;

[0024] Generally speaking, the initial vegetation optical thickness data obtained here is the existing or ready-made vegetation optical thickness data. In order to distinguish it from the vegetation optical thickness data involved in the subsequent process, the vegetation optical thickness data obtained here is recorded as the initial vegetation optical thickness data.

[0025] As for the initial vegetation optical thickness data, it can be understood that it itself is a data product of the existing technology, which is obtained by inverting the vegetation optical thickness based on the observation data of the satellite-borne microwave radiometer. Therefore, due to the limitations of satellite orbit coverage and inversion algorithms in actual situations, there is a problem of obvious missing rate (for example, the missing degree can reach 40%-50%), and there are obvious limitations in both time resolution and spatial resolution. Therefore, it is difficult to provide stable and highly accurate data support for the application of vegetation optical thickness data in fine-scale vegetation monitoring research.

[0026] As an example, the initial vegetation optical thickness data can be specifically the SMAP MCCA VOD product. The SMAP satellite mainly provides soil moisture data and surface freeze-thaw status data. It can observe the world twice a day, including the ascending orbit (6:00 pm local time) and the descending orbit (6:00 am local time). However, due to the failure of the active sensor, only the passive sensor continues to work. The L-band passive microwave radiometer carried by the SMAP satellite can obtain a data set with a spatial resolution of 36km and a global cycle coverage of 2-3 days.

[0027] Taking the 36km SMAP daily VOD product from 2015 to 2022 obtained by inversion using the Multi-Channel Collaborative Inversion Algorithm (MCCA) as an example, this product is projected in the EASE-GRID 2.0 (Equal Area Scalable Ground Grid 2.0) coordinate system. This product includes horizontal and vertical polarization modes. It is the world's first L-band VOD product with polarization difference, but the spatial coverage of the inverted product is only 50%-60%. These incomplete VOD products are not suitable for research in fields such as vegetation dynamic monitoring.

[0028] In addition, in actual situations, the acquisition and processing of the initial vegetation optical thickness data to be reconstructed is usually initiated along with the vegetation optical thickness data reconstruction task. The vegetation optical thickness data reconstruction task directly or indirectly specifies the current initial vegetation optical thickness data as the reconstruction object. It can be initiated manually, by receiving data sent from other devices, or by autonomously initiating according to a preset autonomous task initiation strategy. All of these are possible to meet the flexible and changeable application needs in actual situations.

[0029] Step S102, inputting the initial vegetation optical thickness data into a preset vegetation optical thickness data reconstruction model, so that the vegetation optical thickness data is subjected to corresponding vegetation optical thickness data reconstruction processing based on the initial vegetation optical thickness data, wherein the vegetation optical thickness data reconstruction model is obtained by training the initial model in advance with labeled sample vegetation optical thickness data, and the vegetation optical thickness data reconstruction model is used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input by the model, so as to improve the spatiotemporal resolution of the vegetation optical thickness data input by the model, and the vegetation optical thickness data reconstruction model specifically adopts a model architecture of a partial convolutional neural network;

[0030] It is easy to see that the present application specifically uses a partial convolutional neural network to create a vegetation optical thickness data reconstruction model for reconstructing vegetation optical thickness data, namely, Partial Convolutional Neural Network, PCNN.

[0031] Specifically, compared with traditional convolutional networks, some convolutional neural networks can selectively update the pixel values ​​of the required vegetation optical thickness data when processing the input initial vegetation optical thickness data. The operation formula of ordinary convolution is expressed as follows:

[0032] ,

[0033] in, represents the output pixel value of the convolution result, represents the weight of the convolution kernel, represents the pixel value of the input image, Represents the bias parameter.

[0034] In the process of filling the vegetation optical thickness data, the ordinary convolution method cannot identify whether the information is valid and cannot eliminate the impact of invalid information.

