Deep learning-based SAR (Synthetic Aperture Radar) soil surface water content inversion method and device

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) soil surface water content inversion method and device based on deep learning, and the method comprises the steps: constructing a deep learning regression model for predicting the volume content of soil water through combining SAR data, auxiliary data and in-situ sensor data; a ConvNeXt model is adopted to construct a parallel convolution encoder to solve multi-source data with different input resolutions, and back scattering coefficients, spectral information, digital elevation model high-resolution data and soil sand content, sludge content, clay content and stacking density low-resolution data are fused into soil moisture estimation. And the inference capability of unknown data is improved. The model obtained by training can realize large-range and high-resolution automatic estimation of the water content of the soil surface layer. According to the invention, the influence of radiation transmission noise is avoided, the inversion precision of a vegetation coverage area is improved, and large-range high-resolution automatic processing of soil humidity products is realized.

Description

Technical Field

[0001] The present invention belongs to the field of synthetic aperture radar soil surface moisture content estimation, and in particular relates to a SAR soil surface moisture content inversion method and device based on deep learning. Background Art

[0002] The production of high-resolution soil moisture (SM) products using synthetic aperture radar (SAR) images has extensive scientific and practical application value. In addition to the radar configuration, the backscatter coefficient depends heavily on geophysical parameters, including soil dielectric constant, soil roughness, and vegetation structure. Accurate modeling of these physical parameters is crucial for SM estimation. Traditional soil surface scattering models and variant methods combined with vegetation scattering models are limited by the uncertainties of a large number of geophysical parameters, field calibration, and sub-pixel heterogeneity, making it difficult to generate high-resolution SM products on a large scale.

[0003] At present, deep learning (DL) has been successfully applied to the estimation of SM in SAR remote sensing images, such as artificial neural networks (ANN), long short-term memory (LSTM) models and convolutional neural networks (CNNs), but there are few records of applying deep learning to SM. Existing methods are interfered by the inherent speckle noise of SAR data, which makes it difficult to directly and effectively model radar signals and soil properties. In addition, the volume scattering and surface scattering of vegetation have a nonlinear relationship with the radar echo signal, which reduces the accuracy and stability of the inversion, makes it easy to misestimate the moisture content, and is difficult to achieve robust automated processing. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a SAR soil surface moisture inversion method and device based on deep learning. A soil surface moisture regression model is constructed based on the deep learning method. By integrating SAR backscattering coefficient, spectral information, elevation data, soil characteristics and other multi-source data, the shallow surface soil echo signal dominated by volume scattering under low vegetation coverage conditions is automatically extracted, thereby avoiding the influence of radiation transmission noise on the soil surface moisture inversion method, improving the inversion accuracy in vegetation coverage areas, and realizing the automatic processing of large-scale high-resolution soil moisture products.

[0005] To achieve the above object, on the one hand, the present invention provides a SAR soil surface moisture inversion method based on deep learning, the method comprising the following steps:

[0006] Based on dual-polarization SAR image data, a deep learning regression dataset is constructed by combining auxiliary data and water volume content data obtained by in-situ sensors at a preset soil depth.

[0007] A high-resolution encoder is built based on the soil moisture estimation model of ConvNeXt to extract deep features of three high-resolution data: backscatter coefficient, spectral characteristics, and digital elevation model.

[0008] A low-resolution encoder was built based on the soil moisture estimation model of ConvNeXt to extract soil texture characteristics of four low-resolution data: sand content, silt content, clay content, and bulk density.

[0009] A fusion decoder was constructed to fuse multi-source data features. A ConvNeXt-based soil moisture estimation model was trained on a deep learning regression dataset and then full-image inversion was performed to achieve automatic estimation of soil surface moisture content.

[0010] In a second aspect, the present invention further proposes a SAR soil surface moisture inversion device based on deep learning, the device comprising:

[0011] A data generation module is used to construct a deep learning regression data set based on dual-polarization SAR image data, combined with auxiliary data and water volume content data obtained by in-situ sensors at a preset soil depth;

[0012] High-resolution feature encoding module, which is used to build a high-resolution encoder based on the soil moisture estimation model of ConvNeXt to extract deep features of three high-resolution data: backscatter coefficient, spectral characteristics and digital elevation model;

[0013] The low-resolution feature encoding module is used to construct a low-resolution encoder based on the soil moisture estimation model of ConvNeXt to extract soil texture characteristics of four low-resolution data: sand content, silt content, clay content, and bulk density;

[0014] The feature fusion and training module is used to build a fusion decoder to perform multi-source data feature fusion, train a ConvNeXt-based soil moisture estimation model on a deep learning regression dataset, and then perform full-image inversion to achieve automatic estimation of soil surface moisture content.

