CN112906300B - Polarization SAR soil humidity inversion method based on double-channel convolutional neural network - Google Patents
Polarization SAR soil humidity inversion method based on double-channel convolutional neural network Download PDFInfo
- Publication number
- CN112906300B CN112906300B CN202110182168.9A CN202110182168A CN112906300B CN 112906300 B CN112906300 B CN 112906300B CN 202110182168 A CN202110182168 A CN 202110182168A CN 112906300 B CN112906300 B CN 112906300B
- Authority
- CN
- China
- Prior art keywords
- matrix
- polarization
- layer
- convolution
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000010287 polarization Effects 0.000 title claims abstract description 45
- 239000002689 soil Substances 0.000 title claims abstract description 42
- 238000000034 method Methods 0.000 title claims abstract description 24
- 238000013527 convolutional neural network Methods 0.000 title claims description 29
- 239000011159 matrix material Substances 0.000 claims description 30
- 238000012549 training Methods 0.000 claims description 22
- 210000002569 neuron Anatomy 0.000 claims description 17
- 239000013598 vector Substances 0.000 claims description 16
- 230000004913 activation Effects 0.000 claims description 9
- 238000012795 verification Methods 0.000 claims description 8
- 238000010586 diagram Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 7
- 230000004927 fusion Effects 0.000 claims description 7
- 238000011049 filling Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000002360 preparation method Methods 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 238000000354 decomposition reaction Methods 0.000 claims description 2
- 238000001914 filtration Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 230000001902 propagating effect Effects 0.000 claims description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims 1
- 238000005388 cross polarization Methods 0.000 claims 1
- 238000010606 normalization Methods 0.000 claims 1
- 238000013528 artificial neural network Methods 0.000 abstract description 16
- 239000000523 sample Substances 0.000 description 13
- 230000006870 function Effects 0.000 description 7
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 230000002262 irrigation Effects 0.000 description 2
- 238000003973 irrigation Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001054 cortical effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000001704 evaporation Methods 0.000 description 1
- 230000008020 evaporation Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 238000001556 precipitation Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Geometry (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a polarized SAR soil humidity inversion method based on a dual-channel convolution neural network, which aims at the problems that the soil humidity inversion accuracy of a polarized SAR image is not high and sample data is limited. The main implementation object of the invention is the full polarization SAR image acquired by the synthetic aperture radar, and the main work is to estimate the soil humidity. Experiments show that compared with the traditional neural network, the inversion accuracy is improved by 10.88%, the root mean square error is reduced by 3.2356, and the determinable coefficient is improved by 0.6633.
Description
Technical Field
The invention relates to a polarized SAR soil humidity inversion method based on a double-channel convolutional neural network, and belongs to the field of quantitative inversion of polarized synthetic aperture radars.
Background
High-resolution polarized synthetic aperture radar (PolSAR) data plays an important role in inversion of soil moisture, the earth surface soil moisture is a basic condition for growth and development of crops, and before crop irrigation, the soil moisture is an important index for drought monitoring; after the crops are irrigated, the change condition of soil moisture is an important basis for evaluating the irrigation effect. Soil moisture is also a very important variable in studying land water circulation and energy circulation. The method can influence the distribution proportion of net radiant energy converted into latent heat and sensible heat, and can influence the proportion of precipitation converted into permeation, runoff and evaporation, so that the water content of the accurately obtained soil can be reasonably utilized for the land, and the production level and the production quality can be improved. At present, the accurate measurement of the soil moisture content is carried out by adopting the traditional methods such as probes and the like. The traditional methods can accurately measure the local soil moisture content, but consume a great deal of manpower and material resources, and are not suitable for large-scale soil moisture extraction. PolSAR can make up for the defects of the traditional measurement method by virtue of high sensitivity to soil humidity, and becomes a new method and means for monitoring soil humidity, so far, many researches have been made on using SAR data generation theory, experience and semi-experience back scattering models for estimating soil humidity. Theoretical models include Physical Optics (PO) models, metrology optics (GO) models, and Integral Equation Models (IEM), among others.