[0035] Some convolutional neural networks can solve this problem by introducing a mask. The mask is a matrix of the same size as the input image and can be updated adaptively. The corresponding operation formula can be shown as follows:

[0036] ,

[0037] in, Represents the pixel value of the partial convolution result output, represents the weight of the convolution kernel, represents the pixel value of the input image, represents the bias parameter, represents the input image mask, ⊙ represents and Multiply pixel by pixel, represents the scaling factor, A matrix representing a region whose size is the same as the convolution kernel but all values ​​are 1. Represents the size of the mask area with all values ​​1, 1 represents a valid pixel, and 0 represents an invalid pixel. In the partial convolution operation, the pixel value of the output image is updated only when the pixel at the corresponding position in the mask is 1. By adding a mask, other interference information can be excluded. After each convolution layer is completed, the mask can calculate the output value based on at least one non-zero input value, then the new mask value is 1. The formula for updating the mask is as follows:

[0038] .

[0039] It can be seen that some convolutional neural networks can be trained on masks to reconstruct optical thickness data well without the need for other auxiliary data, thereby significantly improving the spatiotemporal resolution of the original vegetation optical thickness data.

[0040] In this regard, the present application can perform corresponding vegetation optical thickness data reconstruction processing on the current initial vegetation optical thickness data based on a pre-trained vegetation optical thickness data reconstruction model to obtain new vegetation optical thickness data after the initial vegetation optical thickness data is reconstructed.

[0041] Step S103: extracting the vegetation optical thickness data reconstruction result corresponding to the initial vegetation optical thickness data output by the vegetation optical thickness data reconstruction model, and outputting it as the target vegetation optical thickness data.

[0042] It can be understood that after the vegetation optical thickness data reconstruction model completes the reconstruction processing of the initial vegetation optical thickness data, the vegetation optical thickness data reconstruction result can be obtained. The vegetation optical thickness data reconstruction result is obviously also a kind of vegetation optical thickness data, so it can be referred to as the target vegetation optical thickness data.

[0043] At this point, after obtaining the target vegetation optical thickness data, the next step of data application processing can be involved. Specifically, storage, forwarding, content display, output of completed reconstruction prompts, or related vegetation monitoring research based on vegetation optical thickness data can be carried out.

[0044] It is easy to understand that how to use the reconstructed target vegetation optical thickness data with significantly improved usage resolution for data application processing can be processed according to the pre-configured data application strategy and the current real-time configured data application strategy. This application does not make specific limitations on this link.

[0045] from Figure 1 It can be seen from the illustrated embodiments that, with regard to the goal of filling in missing data on vegetation optical thickness, the present application introduces a partial convolutional neural network for the first time into the work of filling in missing data on vegetation optical thickness, and uses a partial convolutional neural network to achieve spatiotemporal reconstruction of the initial vegetation optical thickness data to obtain target vegetation optical thickness data with significantly improved spatiotemporal resolution. This can capture subtle changes in vegetation water content, help to more deeply understand the physiological mechanism of vegetation response to drought, and provide strong data support for the application of vegetation optical thickness data in fine-scale vegetation monitoring research.

[0046] Next, we will continue to Figure 1 Each step of the illustrated embodiment and possible embodiments thereof in practical application are explained in detail.

[0047] As an exemplary embodiment, corresponding to the application of the vegetation optical thickness data reconstruction model of the present application, the present application may also involve the training link of the model. In this regard, the vegetation optical thickness data reconstruction method based on a partial convolutional neural network of the present application may also include:

[0048] Configure the labeled sample vegetation optical thickness data;

[0049] Based on the sample vegetation optical thickness data, an initial model is trained, and the trained model is used as a vegetation optical thickness data reconstruction model.

[0050] It is understandable that the sample vegetation optical thickness data is usually ready-made data. For example, the same data product as the initial optical thickness data to improve the spatiotemporal resolution after the model is put into practical application can be used. Of course, other specific data products can also be used.

[0051] The labeling of sample vegetation optical thickness data can be done manually or using corresponding automatic labeling tools. The purpose of labeling is to indicate the true prediction value, so as to guide the recognition / training of the model in combination with the loss function. The form of the labeling result corresponds to the application of some convolutional neural networks, which can be implemented using masks.

[0052] After configuring the labeled sample vegetation optical thickness data, it is obvious that specific model training processing can be carried out, in which any type of model training architecture can be used and adjusted according to actual needs.

[0053] The model training process usually includes the following processing:

[0054] In each round of model training, a sample is input into the model, so that the model can carry out the corresponding vegetation optical thickness data reconstruction processing to realize forward propagation. Then, the loss function is calculated based on the vegetation optical thickness data reconstruction results output by the model, and the model parameters are optimized according to the loss function calculation results to realize reverse propagation. In this way, when the preset model training requirements such as training time, training times or prediction accuracy are met, the model training can be completed. At this time, the model is a vegetation optical thickness data reconstruction model that can be put into practical use.