[0015] In a third aspect, the present invention also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned deep learning-based SAR soil surface moisture inversion method.

[0016] In a fourth aspect, the present invention further proposes a storage medium, on which a SAR soil surface moisture inversion program is stored. When the SAR soil surface moisture inversion program is executed by a processor, the steps of the SAR soil surface moisture inversion method based on deep learning as described above are implemented.

[0017] The beneficial effects of the present invention are:

[0018] The present invention realizes automatic estimation of large-scale and high-resolution soil surface moisture content, thereby solving the problem in the prior art that field measured data calibration is required and is limited by speckle noise and vegetation cover, improves the accuracy of estimating soil surface moisture content, and realizes automatic processing of large-scale and high-resolution products. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the SAR soil surface moisture inversion method based on deep learning of the present invention;

[0020] Figure 2 This is a schematic diagram of the network architecture of the ConvNeXt model of the present invention;

[0021] Figure 3 It is a structural diagram of the basic block Block of the ConvNeXt model of the present invention;

[0022] Figure 4 Schematic diagram of the ConvNeXt model decoder network structure of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0024] like Figure 1 As shown, it is a flow chart of the SAR soil surface moisture inversion method based on deep learning of the present invention, which includes the following steps:

[0025] Step S10: Based on the dual-polarization SAR image data, a deep learning regression data set is constructed through auxiliary data and the water volume content obtained by the in-situ sensor at a preset soil depth.

[0026] It should be noted that the dual-polarization SAR image data uses C-band cross-polarization data. Since different bands of SAR data have different penetration into the ground surface, the soil surface moisture content in the method refers to the volume fraction of moisture content from the top soil layer to a depth of 5 cm. Since each input source has a different revisit interval and is not completely consistent with each other, the present invention anchors the data to the SAR data in time and sets time boundaries for other sources to obtain meaningful information. The following will introduce how to normalize the data to create a regression data set.

[0027] For the dual-polarization SAR image data, the ground-based multi-view image product is used to perform thermal noise removal and radiation calibration terrain correction preprocessing to obtain VV, VH and local incident angle data. A linear normalization scheme is used to normalize the VV, VH bands and local incident angle Angle data based on the maximum-minimum stretching method:

[0028] ,

[0029] ,

[0030] ,

[0031] in, , , These are the normalized VV, VH, and Angle data. , , The original VV, VH, and Angle data are as follows. and Respectively refer to the minimum and maximum valid values ​​of each type of data.

[0032] For auxiliary optical data, the atmospheric apparent reflectance product is used, and the visible, near-infrared and short-wave infrared components are selected. The following nonlinear normalization scheme is adopted, where the normalization is applied in logarithmic space, which ensures a wider dynamic range of the cloud-free image after normalization:

[0033] ,

[0034] in, is the band that needs to be scaled. is the original band before normalization, and After calculating the statistics of the data set, the mean and variance are logarithmic.

[0035] ,

[0036] in, This is the normalized band. Here all data are scaled to the range of 0-255 for subsequent calculations.

[0037] For the auxiliary digital elevation model data, the surface water capacity vector is constructed, and the linear normalization scheme is used to normalize the DEM data based on the maximum-minimum stretching method.

[0038] ,

[0039] in, and Refers to the normalized and original DEM data, and are the minimum and maximum effective elevation values, respectively. Here, and Set to 0 and 3000 respectively to accommodate more terrain.

[0040] For auxiliary soil texture data, four parameters, namely soil sand content, silt content, clay content and bulk density, were selected to simulate soil texture. A linear normalization scheme was used to normalize the soil texture number SOIL based on the maximum-minimum stretching method:

[0041] ,

[0042] in, and The normalized and original soil texture data are shown in order. and are the minimum and maximum effective soil texture values, respectively, where and Set them to 0 and 800 respectively.

[0043] Resample the auxiliary optical, digital elevation model and soil texture data to the resolution of the SAR image data. Align the above multi-source data with the actual measurements obtained by the in-situ sensors from the International Soil Moisture Network, and use the WGS-84 longitude and latitude coordinate system as the geographic coordinate system of the deep learning regression dataset. By setting ground control points and second-order polynomials, the root mean square error after alignment is reduced to less than two pixels. Finally, apply the above normalization and cropping operations to use the SAR image data and the auxiliary optical, digital elevation model, and soil texture data as the input sources of the model, and the actual measurements as the output values ​​of the model to create a TFRecord format deep learning regression dataset. Through the alignment, normalization and cropping operations, the errors caused by the dimensions of different source data can be effectively eliminated, numerical problems can be prevented, the stability of training can be increased, and the convergence of the algorithm can be accelerated.