With the rise of neural networks, modeling between polarization parameters and soil humidity by using the neural networks becomes an important means for inversion of soil humidity, but the traditional neural networks only use the polarization parameters of sample points. In recent years, deep learning is the most hot technique in the current age, and convolutional neural networks (Convolutional Neural Network, CNN) are an important branch of deep learning, and are widely used in the image field because of their superior feature extraction capability. The convolution neural network is applied to the inversion of the polarized SAR soil humidity, and the nonlinear relation between the polarization parameters and the soil humidity is fitted by adopting the two-channel convolution neural network, so that the inversion accuracy is improved.
Disclosure of Invention
The invention mainly aims at solving the problems of low accuracy and limited samples in soil moisture inversion by using polarized SAR images, and provides a polarized SAR soil moisture inversion method based on a double-channel convolution neural network and a coarse-fine granularity inversion system applied to various scenes based on the neural network. According to the method, a double-channel convolutional neural network is used, the space characteristics of a polarized SAR image are fully utilized, the inversion accuracy is improved while training samples are reduced, dropout is introduced into a convolutional layer and a full-connection layer, the robustness of the method is improved, and a coarse-granularity classification network and a fine-granularity regression network are designed through the network. The main implementation object of the invention is the full polarization SAR image acquired by the synthetic aperture radar, and the main work is to estimate the soil humidity. Compared with the traditional neural network, the experimental result shows that the average inversion precision is improved by 10.88%, the root mean square error is reduced by 3.2356, and the determinable coefficient is improved by 0.6633.
The technical scheme of the invention specifically mainly comprises the following contents:
1. extracting polarization parameters: performing fine Lee filtering treatment with window size of 7*7 on the PolSAR image, extracting 6 polarization parameters, and applying to soil humidity inversion
2. Training sample and verification sample preparation: sample partitioning for polarization parameters using 7*7 sliding window
3. Establishing a double-channel convolutional neural network and training: the invention constructs a double-channel convolutional neural network, divides 6 polarization parameters in each training sample into two groups, respectively sends the two groups of the training samples into the two channels of convolutional neural network for training, and generates a prediction model after the features of the two channels are fused after the features are extracted.
4. Predicting the verification sample: and predicting the verification samples through the trained double-channel convolutional neural network, and calculating the average accuracy of the classification network, the root mean square error of the regression network and the determinable coefficient.
Drawings
FIG. 1 is a schematic diagram of training and validation sample preparation.
FIG. 2 is a block diagram of a two-channel convolutional neural network.
FIG. 3 is an overall flow chart for SAR soil moisture inversion based on dual channel convolutional neural network polarization.
Detailed Description
The present invention will be described in detail below with reference to the drawings and examples.
The polarized SAR soil humidity inversion method based on the double-channel convolutional neural network comprises the following steps,
step 1, extracting polarization parameters from a PolSAR image:
for fully polarized data, the scattering matrix of the target is:
wherein S is HH 、S VV Is the echo power of the homopolar channel, S HV 、S VH Is the echo power of the cross-polarized channels. When the reciprocity theorem is satisfied, S HV ＝S VH 。
Through [ S ]]The matrix can obtain the polarization covariance C of the target point 3 A matrix.
Selecting C 3 Elements in the matrix:<|S HH | 2 >、<|S VV | 2 >as the first two polarization parameters, wherein delta is made HH ＝<|S HH | 2 >、δ VV ＝<|S VV | 2 >And calculates the co-polarization ratio delta HH /δ VV As a third polarization parameter.
Matrix-transforming polarization covariance matrix into polarization coherence matrix T 3 ：
T 3 ＝U 3 C 3 U 3 -1 (3)
For T 3 The matrix carries out the decomposition of the eigenvalues and eigenvectors:
in the method, in the process of the invention,is T 3 Eigenvectors, lambda of matrix i Is T 3 Eigenvalues of the matrix. The scattering probability p can be obtained by normalizing the absolute scattering amplitude of the eigenvalue i ：
Using probability p i Characteristic value lambda i The fourth and fifth polarization parameters used can be obtained: scattering entropy (H) and anisotropy (a), defined as:
the scattering entropy (H) represents the degree of scattering polarization of the target point, and the degree of anisotropy (a) defines the relationship between the second characteristic value and the third characteristic value.