[0055] In addition, the present application has further optimized settings for the vegetation optical thickness data reconstruction model, rather than being limited to introducing part of the convolutional neural network into the configuration work of the vegetation optical thickness data reconstruction model.

[0056] Specifically, the previous step S103 trains an initial model based on the sample vegetation optical thickness data, and uses the trained model as a vegetation optical thickness data reconstruction model, which may include:

[0057] Based on the sample vegetation optical thickness data, the initial model was trained with 9 consecutive days of data as units, and the trained model was used as the vegetation optical thickness data reconstruction model.

[0058] It can be seen from here that the vegetation optical thickness data reconstruction model configured in the present application, on the one hand, carries out specific reconstruction work in units of days (days), so it has a daily / 24-hour time resolution, and on the other hand, during the training process, it is trained in units of data with a time span of 9 consecutive days. In this way, under actual circumstances, the reconstruction work of the vegetation optical thickness data has good time / time series continuity, which can bring a smooth and precise reconstruction effect.

[0059] As an example, before importing the data into the model, 40%-50% of the vegetation optical thickness data needs to be eliminated to generate the corresponding mask, and then the vegetation optical thickness data of the target date T and the 4 days before and after (T-4, T+4) and the corresponding 9-day mask are simultaneously imported into the model as the input data for training.

[0060] In addition, the present application provides a loss function used in the model training process as an exemplary embodiment. The loss function used in the model training process may be:

[0061] ,

[0062] ,

[0063] ,

[0064] ,

[0065] in, is the uniform loss value (i.e., overall loss), is the first loss balance factor (i.e. the first weight), is the sum of the loss values ​​of the remaining effective areas except the invalid areas of holes in the effective area, is the second loss balance factor (i.e., the second weight), The sum of the loss values ​​of the holes and invalid areas created in the valid area, is the third loss balance factor (i.e., the third weight), To smooth the loss,

[0066] The value in the valid area is 1. The value in the invalid area is 0. is the predicted value of vegetation optical thickness, is the true value of vegetation optical thickness,

[0067] To enter the forecast data for the next four days after the target date T, Enter the forecast data for four days before the target date T.

[0068] For the above loss function combination settings, it can be understood that while a set of specific implementation solutions are given for the corresponding part of the convolutional neural network characteristics, a more suitable smoothing loss configuration solution is given based on the aforementioned 9-day training sample unit, so that the model can be trained more efficiently and accurately, and the trained model can have better reconstruction / prediction effects.

[0069] In addition, the present application also provides a more specific implementation solution for the model architecture of the vegetation optical thickness data reconstruction model / initial model based on the working logic of some convolutional neural networks.

[0070] For the convenience of explanation, the specific working logic of the model can be combined for explanation. Specifically, as an exemplary embodiment, the vegetation optical thickness data reconstruction model includes the following processing contents during the working process:

[0071] Through 3 layers of 3D partial convolutional neural network units and rectified linear units (i.e. ReLU activation function), 3 layers of convolution are performed.

[0072] Update the mask after each layer of convolution and import the updated mask into the next convolution operation.

[0073] In each layer of convolution, the dimension of the convolution kernel is defined as 3×3×3. If the transition between the filled result and the original image is smooth, the convergence result is considered qualified, and the processed image is finally output.

[0074] Among them, it can be understood that because each pixel output by the 3D partial convolutional neural network model when reconstructing vegetation optical thickness data only depends on the pixels at the corresponding position in the input vegetation optical thickness data, this operation can maintain spatial consistency and make the reconstructed vegetation optical thickness data more accurate.

[0075] At the same time, the spatiotemporal resolution that can be achieved by the vegetation optical thickness data reconstruction model of the present application in its reconstruction work can be specifically a global daily scale resolution, that is, on the one hand, it is a daily, 24-hour temporal resolution, and on the other hand, it is a global spatial resolution, thus having a powerful vegetation optical thickness data reconstruction effect.

[0076] In addition, in order to further improve the reconstruction effect, the present application can continue to specially configure new influencing factors to participate in the reconstruction of the vegetation optical thickness data input into the model.