[0044] Step S20: Based on the soil moisture estimation model of ConvNeXt, a high-resolution encoder is constructed to extract deep features of high-resolution data such as backscattering coefficient, spectral characteristics and elevation.

[0045] In the specific implementation, Figure 2 As shown in Figure 1, the high-resolution encoder HS-Encoder uses ConvNeXt as the basic architecture and is responsible for processing high-resolution input. Figure 3 The structure of ConvNeXt and basic blocks is shown. The basic block uses a residual structure design, which contains three convolution layers. In the first convolution layer, a depth-separable convolution with a kernel size of 7×7 is used, the number of channels is set to 96, and then layer normalization (LN) is used to process the intermediate features. In the second convolution layer, a normal convolution with a kernel size of 1×1 is used, the number of channels is set to 384, and then a Gaussian error linear unit (GELU) is used as the activation function. In the third convolution layer, a normal convolution with a kernel size of 1×1 is used, and the number of channels is set to 96. In the high-resolution encoder, a total of four stages are used to stack basic blocks, and the number of stacked basic blocks is set according to the ratio of 1:1:9:1. In the encoder, the number of basic blocks stacked from stage 1 to stage 4 is (3,3,27,3). The initial number of channels in each stage is set to (128, 256, 512, 1024).

[0046] HS-Encoder uses the activation function GELU to replace the most commonly used RELU function in convolutional neural networks, and only uses the activation function in the middle layer of the basic block, rather than after each fully connected layer or convolutional layer. In the depthwise separable convolution, the number of groups and channels are set to be the same to achieve a weighted summation operation similar to the self-attention mechanism in the Transformer. And the use of depthwise separable convolution can effectively reduce the amount of computation (FLOPs). By observing Figure 3 From the basic block structure shown, we can see that the module adopts the Inverted Bottleneck architecture, and the number of channels of the input feature map will increase first and then decrease. In the basic block of the ordinary ResNet model, the 3×3 convolution structure is usually in the middle layer, while in ConvNeXt, it is advanced to the first layer and a larger convolution kernel size is used to increase the size of the receptive field.

[0047] In previous convolutional neural networks, the initial stem layer (Stem) is usually composed of a convolution layer with a convolution kernel size of 7×7 and a stride of 2 and a maximum pooling downsampling with a stride of 2, and the height and width are downsampled by 4 times. However, the stem layer of HS-Encoder uses a convolution layer with a convolution kernel size of 3×3 and a stride of 4 to reduce the size of the feature map.

[0048] In summary, the high-resolution encoder processes three types of high-resolution input data, namely backscatter coefficient, spectral features, and digital elevation, based on the ConvNeXt architecture. First, the initial part of the encoder, namely the backbone layer, reduces the size of the feature map through a convolution layer with a convolution kernel size of 3×3 and a stride of 4 to achieve preliminary feature extraction and downsampling. Subsequently, the encoder learns the multi-scale features of the input image by integrating four stages, each of which stacks a different number of basic blocks. In these basic blocks, the activation function GELU is only used in the middle layer of the basic block to enhance the nonlinear expression ability of the model. In addition, the use of depthwise separable convolution not only simulates the self-attention mechanism in the Transformer, but also effectively reduces the amount of computation. After this series of processing, the HS-Encoder is able to extract deep features from the input high-resolution data, which are crucial for tasks such as understanding surface conditions and estimating soil moisture content.

[0049] Step S30: Based on the ConvNeXt soil moisture estimation model, a low-resolution encoder is constructed to extract soil texture characteristics of low-resolution data such as soil sand content, silt content, clay content, and bulk density, so that the model can learn rich soil attribute information.

[0050] It should be noted that Figure 2 The figure shows the position of the low-resolution encoder LS-Encoder in the overall model framework. Similar to HS-Encoder, LS-Encoder also uses ConvNeXt as the basic architecture. Low-resolution images have fewer input pixels, which contain richer semantic features, so the size of the basic block required by LS-Encoder is reduced here.

[0051] The low-resolution encoder also uses four stages to stack basic blocks, but the number of stacked basic blocks is set in the ratio of 1:1:3:1. In the encoder, the number of basic blocks stacked from stage 1 to stage 4 is (3,3,9,3). The initial number of channels in each stage is set to (96, 192, 384, 768).