The sixth polarization parameter can be obtained by the eigenvector through the target scattering mechanism alpha i The average scatter angle α can be obtained, which is defined as:
alpha may represent the type of mean scattering process for the target point.
The extracted 6 polarization parameters have certain nonlinear relation with the soil humidity, so that the soil humidity is inverted by using the polarization parameters.
Step 2, training samples and manufacturing verification samples:
after extracting characteristic parameters of experimental data, 6 characteristic parameter matrixes can be obtained and are obtained according to H, A, alpha and delta HH 、δ VV 、δ HH /δ VV M x n x 6, where m, n are the width and height of the matrix and 6 are the depth. Next, as shown in fig. 1, a 7×7×6 sliding window is used to divide the matrix pixel by pixel, the matrix extracted by each window is the sample, and the center point of the sample corresponds toSoil moisture is the label of this point. According to the invention, 6000 samples are randomly selected from each type of humidity samples, 10% of the samples are randomly selected as training samples, and 90% of the samples are selected as verification samples.
Step 3, establishing a double-channel convolutional neural network and training:
convolutional neural networks are inspired by the visual imaging principle, which utilizes different convolutional kernels to replace the perception of visual cortical cells on a target object, and then the obtained perception result is subjected to a nonlinear activation function to obtain the complex characteristics of the target object. The two-channel convolutional neural network designed by the invention is shown in figure 2. The 6 polarization parameters of the sample are divided into two groups and respectively put into two channels, wherein for the three polarization parameters of H, A and alpha of the first channel, 4 convolution layers and one full connection layer are selected. Second channel input delta HH 、δ VV 、δ HH /δ VV The three parameters are two convolution layers and one full connection layer. The nature of the convolution layer (conv layer) is a filter that performs local feature extraction on the input data, where the input data I is convolved with a set of predetermined filters W and added to the offset b. And finally, the convolution addition result is passed through a nonlinear activation function, as shown in a formula (23), so that the local feature F can be obtained.
Wherein f (·) is a nonlinear activation function that is a modified linear unit (ReLU), the ReLU function can well avoid gradient dissipation and can reduce training time [11] It is defined as:
f(x ij )＝max{0,x ij } (10)
the Full Connection Layer (FCL) maps the feature map after a plurality of convolution operations to a sample marking space, so that the influence of the feature position on classification can be greatly reduced. And the Dropout layer is mixed between every two layers, and the Dropout layer has the function of stopping the operation of the activation value of a certain neuron with a certain probability p when propagating forwards, so that the model can be reduced to be overfitted and has stronger generalization. A corrected linear activation function (ReLU) is accessed after each layer. And carrying out feature extraction on the two channels, carrying out feature fusion on the channels, accessing the fused features into a full-connection layer, and finally carrying out classification and regression.