[0077] Correspondingly, as an exemplary embodiment, the vegetation optical thickness data reconstruction model can be specifically used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input into the model and the surface cover type corresponding to the specific area.

[0078] Similarly, when conducting model training, it is also necessary to label the surface cover type corresponding to the area involved in the sample vegetation optical thickness data phase.

[0079] It can be understood that the present application believes that in terms of details, different surface cover types will have slight differences in the reconstruction of the model. Therefore, by indicating the surface cover types corresponding to different areas, the model can be guided to make more delicate vegetation optical thickness data reconstruction processing in terms of details, which will help avoid subtle disturbances and achieve more accurate reconstruction effects.

[0080] As for the surface cover type corresponding to the specific areas involved, it can be manually configured or indicated by some existing data products.

[0081] In this regard, as an exemplary embodiment, the land cover type corresponding to the specific area involved in the present application can be determined by a global land cover type information product obtained by observing the global land cover situation using satellite remote sensing technology.

[0082] As an example, the global land cover type information product can adopt the MODIS IGBP product, which is based on MODIS sensor data and aims to provide global land cover type information. The product is updated annually and provides global surface cover data with a spatial resolution of 500 meters. The surface is classified into 17 main types, including forests, grasslands, farmlands, urban areas, etc., which can bring more convenient processing effects.

[0083] Of course, in practical applications, it can be understood that the number of surface cover types to be configured and how to divide different areas can be adjusted according to actual needs.

[0084] As an example, the surface cover types involved in the present application may specifically include bare land type, ice and snow type, mixed land type of crops and natural vegetation, urban type, crop type, permanent wetland type, grassland type, savanna type, tree-filled grassland type, open shrub type, closed shrub type, mixed forest type, deciduous broad-leaved forest type, deciduous coniferous forest type, evergreen broad-leaved forest type, evergreen coniferous forest type and water body type.

[0085] In addition, in addition to the land cover types corresponding to specific areas, this application can continue to introduce new influencing factors.

[0086] Correspondingly, as an exemplary embodiment, the vegetation optical thickness data reconstruction model can be specifically used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input by the model and the surface cover type corresponding to the specific area, as well as the corresponding regional factors and environmental factors.

[0087] It can be understood that the present application believes that in some detailed cases, different regions and different environments may also have subtle effects on the reconstruction of vegetation optical thickness data. Therefore, it can provide certain references for fine-tuning to obtain a clearer reconstruction effect.

[0088] Regarding the above solution content, it can be understood that in the verification phase, it can be found that the vegetation optical thickness data reconstruction model configured in the present application is effective in filling the vegetation optical thickness for about 50% missing data. At the same time, it can be observed that not only the boundary between the missing area and the original data shows excellent spatial continuity, but the texture of the reconstruction result inside the missing area is also highly consistent with the original data. The filled data also shows a high consistency with the original data in seasonality and annual periodicity. The relevant evaluation indicators are significantly better than existing reconstruction schemes such as the DCT-PLS filling method, and the filling / reconstruction effect is best during the combined period of ascending and descending orbits. This fully proves the application value of the solution of the present application. By reconstructing vegetation optical thickness data with high temporal and spatial resolution, it is possible to capture subtle changes in vegetation water content, which helps to more deeply understand the physiological mechanism of vegetation response to drought, and provides strong data support for the application of vegetation optical thickness data in fine-scale vegetation monitoring research.

[0089] The above is an introduction to the vegetation optical thickness data reconstruction method based on a partial convolutional neural network provided in the present application. In order to facilitate better implementation of the vegetation optical thickness data reconstruction method based on a partial convolutional neural network provided in the present application, the present application also provides a vegetation optical thickness data reconstruction device based on a partial convolutional neural network from the perspective of functional modules.

[0090] See also Figure 2 , Figure 2 This is a structural schematic diagram of a vegetation optical thickness data reconstruction device based on a partial convolutional neural network in the present application. In the present application, the vegetation optical thickness data reconstruction device 200 based on a partial convolutional neural network may specifically include the following structure:

[0091] An acquisition unit 201 is used to acquire initial vegetation optical thickness data to be reconstructed, wherein the initial vegetation optical thickness data is a data product obtained by inverting vegetation optical thickness based on observation data of a satellite-borne microwave radiometer;

[0092] The reconstruction unit 202 is used to input the initial vegetation optical thickness data into a preset vegetation optical thickness data reconstruction model, so that the vegetation optical thickness data is reconstructed based on the initial vegetation optical thickness data, wherein the vegetation optical thickness data reconstruction model is obtained by training the initial model with labeled sample vegetation optical thickness data in advance, and the vegetation optical thickness data reconstruction model is used to carry out corresponding vegetation optical thickness data reconstruction based on the vegetation optical thickness data input by the model, so as to improve the spatiotemporal resolution of the vegetation optical thickness data input by the model, and the vegetation optical thickness data reconstruction model specifically adopts a model architecture of a partial convolutional neural network;

[0093] The extraction unit 203 is used to extract the vegetation optical thickness data reconstruction result corresponding to the initial vegetation optical thickness data output by the vegetation optical thickness data reconstruction model, and output it as the target vegetation optical thickness data.

[0094] As an exemplary embodiment, the apparatus further includes a training unit 204, configured to:

[0095] Configure the labeled sample vegetation optical thickness data;

[0096] Based on the sample vegetation optical thickness data, an initial model is trained, and the trained model is used as a vegetation optical thickness data reconstruction model.

[0097] As another exemplary embodiment, the training unit 204 is specifically configured to:

[0098] Based on the sample vegetation optical thickness data, the initial model was trained with 9 consecutive days of data as units, and the trained model was used as the vegetation optical thickness data reconstruction model.

[0099] As another exemplary embodiment, the loss function used in the model training process is specifically:

[0100] ,

[0101] ,

[0102] ,

[0103] ,

[0104] in, To unify the loss value, is the first loss balance factor, is the sum of the loss values ​​of the remaining effective areas except the invalid areas of holes in the effective area, is the second loss balance factor, The sum of the loss values ​​of the holes and invalid areas created in the valid area, is the third loss balance factor, To smooth the loss,

[0105] The value in the valid area is 1. The value in the invalid area is 0. is the predicted value of vegetation optical thickness, is the true value of vegetation optical thickness,

[0106] To enter the forecast data for the next four days after the target date T, Enter the forecast data for four days before the target date T.

[0107] As another exemplary embodiment, the vegetation optical thickness data reconstruction model includes the following processing contents during its working process:

[0108] Through 3 layers of 3D partial convolutional neural network units and rectified linear units, 3 layers of convolution are performed.

[0109] Update the mask after each layer of convolution and import the updated mask into the next convolution operation.

[0110] In each layer of convolution, the dimension of the convolution kernel is defined as 3×3×3. If the transition between the filled result and the original image is smooth, the convergence result is considered qualified, and the processed image is finally output.

[0111] As another exemplary embodiment, the temporal and spatial resolution is specifically a global daily scale resolution.

[0112] As another exemplary embodiment, the vegetation optical thickness data reconstruction model is specifically used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input into the model and the surface cover type corresponding to the specific area.

[0113] As another exemplary embodiment, the land cover type corresponding to a specific area is determined by using a global land cover type information product obtained by observing the global land cover situation using satellite remote sensing technology.

[0114] As another exemplary embodiment, the vegetation optical thickness data reconstruction model is specifically used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input into the model and the surface cover type corresponding to the specific area, as well as the corresponding regional factors and environmental factors.

[0115] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the soil moisture spatiotemporal prediction device for geospatial sensor network and its corresponding units described above can refer to Figure 1 The description of the spatiotemporal prediction method of soil moisture for a geospatial sensor network in the corresponding embodiment will not be repeated here.

[0116] The above is a detailed introduction to the vegetation optical thickness data reconstruction method and device based on a partial convolutional neural network provided by the present application. 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 method of the present application and its core idea; 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 method for reconstructing vegetation optical thickness data based on partial convolutional neural network, characterized in that: The method comprises: Acquire initial vegetation optical thickness data to be reconstructed, wherein the initial vegetation optical thickness data is a data product obtained by inverting vegetation optical thickness based on observation data of a satellite-borne microwave radiometer; Inputting the initial vegetation optical thickness data into a preset vegetation optical thickness data reconstruction model, so that the vegetation optical thickness data is subjected to corresponding vegetation optical thickness data reconstruction processing based on the initial vegetation optical thickness data, wherein the vegetation optical thickness data reconstruction model is obtained by training an initial model in advance with labeled sample vegetation optical thickness data, and the vegetation optical thickness data reconstruction model is used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input by the model, so as to improve the spatiotemporal resolution of the vegetation optical thickness data input by the model, and the vegetation optical thickness data reconstruction model specifically adopts a model architecture of a partial convolutional neural network; The vegetation optical thickness data reconstruction result corresponding to the initial vegetation optical thickness data output by the vegetation optical thickness data reconstruction model is extracted and output as the target vegetation optical thickness data.

2. The method according to claim 1, characterized in that: The method further comprises: Configure the labeled optical thickness data of the sample vegetation; Based on the sample vegetation optical thickness data, the initial model is trained, and the trained model is used as the vegetation optical thickness data reconstruction model.

3. The method according to claim 2, characterized in that The step of training the initial model based on the sample vegetation optical thickness data and using the trained model as the vegetation optical thickness data reconstruction model comprises: On the basis of the sample vegetation optical thickness data, the initial model is trained with 9 consecutive days of data as a unit, and the trained model is used as the vegetation optical thickness data reconstruction model.

4. The method according to claim 3, characterized in that The loss function used in the model training process is as follows: ， ， ， ， in, To unify the loss value, is the first loss balance factor, is the sum of the loss values ​​of the remaining effective areas except the invalid areas of holes in the effective area, is the second loss balance factor, The sum of the loss values ​​of the holes and invalid areas created in the valid area, is the third loss balance factor, To smooth the loss, The value in the valid region is 1, The value in the invalid area is 0. is the predicted value of vegetation optical thickness, is the true value of vegetation optical thickness, To enter the forecast data for the next four days after the target date T, The forecast data for the four days before the input target date T is obtained.

5. The method according to claim 1, characterized in that The vegetation optical thickness data reconstruction model includes the following processing contents during the working process: Through 3 layers of 3D partial convolutional neural network units and rectified linear units, 3 layers of convolution are performed. Update the mask after each layer of convolution and import the updated mask into the next convolution operation. In each layer of convolution, the dimension of the convolution kernel is defined as 3×3×3. If the transition between the filled result and the original image is smooth, the convergence result is considered qualified, and the processed image is finally output.

6. The method according to claim 1, characterized in that The temporal and spatial resolution is specifically a global daily scale resolution.

7. The method according to claim 1, characterized in that The vegetation optical thickness data reconstruction model is specifically used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input into the model and the surface cover type corresponding to the specific area.

8. The method according to claim 7, characterized in that The land cover type corresponding to the specific area is determined by a global land cover type information product obtained by observing the global land cover situation using satellite remote sensing technology.

9. The method according to claim 1, characterized in that: The vegetation optical thickness data reconstruction model is specifically used to carry out corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input into the model and the surface cover type corresponding to the specific area, as well as the corresponding regional factors and environmental factors.

10. A vegetation optical thickness data reconstruction device based on partial convolutional neural network, characterized in that: The device comprises: An acquisition unit, used for acquiring initial vegetation optical thickness data to be reconstructed, wherein the initial vegetation optical thickness data is a data product obtained by inverting the vegetation optical thickness based on observation data of a satellite-borne microwave radiometer; A reconstruction unit, used for inputting the initial vegetation optical thickness data into a preset vegetation optical thickness data reconstruction model, so that the vegetation optical thickness data is subjected to corresponding vegetation optical thickness data reconstruction processing based on the initial vegetation optical thickness data, wherein the vegetation optical thickness data reconstruction model is obtained by training an initial model in advance with labeled sample vegetation optical thickness data, and the vegetation optical thickness data reconstruction model is used for conducting corresponding vegetation optical thickness data reconstruction processing based on the vegetation optical thickness data input by the model, so as to improve the spatiotemporal resolution of the vegetation optical thickness data input by the model, and the vegetation optical thickness data reconstruction model specifically adopts a model architecture of a partial convolutional neural network; The extraction unit is used to extract the vegetation optical thickness data reconstruction result corresponding to the initial vegetation optical thickness data output by the vegetation optical thickness data reconstruction model, and output it as the target vegetation optical thickness data.

Images