[0052] The low-resolution encoder LS-Encoder processes four features of soil: sand content, silt content, clay content, and bulk density based on the ConvNeXt architecture. These key features are closely related to the physical properties of soil and can reflect the pore structure and permeability of soil. First, the low-resolution encoder reduces the size of the feature map through a backbone layer containing convolutions of size 3×3 and stride 4 to achieve preliminary feature extraction and downsampling. Then, the multi-scale features of the input image are learned by integrating four stages, each of which stacks a different number of basic blocks. Compared with the HS-Encoder, the LS-Encoder stacks fewer basic blocks in each stage. Based on multiple convolutional layers and activation functions of the ConvNeXt architecture, the model can learn complex patterns and relationships of soil texture. These layers are able to capture the subtle differences in soil texture and encode this information into feature vectors that can be used for further analysis.

[0053] The high-resolution encoder HS-Encoder and the low-resolution encoder LS-Encoder form the structure of a parallel convolutional encoder to cope with data from different sources and different resolutions.

[0054] Step S40: Extract the features of SAR image data and auxiliary data through the above-mentioned parallel convolution encoder, build a fusion decoder to realize multi-source data fusion, and finally train the ConvNeXt-based soil moisture estimation model on the deep learning regression dataset and perform full-image inversion to realize large-scale high-resolution automatic estimation of soil surface moisture content.

[0055] It should be noted that Figure 2 The parallel convolutional encoder backend shown is connected to a fusion decoder, such as Figure 4 As shown in the figure, for the feature maps from the parallel convolutional encoder, they need to be merged into one feature map through a concatenation operation, and then the fully connected layer, batch normalization layer, activation layer and Dropout layer are stacked to form a fusion decoder. The proposed fusion decoder adopts a serial structure. First, the concatenated feature map is passed through N layers of convolutional structure to obtain deep fusion features. Then these high-level semantic features are mapped to soil moisture prediction values ​​through fully connected layers. It is worth noting that in the second step, only the Dropout layer and the fully connected layer are used, and the activation layer and batch normalization layer are not used. The fusion decoder realizes the nonlinear fusion of the serial input features, which can be configured by specifying the number of output channels of each module and the number of modules.

[0056] In a second aspect, the present invention further proposes a SAR soil surface moisture inversion device based on deep learning, wherein each module included in the device can execute each of the above steps. Specifically, the device includes:

[0057] A data generation module is used to construct a deep learning regression data set based on dual-polarization SAR image data, combined with auxiliary data and water volume content data obtained by in-situ sensors at a preset soil depth;

[0058] High-resolution feature encoding module, which is used to build a high-resolution encoder based on the soil moisture estimation model of ConvNeXt to extract deep features of three high-resolution data: backscatter coefficient, spectral characteristics and digital elevation model;

[0059] The low-resolution feature encoding module is used to construct a low-resolution encoder based on the soil moisture estimation model of ConvNeXt to extract soil texture characteristics of four low-resolution data: sand content, silt content, clay content, and bulk density;

[0060] The feature fusion and training module is used to build a fusion decoder to perform multi-source data feature fusion, train a ConvNeXt-based soil moisture estimation model on a deep learning regression dataset, and then perform full-image inversion to achieve automatic estimation of soil surface moisture content.

[0061] In a third aspect, the present invention also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned deep learning-based SAR soil surface moisture inversion method.

[0062] In a fourth aspect, the present invention further proposes a storage medium, on which a SAR soil surface moisture inversion program is stored. When the SAR soil surface moisture inversion program is executed by a processor, the steps of the SAR soil surface moisture inversion method based on deep learning as described above are implemented.

[0063] It should be noted that, in the present invention, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0064] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0065] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0066] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A SAR soil surface moisture inversion method based on deep learning, characterized in that: The method comprises: Based on dual-polarization SAR image data, a deep learning regression dataset is constructed by combining auxiliary data and water volume content data obtained by in-situ sensors at a preset soil depth. A high-resolution encoder is built based on the soil moisture estimation model of ConvNeXt to extract deep features of three high-resolution data: backscatter coefficient, spectral characteristics, and digital elevation model. A low-resolution encoder was built based on the soil moisture estimation model of ConvNeXt to extract soil texture characteristics of four low-resolution data: sand content, silt content, clay content, and bulk density. A fusion decoder was constructed to fuse multi-source data features. A ConvNeXt-based soil moisture estimation model was trained on a deep learning regression dataset and then full-image inversion was performed to achieve automatic estimation of soil surface moisture content.