Specific setting details of the invention are shown in tables 1 and 2, the invention aims at H, A and alpha channels firstly passing through a first group of convolution kernels with 3, step length of 1 and filling 1 convolution layers to generate 8 feature maps with the size of 7*7, wherein filling 1 is to expand the periphery of original data by 0 circle, and the purpose of the invention is to fully utilize edge information of the data and keep the size of the data unchanged after convolution operation. And then passing through a convolution layer with a second group of convolution kernels of 3 and a step length of 1 to generate 16 characteristic diagrams with the size of 5*5. Then, through a third group of convolution kernels of 3, step sizes of 1 and a convolution layer filled with 1, 24 feature maps of 5*5 are generated. Through a fourth set of convolution layers with a convolution kernel of 3 and a step size of 1, 32 feature maps with a size of 5*5 are generated. Finally, the 32 feature maps of size 3*3 are changed into one-dimensional vectors with 120 neurons through a full connection layer. For delta HH 、δ VV 、δ HH /δ VV The channel first passes through a first set of convolution layers with a 3-ary convolution kernel and a 1-step size, generating 8 signature graphs of size 5*5. And then passing through a convolution layer with a second group of convolution kernels of 3 and a step length of 1 to generate 16 characteristic diagrams with the size of 3*3. Finally, the 24 3*3 feature maps are changed into one-dimensional vectors with 120 neurons through one full connection layer. After each operation, a Dropout layer is accessed and the ratio is set to 0.2. The two branches pass through the full connection layer to obtain a one-dimensional feature vector with 120 neurons extracted from respective data, and the two one-dimensional vectors are connected and fused to obtain a one-dimensional fusion feature vector with 240 neurons. The 240 obtained neurons are passed through a full connection layer, and the characteristics of the connection are rearranged, so that a new characteristic vector which is more satisfactory to 84 neurons after fusion is obtained. When performing classification tasks, output a vector through a full connected layer and output a vector with C neurons, where C is the classNumber of parts. When performing the regression task, the output passes through two fully connected layers, the first connected layer outputting a vector of 32 neurons and the second connected layer outputting the regression value. The total training of 100 epochs during training, the model with highest accuracy is stored, and an Adma optimization algorithm is adopted during training [26] Each batch size has 256 samples, the initial learning rate is set to be 0.001, the algorithm optimizes a random objective function based on a first-order gradient through low-order self-adaptive moment estimation, the calculation is efficient, the memory requirement is small, and the automatic adjustment of the learning rate can be realized generally without adjusting super parameters.
Table 1 two-channel classified convolutional neural network structure parameter table
Table 2 two-channel regression convolutional neural network structure parameter table
Step 4, predicting the verification sample:
in the invention, three indexes are adopted to evaluate the effect
1) Inversion accuracy
2) Root mean square error
3) Coefficient of determinability
Where N is the total number of validation samples, y p tame i t To invert the output value, y Eal (V) Is the true value of the soil humidity,the true value average value of the soil humidity is obtained.
The invention uses E-SAR full polarization data of the German North DEMIN region to train and verify, and calculates the three indexes. The experimental results are shown in tables 2 and 3. From experimental result analysis, the inversion precision of the two-channel convolutional neural network is 99.42%, the inversion precision of the traditional neural network is 88.54%, and compared with the traditional neural network, the two-channel convolutional neural network is improved by 10.88% in a same ratio. It can be obtained that the inversion accuracy of the method provided by the invention is obviously improved. The root mean square error is reduced by 3.2356 compared with the traditional neural network, and the determinable coefficient is improved by 0.6633 compared with the traditional neural network.
Table 2 different network inversion accuracy
| Network type | Inversion accuracy |
| Double-channel convolutional neural network | 99.42％ |
| Traditional neural network | 88.54％ |
TABLE 3 root mean square error and determinable coefficient for different networks
| Network type | Root mean square error | Coefficient of determinability |
| Double-channel convolutional neural network | 0.4194 | 0.9912 |
| Traditional neural network | 3.6550 | 0.3279 |
Claims (2)
1. The method for inverting the soil humidity based on the polarized SAR of the double-channel convolutional neural network is characterized by comprising the following steps of: the method comprises the following steps:
step (1) extracting polarization parameters: carrying out Lee filtering treatment with a window size of 7*7 on the PolSAR image, extracting 6 polarization parameters, and applying the parameters to soil humidity inversion;
and (2) training samples and verifying sample preparation: dividing samples by adopting 7*7 sliding windows for 6 polarization parameters extracted in the step (1);
step (3) establishing a double-channel convolutional neural network and training: constructing a double-channel convolutional neural network, dividing 6 polarization parameters in each training sample in the step (2) into two groups, respectively sending the two groups of polarization parameters into the two-channel convolutional neural network for training, and generating a prediction model after feature extraction and feature fusion of the two channels;
and (4) predicting the verification sample: predicting the verification sample through a trained double-channel convolutional neural network, and calculating average inversion accuracy, root mean square error and determinable coefficient;
in step (1), for the fully polarized data, the scattering matrix of the target is:
wherein S is HH 、S VV Is the echo power of the homopolar channel of polarized SAR data used for soil humidity inversion, S HV 、S VH The echo power of a polarized SAR data cross polarization channel used for soil humidity inversion is obtained; when the reciprocity theorem is satisfied, S HV ＝S VH ；
Through [ S ]]Matrix to obtain polarization covariance C of target point 3 A matrix;
at C 3 Representing conjugate operations in a matrix, whereinIs S HH Conjugate complex number of->Is S VV Complex conjugate of (2); selecting C 3 Elements in the matrix:<|S HH | 2 >、<|S VV | 2 >as the first two polarization parameters, wherein delta is made HH ＝<|S HH | 2 >、δ VV ＝<|S VV | 2 >And calculates the co-polarization ratio delta HH /δ VV As a third polarization parameter;
matrix-transforming polarization covariance matrix into polarization coherence matrix T 3 ：
T 3 ＝U 3 C 3 U 3 -1 (3)
Wherein C is 3 U as polarization covariance matrix 3 As unitary matrix, it is defined as:
for T 3 The matrix carries out the decomposition of the eigenvalues and eigenvectors:
in the method, in the process of the invention,is T 3 Eigenvectors of matrix, where α i Representing the scattering mechanism of the target beta i For the target azimuth angle phi i 、δ i And gamma i Is the target phase angle; i is 1, 2, 3, wherein lambda i Represents the ith eigenvalue; lambda (lambda) i Is T 3 Eigenvalues of the matrix and satisfy lambda 1 ≥λ 2 ≥λ 3 ；
The absolute scattering amplitude normalization is carried out on the eigenvalue to obtain scattering probability p i ：
Using probability p i Characteristic value lambda i Obtaining fourth and fifth polarization parameters used: the scattering entropy H and the anisotropy degree a are defined as:
the scattering entropy H represents the scattering polarization degree of the target point, and the anisotropy degree A defines the relation between the second characteristic value and the third characteristic value;
the sixth polarization parameter is obtained by the eigenvector and is obtained by the target scattering mechanism alpha i The average scatter angle α is obtained, defined as:
the average scattering angle α represents the type of average scattering process of the target point; the extracted 6 polarization parameters have a certain nonlinear relation with the soil humidity, so that inversion is carried out on the soil humidity by utilizing the polarization parameters;
in the step (2), after extracting the characteristic parameters of the experimental data, a matrix of 6 characteristic parameters is obtained, and the matrix is processed according to H, A, alpha and delta HH 、δ VV 、δ HH /δ VV Is constructed as a large matrix of m x n x 6, where m, n are the width and height of the matrix and 6 is the depth; then, a 7 multiplied by 6 sliding window is adopted to scratch a large matrix pixel by pixel, the matrix extracted by each window is a sample, and the soil humidity corresponding to the sample center point is the label of the sample center point;
in the step (3), the 6 polarization parameters of the sample are divided into two groups and respectively put into two channels, wherein for the three polarization parameters of H, A and alpha of the first channel, 4 convolution layers and one full connection are selectedA joint layer; second channel input delta HH 、δ VV 、δ HH /δ VV The three parameters are two convolution layers and a full connection layer; the essence of the convolutional layer conv layer is a filter, which performs local feature extraction on the input data, and in the convolutional layer, the input data I i And a set of preset filters W ij Convolving with offset b j Adding; finally, the convolution addition result is passed through a nonlinear activation function, as shown in a formula (10), so as to obtain local characteristics; wherein F is j The j-th feature diagram is output after convolution, M is the channel number of the input data;