2. The SAR soil surface moisture inversion method based on deep learning as claimed in claim 1, characterized in that: The deep learning regression data set is constructed based on dual-polarization SAR image data, combined with auxiliary data and water volume content data obtained by in-situ sensors at a preset soil depth, including: Based on dual-polarization SAR image data, using ground-range multi-view image products, thermal noise removal and radiation calibration terrain correction preprocessing are performed to obtain VV, VH and local incidence angle data; Based on auxiliary optical data, the atmospheric apparent reflectance product is used to select visible, near-infrared and short-wave infrared components; Using auxiliary digital elevation model data, construct surface water capacity vectors; Based on auxiliary soil texture data, four parameters, namely soil sand content, silt content, clay content, and bulk density, were selected to simulate soil texture; The auxiliary optical, digital elevation model and soil texture data were resampled to the resolution of dual-polarization SAR image data, aligned with the actual measurements obtained by in situ sensors from the International Soil Moisture Network, normalized and then cropped to a fixed size.

3. The SAR soil surface moisture inversion method based on deep learning as claimed in claim 1, characterized in that: The high-resolution data in the deep learning regression dataset includes SAR data, optical data, and digital elevation model data; the low-resolution data in the deep learning regression dataset includes soil texture data, and the soil texture data includes soil sand content, silt content, clay content, and bulk density.

4. The SAR soil surface moisture inversion method based on deep learning as claimed in claim 1, characterized in that: The soil moisture estimation model based on ConvNeXt constructs a high-resolution encoder for extracting deep features of three high-resolution data: backscatter coefficient, spectral features, and digital elevation model. The features include: Set the number of layers of stacked basic blocks for high-resolution encoder stages 1 to 4; Set the convolution kernel size and stride size used for downsampling in the backbone layer to the same value; Set the convolution module to depthwise separable convolution with the same number of groups and channels; Based on the high-resolution data, deep feature extraction is performed on VV, VH and local incident angle data from SAR images, visible light, near infrared and short-wave infrared components from optical images, and digital elevation models.

5. The SAR soil surface moisture inversion method based on deep learning as claimed in claim 4 is characterized in that: The proposed high-resolution encoder consists of a backbone layer for initial downsampling and four stages with different numbers of basic blocks to learn multi-scale features of the input image.

6. The SAR soil surface moisture inversion method based on deep learning as claimed in claim 1, characterized in that: The soil moisture estimation model based on ConvNeXt constructs a low-resolution encoder for extracting soil texture characteristics of four low-resolution data, namely, soil sand content, silt content, clay content, and bulk density, including: Set the number of layers of stacked basic blocks for low-resolution encoder stages 1 to 4; Set the convolution kernel size, number, and stride for downsampling in the backbone layer; Use depth-wise separable convolution and set the number of groups to be the same as the number of channels; Based on the low-resolution data, soil texture features of the low-resolution data including soil sand content, silt content, clay content, and bulk density from soil grid products are extracted.

7. The SAR soil surface moisture inversion method based on deep learning as claimed in claim 1, characterized in that: The construction of the fusion decoder to fuse multi-source data features, training a ConvNeXt-based soil moisture estimation model on a deep learning regression data set, and then performing full-image inversion to achieve automatic estimation of soil surface moisture content includes: Stack the fully connected layer, batch normalization layer, and activation layer to form a fusion decoder, set the number of fusion decoder layers, and use Dropout to prevent overfitting; Set the loss function to mean square error and mean absolute error; Divide the deep learning regression dataset into training, testing, and validation sets according to a certain ratio; The learning rate, number of training rounds, and batch size hyperparameters are set to train and verify the ConvNeXt-based soil moisture estimation model.

8. A SAR soil surface moisture inversion device based on deep learning, characterized in that: The device comprises: A data generation module is used to construct a deep learning regression data set based on dual-polarization SAR image data, combined with auxiliary data and water volume content data obtained by in-situ sensors at a preset soil depth; High-resolution feature encoding module, which is used to build a high-resolution encoder based on the soil moisture estimation model of ConvNeXt to extract deep features of three high-resolution data: backscatter coefficient, spectral characteristics and digital elevation model; The low-resolution feature encoding module is used to construct a low-resolution encoder based on the soil moisture estimation model of ConvNeXt to extract soil texture characteristics of four low-resolution data: sand content, silt content, clay content, and bulk density; The feature fusion and training module is used to build a fusion decoder to perform multi-source data feature fusion, train a ConvNeXt-based soil moisture estimation model on a deep learning regression dataset, and then perform full-image inversion to achieve automatic estimation of soil surface moisture content.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Among them, when one or more programs are executed by the one or more processors, the one or more processors implement the SAR soil surface moisture inversion method based on deep learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the SAR soil surface moisture inversion method based on deep learning as described in any one of claims 1-7.

Images