wherein f (·) is a nonlinear activation function that corrects the linear unit ReLU, which avoids gradient dissipation and reduces training time, defined as:
f(x)＝max{0,x} (11)
wherein x represents a feature map obtained after convolution;
the full connection layer FCL maps the characteristic diagram after a plurality of convolution operations to a sample marking space, and is mixed with a Dropout layer between every two layers, wherein the Dropout layer has the function of stopping the operation of an activation value of a certain neuron with a certain probability p when propagating forwards; accessing a corrected linear activation function ReLU behind each layer; carrying out feature extraction on the two channels, carrying out feature fusion on the channels, accessing the fused features into a full-connection layer, and finally selecting a softmax classifier for classification by an output layer;
for H, A and alpha channels, firstly, a first group of convolution layers with convolution kernels of 3 and step length of 1 and filling 1 are passed through to generate 8 feature graphs with the size of 7*7, wherein the filling 1 is that the periphery of original data is subjected to one circle of expansion by 0; then, 16 characteristic graphs with the size of 5*5 are generated through a convolution layer with the second group of convolution kernels of 3 and the step length of 1; then, generating 24 characteristic graphs with the size of 5*5 through a convolution layer with the third group of convolution kernels of 3 and the step length of 1 and filling 1; warp yarnGenerating 32 characteristic graphs with the size of 3*3 by a convolution layer with the fourth group of convolution kernels of 3 and the step length of 1; finally, changing the 32 feature maps with the size of 5*5 into one-dimensional vectors with 120 neurons through a full connection layer; for delta HH 、δ VV 、δ HH /δ VV The channel firstly passes through a first group of convolution layers with convolution kernels of 3 and step length of 1 to generate 8 characteristic diagrams with the size of 5*5; then, 16 characteristic graphs with the size of 3*3 are generated through a convolution layer with the second group of convolution kernels of 3 and the step length of 1; finally, changing 24 3*3 feature maps into one-dimensional vectors with 120 neurons through a full connection layer; after each operation, a Dropout layer is connected, and the ratio is set to be 0.2; the two branches are subjected to full connection layer to obtain a one-dimensional feature vector with 120 neurons extracted from respective data, and the two one-dimensional vectors are connected and fused to obtain a one-dimensional fusion feature vector with 240 neurons; the obtained 240 neurons are passed through a full connection layer, the characteristics of the connection are rearranged, a new characteristic vector which is more in line with the requirements of 84 neurons after fusion is obtained, and finally, the output layer outputs a vector with C neurons through the full connection layer, wherein C is the category number; and training 100 epochs in total during training, storing a model with highest accuracy, and adopting an Adma optimization algorithm during training.
2. The dual-channel convolutional neural network polarization-based SAR soil humidity inversion method as claimed in claim 1, wherein the method is characterized by comprising the following steps: in the step (4), three indexes are adopted to evaluate the effect; wherein the inversion accuracy IA is used to describe the accuracy of the classification, defined as equation (12), the deviation between the predicted value and the true value is represented as root mean square error RMSE, defined as equation (13); the determinable coefficient r 2 Judging the fitting degree of the model by the formulas (14) - (16);
1) Inversion accuracy
2) Root mean square error
3) Coefficient of determinability
Where N is the total number of validation samples,for inverting the output value +.>True soil moisture value>The true value average value of the soil humidity is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110182168.9A CN112906300B (en) | 2021-02-09 | 2021-02-09 | Polarization SAR soil humidity inversion method based on double-channel convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110182168.9A CN112906300B (en) | 2021-02-09 | 2021-02-09 | Polarization SAR soil humidity inversion method based on double-channel convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112906300A CN112906300A (en) | 2021-06-04 |
| CN112906300B true CN112906300B (en) | 2023-11-21 |
Family
ID=76123348
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110182168.9A Active CN112906300B (en) | 2021-02-09 | 2021-02-09 | Polarization SAR soil humidity inversion method based on double-channel convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112906300B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113406584B (en) * | 2021-06-22 | 2022-11-15 | 北京化工大学 | Cross calibration method for synthetic aperture radar |
| CN113744237A (en) * | 2021-08-31 | 2021-12-03 | 华中科技大学 | Deep learning-based automatic detection method and system for muck fluidity |
| CN114154886B (en) * | 2021-12-08 | 2024-05-10 | 重庆大学 | Land transition simulation method based on delay communication CA model |
| CN116821694B (en) * | 2023-08-30 | 2023-12-01 | 中国石油大学(华东) | Soil humidity inversion method based on multi-branch neural network and segmented model |
| CN117114627A (en) * | 2023-10-18 | 2023-11-24 | 日照市自然资源和规划局 | land resource management system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2006086566A2 (en) * | 2005-02-09 | 2006-08-17 | Inlight Solutions, Inc. | Methods and apparatuses for noninvasive determinations of analytes |
| CN105930664A (en) * | 2016-04-26 | 2016-09-07 | 中国科学院地理科学与资源研究所 | Method for estimating instantaneous surface emissivity from passive microwave data |
| WO2018028255A1 (en) * | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
| CN108564006A (en) * | 2018-03-26 | 2018-09-21 | 西安电子科技大学 | Based on the polarization SAR terrain classification method from step study convolutional neural networks |
-
2021
- 2021-02-09 CN CN202110182168.9A patent/CN112906300B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2006086566A2 (en) * | 2005-02-09 | 2006-08-17 | Inlight Solutions, Inc. | Methods and apparatuses for noninvasive determinations of analytes |
| CN105930664A (en) * | 2016-04-26 | 2016-09-07 | 中国科学院地理科学与资源研究所 | Method for estimating instantaneous surface emissivity from passive microwave data |
| WO2018028255A1 (en) * | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
| CN108564006A (en) * | 2018-03-26 | 2018-09-21 | 西安电子科技大学 | Based on the polarization SAR terrain classification method from step study convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| SVRM辅助的北斗GEO卫星反射信号土壤湿度反演方法;杨磊;吴秋兰;张波;梁勇;洪学宝;邹文博;;北京航空航天大学学报(第06期);论文全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112906300A (en) | 2021-06-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112906300B (en) | Polarization SAR soil humidity inversion method based on double-channel convolutional neural network | |
| Drummond et al. | Statistical and neural methods for site–specific yield prediction | |
| CN108710974B (en) | Water ammonia nitrogen prediction method and device based on deep belief network | |
| CN108981957A (en) | Submarine temperatures field reconstructing method based on self organizing neural network and Empirical Orthogonal Function | |
| CN107576949B (en) | SVDD radar target one-dimensional range profile identification method based on density weight and mixed kernel function | |
| CN104361346B (en) | Classification of Polarimetric SAR Image method based on K SVD and rarefaction representation | |
| Li et al. | Comparison of parameter uncertainty analysis techniques for a TOPMODEL application | |
| CN113255972B (en) | Short-term rainfall prediction method based on Attention mechanism | |
| CN112180369B (en) | Depth learning-based sea surface wind speed inversion method for one-dimensional synthetic aperture radiometer | |
| CN107301641A (en) | A kind of detection method and device of Remote Sensing Imagery Change | |
| CN110954860B (en) | DOA and polarization parameter estimation method | |
| CN102914774B (en) | Synthetic aperture interferometric radiometer image inversion method | |
| Huang et al. | A data-driven method for hybrid data assimilation with multilayer perceptron | |
| Sihvonen et al. | Spectral profile partial least-squares (SP-PLS): Local multivariate pansharpening on spectral profiles | |
| CN116299247B (en) | InSAR atmospheric correction method based on sparse convolutional neural network | |
| Skalová et al. | Comparison of three regression models for determining water retention curves | |
| CN113392871A (en) | Polarized SAR terrain classification method based on scattering mechanism multichannel expansion convolutional neural network | |
| CN116994144A (en) | High-resolution winter wheat yield estimation method, system, storage medium and electronic equipment based on synthetic aperture radar image | |
| CN109063760B (en) | Polarization SAR classification method based on random forest multi-scale convolution model | |
| CN114994674B (en) | Intelligent microwave staring correlated imaging method and equipment and storage medium | |
| CN107271981B (en) | Method for establishing target electromagnetic scattering data model | |
| CN114611699A (en) | Soil moisture downscaling method and device, electronic equipment and storage medium | |
| CN114047563A (en) | All-weather assimilation method of infrared hyperspectrum | |
| CN108830290B (en) | SAR image classification method based on sparse representation and Gaussian distribution | |
| CN111914402B (en) | Dynamic topology estimation system and method based on signal characteristics and topology change priori |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |