Snow depth inversion method and system based on fusion of underlying surface information of sar image

Abstract

The application discloses a snow depth inversion method and system based on underlying surface information fusion of a SAR image, relates to the technical field of ice and snow monitoring based on satellite remote sensing technology, and comprises the following steps: collecting orbit data, pre-processing the collected orbit data, removing thermal noise and speckle noise, performing terrain correction, and obtaining a backscattering characteristic parameter and a polarization characteristic parameter according to a radar equation for radiation calibration, performing feature extraction on the corrected data, optimizing a feature data set, constructing a neural network model, inputting the optimized feature data set into a residual neural network, training to obtain a snow depth inversion model, and introducing a time variable to predict the snow depth. The application effectively solves the problems of snow depth inversion and prediction by fusing advanced SAR data processing technology and a machine learning method, and provides an important basis for decision support.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ice and snow monitoring based on satellite remote sensing technology, and particularly relates to a method and system for snow depth inversion based on SAR image and underlying surface information fusion. BACKGROUND

[0002] Snow is a main component of the surface hydrological system, and snow cover can effectively maintain the ground temperature in winter and provide water resources for next year's biology after melting. Snow depth is an important input variable for a large number of climate, hydrology, agriculture and ecology models. Accurate acquisition of snow depth is helpful for rational utilization of snow resources, and is very important for human production activities and research on global climate change. Traditional snow depth measurement methods usually use fixed instruments at ground weather stations or portable devices for field experiment measurement, and these methods usually require a large amount of manpower and material resources and a long time. With the development of satellite remote sensing technology, remote sensing observation data can be used to observe regional and global scale surface snow changes in real time and quickly.

[0003] At present, snow depth inversion based on satellite remote sensing technology mainly includes microwave remote sensing and optical remote sensing. In the microwave remote sensing inversion of snow depth, there are passive microwave remote sensing and active microwave remote sensing inversion. The principle of optical remote sensing is mainly to extract relevant data according to the relationship between the spectral characteristics of snow and the snow depth, but optical remote sensing is easily affected by solar radiation, clouds and fog, and it is difficult to explore the internal information of snow, so the accuracy of snow depth inversion is limited. Passive microwave can penetrate the snow surface layer and obtain the internal information of snow, and it is the most common data source for snow depth and other snow parameters research and long-term monitoring, but the existing passive microwave remote sensing image has low spatial resolution, which cannot meet the high-resolution and high-precision snow parameter research. Compared with optical remote sensing image and passive microwave remote sensing, active radar remote sensing usually has the ability of all-time and all-weather observation of the earth, and can penetrate the snow surface layer without being affected by clouds and fog. In recent years, this method has been gradually used to explore the internal information of snow and to invert snow depth. However, at present, the microwave backscattering coefficient is greatly affected by the underlying surface conditions, so it is suitable for small-scale snow depth inversion, and the accuracy of large-scale snow depth inversion needs to be improved. SUMMARY

[0004] In view of the above existing problems, the present application provides a method and system for snow depth inversion based on SAR image and underlying surface information fusion, to solve the problem that a large amount of manpower and material resources and a long time are required in the prior art, and regional and global scale surface snow changes cannot be observed in real time and quickly.

[0005] To solve the above technical problems, a method for snow depth inversion based on SAR image and underlying surface information fusion is provided, which comprises,

[0006] Collecting track data, preprocessing the collected track data, and removing thermal noise and speckle noise; performing terrain correction, and obtaining backscattering characteristic parameters and polarization characteristic parameters according to a radar equation, extracting features from the corrected data, and optimizing the feature data set; constructing a neural network model, inputting the optimized feature data set into a residual neural network, training to obtain a snow depth inversion model, and introducing a time variable to predict the snow depth.

[0007] As a preferred scheme of the snow depth inversion method based on SAR image-based underlying surface information fusion of the application, the collected track data includes satellite position data, satellite speed data, track attitude data, time data, track dynamics parameters and track error correction data, and the collected track data is preprocessed.

[0008] The satellite position data includes longitude, latitude, altitude, track inclination and ascending node longitude; the satellite speed data includes speed along the track and speed perpendicular to the track; the track attitude data includes pitch angle, yaw angle and roll angle; the time data includes ephemeris time and data collection time; and the track dynamics parameters include track semi-major axis, track eccentricity, perigee angle and mean anomaly.

[0009] As a preferred scheme of the snow depth inversion method based on SAR image-based underlying surface information fusion of the application, the preprocessing includes track correction, thermal noise removal, speckle filtering and decibel processing of the obtained track data.

[0010] The track correction includes error estimation by polynomial fitting of track error, formation of error data set by calculating the difference between actual track data and reference data, selection of polynomial order to fit error data, determination of polynomial coefficients by least square method, i.e., the transpose of matrix multiplied by matrix multiplied by coefficient vector is equal to the transpose of matrix multiplied by error data vector, calculation of polynomial value according to time stamp after the coefficients are determined, and addition of the value as a correction to the original track data for track correction.

[0011] The thermal noise removal includes linear interpolation of the thermal noise vector of any given azimuth time through two nearest distance noise vectors, derivation of two noise vectors from the detection range, and removal by linear interpolation.

[0012] The speckle filtering includes speckle filtering of SAR remote sensing images by filter and updating of center pixel value by local statistical characteristics of the image, and the filter window is 7x7.

[0013] The decibel processing formula is:

[0014] ε = 10 log 10 σ o

[0015] wherein σ o is a dimensionless backscattering coefficient, and ε is a normalized backscattering coefficient.

[0016] As a preferred scheme of the snow depth inversion method based on the SAR image-based underlying surface information fusion of the present application, the terrain correction includes eliminating the radiation brightness error by using the SCS+C correction method of the Lambert reflectivity model based on the elevation information DEM, so that the same ground objects with the same reflectivity have the same brightness value in the image.

[0017] The SCS+C correction method is expressed by the following formula:

[0018] ρ H = ρ T (cos θ p cos θ z +C λ ) / (cos γ i +C λ )

[0019] ρ T = a λ +b λ cos γ i

[0020] C λ = a λ / b λ

[0021] wherein ρ H is the horizontal ground reflection, ρ T is the slope ground reflection, θ p is the slope angle, θ z is the solar zenith angle, γ i is the solar incident angle, C λ is the correction coefficient, a λ is the slope of the regression line, and b λ is the intercept of the regression line.

[0022] The radiation scaling formula is expressed by the following formula:

[0023]

[0024] wherein P d is the backscattering intensity received by the sensor, P t is the transmission power, P n is the additional power, is the perspective antenna gain, is the receiving antenna gain, G is the current gain of the radar receiver p R is the distance propagation loss, and θ el θ is the antenna elevation angle oc L is the antenna azimuth angle s L is the atmospheric loss a L is the system loss, A is the scattering area, and λ is the wavelength o is the dimensionless backscattering coefficient.

[0025] As a preferred scheme of the snow depth inversion method based on the SAR image-based underlying surface information fusion according to the present application, the polarized characteristic parameter acquisition includes adopting dual-polarization ratio data, and calculating the polarization entropy, anisotropy and scattering angle through the eigenvalues of three coherent matrix sub-components obtained by the coherent matrix eigenvalue decomposition.

[0026] The polarized characteristic parameter acquisition formula is:

[0027]

[0028]

[0029] P i is the proportion of the eigenvalue corresponding to the i-th eigenvector in the total eigenvalue, λ i is the i-th eigenvalue, H is the polarization entropy, A is the anisotropy, α is the scattering angle, i is the variable index, λ1 is the volume scattering of the target, λ2 is the surface scattering of the target, and λ3 is the multiple scattering component of the target.

[0030] As a preferred scheme of the snow depth inversion method based on the SAR image-based underlying surface information fusion according to the present application, the optimized feature data set includes pre-processing the geographic feature data to obtain slope and aspect data, pre-processing the meteorological data, performing feature extraction on the corrected data, and performing feature optimization using the Pearson correlation coefficient.

[0031] The slope and aspect data acquisition formula is:

[0032]

[0033] Slope is the slope, Aspect is the aspect, Slope x is the slope in the horizontal direction of the two-dimensional image, Slope y is the slope in the vertical direction of the two-dimensional image.

[0034] The Pearson correlation coefficient formula is:

[0035]

[0036] wherein Y i and X i are characteristic parameters, and are average values of the characteristic parameters, n is the number of parameters, i is the variable index, PCC Y-X is the Pearson correlation coefficient; the Pearson correlation coefficient threshold is selected as 0.9, when the calculated correlation coefficient is greater than or equal to 0.9, the characteristic parameter is judged to be a parameter related to the snow depth, and is used as an input characteristic parameter of the snow depth inversion model.

[0037] As a preferred scheme of the snow depth inversion method based on SAR image underlying surface information fusion according to the present application, the predicted snow depth comprises that the snow depth inversion models are respectively constructed based on different neural network model frameworks, a residual neural network ResNet model is selected for training to obtain the snow depth inversion model, model training is performed, and the difference between the inversion model result and the measured value, i.e., the inversion accuracy, is obtained through testing, the model parameters are adjusted according to the result, the model parameters with the highest inversion accuracy are selected, the snow depth inversion model is constructed based on the parameters, and the time variable is input into a long short-term memory network LSTM model for training to obtain a snow depth long time sequence inversion model, and the snow depth is predicted.

[0038] The model training comprises defining a model structure, inputting an experimental data set optimized in characteristics, initializing a learning rate, an optimizer and an iteration number, capturing a nonlinear relationship between the input characteristic parameters and the measured snow depth, and obtaining the difference between the inversion model result and the measured value, i.e., the accuracy of the inversion result, through testing.

[0039] Another object of the present application is to provide a snow depth inversion system based on SAR image underlying surface information fusion, which combines SAR data processing technology and machine learning method to solve the problems of snow depth inversion and prediction; the system of the present application improves data quality through data acquisition and preprocessing of synthetic aperture radar images, including orbit data collection, thermal noise removal and speckle filtering; terrain correction and radiation calibration are performed to ensure data standardization; the system extracts and optimizes the advanced polarization characteristic parameters including the dual polarization ratio and the Cloude-Pottier polarization decomposition based polarization parameters to accurately describe the snow characteristics.

[0040] As a preferred scheme of the snow depth inversion system based on SAR image underlying surface information fusion according to the present application, it comprises a data acquisition module, a terrain correction and radiation calibration module, a characteristic extraction and optimization module, a neural network model construction and training module, and a snow depth prediction module.

[0041] The data acquisition module is configured to collect satellite position, speed, orbit attitude, time data and dynamic parameters, and perform orbit correction, thermal noise removal, coherent spot filtering and decibel processing, thereby reducing errors and noises in the data.

[0042] The terrain correction and radiation calibration module is configured to eliminate radiation brightness errors caused by terrain undulations by using the SCS+C correction method, ensure consistency of the reflectivity of different terrains, and convert the backscattering intensity received by the sensor into a dimensionless backscattering coefficient according to the radar equation.

[0043] The feature extraction and optimization module is configured to calculate polarization entropy, anisotropy and scattering angle by dual polarization ratio data and eigenvalue decomposition of a coherence matrix, obtain polarization information reflecting snow characteristics, extract slope and slope direction from elevation data, and extract environmental temperature and humidity data, and screen features by using the Pearson correlation coefficient to retain features related to snow depth.

[0044] The neural network model construction and training module is configured to define the structure of a residual neural network ResNet model, input the optimized feature data set, initialize the model parameters, capture the nonlinear relationship between the feature parameters and the snow depth by training, and evaluate the accuracy of the model.

[0045] The snow depth prediction module is configured to input a SAR remote sensing image based on the ResNet model obtained by training, obtain snow depth by inversion, introduce a time variable into an LSTM model, train a prediction model of snow depth changing with time, and perform long-term prediction of snow depth.

[0046] A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method for snow depth inversion based on SAR image-based underlying surface information fusion when executing the computer program.

[0047] A computer readable storage medium having a computer program stored thereon, characterized in that the computer program implements the steps of the method for snow depth inversion based on SAR image-based underlying surface information fusion when executed by a processor.

[0048] The present application is different from optical remote sensing technology which is susceptible to natural environmental conditions such as cloud, light, etc., and can work all day, all weather, and for a long time, obtaining large-scale data and rich image effective information, and the result of snow depth inversion based thereon is more stable and reliable; different from passive microwave remote sensing technology with low spatial resolution, the present application adopts synthetic aperture radar (SAR) remote sensing technology, which can obtain higher spatial resolution, so that the snow depth inversion is not affected by insufficient spatial resolution, and high-precision and high-resolution snow depth inversion can be carried out, the result of snow depth inversion based thereon is more reliable; and the information such as backscattering characteristics and polarization characteristics of the underlying surface is fused, the problem of limited snow depth inversion accuracy in a large scale range caused by the influence of the underlying surface condition on the SAR image is overcome, the influence of geographical factors and meteorological factors on the snow depth is considered, and a snow depth inversion model is constructed by combining the fusion of multi-source data, which is beneficial to the improvement of snow depth inversion accuracy; and a machine learning model with memory function is introduced, a long time series snow depth inversion model is constructed by combining multi-source data and time series, and the prediction of snow depth can be realized, which has a more in-depth application in disaster warning and the like. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0050] Figure 1 The overall flowchart of the underlying surface information fusion snow depth inversion method based on the SAR image provided by one embodiment of the present application.

[0051] Figure 2 The main network ResNet model schematic diagram of the snow depth inversion model of the underlying surface information fusion snow depth inversion method based on the SAR image provided by one embodiment of the present application.

[0052] Figure 3 The main network LSTM model schematic diagram of the long time series snow depth inversion model of the underlying surface information fusion snow depth inversion method based on the SAR image provided by one embodiment of the present application.

[0053] Figure 4 The system scheme flowchart of the underlying surface information fusion snow depth inversion system based on the SAR image provided by one embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described in detail below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0055] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in other embodiments that depart from the details described herein. In other instances, detailed descriptions of well-known methods and devices are not provided in order to avoid obscuring the present application.

[0056] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor does it mean that the embodiment is mutually exclusive with other embodiments.

[0057] The present application is described in detail in conjunction with the schematic diagram. In the detailed description of the embodiments of the present application, the cross-sectional view of the device structure is partially enlarged without the general proportion for the convenience of description, and the schematic diagram is only an example, which should not limit the scope of protection of the present application. In addition, three-dimensional spatial dimensions including length, width and depth should be included in actual manufacture.

[0058] Meanwhile, in the description of the present application, it should be noted that the terms "up, down, inner and outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the indicated device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first, second or third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0059] Unless otherwise specifically defined and limited, the terms "mounting, connecting, connection" in the present application should be understood broadly, for example: it can be fixed connection, detachable connection or integral connection; it can also be mechanical connection, electrical connection or direct connection, it can also be indirectly connected through intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0060] Embodiment 1, refer to Figures 1-3 For the first embodiment of the present application, the embodiment provides a snow depth inversion method based on the fusion of underlying surface information of SAR image, comprising:

[0061] S1: collecting track data, pre-processing the collected track data, and removing thermal noise and coherent speckle noise.

[0062] The collecting track data comprises collecting satellite position data, satellite speed data, track attitude data, time data, track dynamics parameters and track error correction data, and pre-processing the collected track data;

[0063] The satellite position data comprises longitude, latitude, height, track inclination and ascending node longitude; the satellite speed data comprises speed along the track and speed perpendicular to the track; the track attitude data comprises pitch angle, yaw angle and roll angle; the time data comprises ephemeris time and data collection time; the track dynamics parameters comprise track semi-major axis, track eccentricity, perigee angle and mean anomaly.

[0064] It should be noted that the pre-processing comprises track correction, thermal noise removal, coherent speckle filtering and decibel processing of the acquired track data;

[0065] The track correction comprises error estimation by polynomial fitting of track error, forming an error data set by calculating the difference between actual track data and reference data, selecting a polynomial order to fit the error data, determining the polynomial coefficients by least squares method, i.e. the transpose of the matrix multiplied by the matrix multiplied by the coefficient vector is equal to the transpose of the matrix multiplied by the error data vector, calculating the value of the polynomial according to the time stamp after the coefficient is determined, and adding the value as the correction amount to the original track data for track correction;

[0066] The thermal noise removal comprises obtaining the thermal noise vector for any given orientation time by linear interpolation of the two closest noise vectors, deriving two noise vectors from the detection range for the thermal noise vector, and removing from the detection range by linear interpolation;

[0067] The coherent speckle filtering comprises coherent speckle filtering of SAR remote sensing images by filter and updating the center pixel value using the local statistical properties of the image, and the filter window is 7x7;

[0068] The decibel processing formula is:

[0069] ε = 10log 10 σ o

[0070] where σ o is the dimensionless backscattering coefficient, and ε is the normalized backscattering coefficient.

[0071] S2: terrain correction is carried out, and radiation calibration is carried out according to the radar equation to obtain backscattering characteristic parameters and polarization characteristic parameters, the corrected data is carried out feature extraction, and the feature data set is optimized.

[0072] Further, the terrain correction includes eliminating radiation brightness error by using SCS+C correction method of Lambert reflectivity model based on elevation information DEM, so that the same ground objects with the same reflection properties have the same brightness value in the image.

[0073] The formula of the SCS+C correction method is:

[0074] ρ H =ρ T (cosθ p cosθ z +C λ ) / (cosγ i +C λ )

[0075] ρ T =a λ +b λ cosγ i

[0076] C λ =a λ / b λ

[0077] Wherein, ρ H is the horizontal ground reflection, ρ T is the slope ground reflection, θ p is the slope angle, θ z is the solar zenith angle, γ i is the solar incident angle, C λ is the correction coefficient, a λ is the slope of the regression line, and b λ is the intercept of the regression line.

[0078] The radiation calibration formula is:

[0079]

[0080] Wherein, P d is the backscattering intensity received by the sensor, P t is the transmission power, P n is the additional power, is the perspective antenna gain, is the receiving antenna gain, is the current gain of the radar receiver, G p is the processor constant, R is the distance propagation loss, θ el is the antenna elevation angle, and θoc is the azimuth angle of the antenna, L s is the loss of the atmosphere, L a is the loss of the system, A is the scattering area, λ is the wavelength, σ o is the dimensionless backscattering coefficient.

[0081] The polarization characteristic parameters include polarization entropy, anisotropy and scattering angle, and the polarization entropy, the anisotropy and the scattering angle are calculated by using the dual-polarization ratio data and eigenvalues of three sub-components of the coherence matrix obtained by eigen-decomposition of the coherence matrix;

[0082] The formula for obtaining the polarization characteristic parameters is:

[0083]

[0084] wherein, P i is the proportion of the eigenvalue corresponding to the i-th eigenvector in the total eigenvalue, λ i is the i-th eigenvalue, H is the polarization entropy, A is the anisotropy, α is the scattering angle, i is the variable index, λ1 is the volume scattering of the target, λ2 is the surface scattering of the target, and λ3 is the multiple scattering component of the target.

[0085] Further, the optimization of the feature data set includes preprocessing of geographical feature data to obtain slope and aspect data, preprocessing of meteorological data, feature extraction of the corrected data, and feature optimization using the Pearson correlation coefficient;

[0086] The auxiliary data preprocessing process considers the influence of meteorological factors and topography on snow depth, and introduces original elevation data and reanalysis meteorological data when constructing the experimental data set. The original elevation data can obtain geographical feature data such as slope and aspect after preprocessing, and the reanalysis meteorological data can obtain meteorological features such as environmental temperature, environmental humidity, snow temperature and snow density after preprocessing.

[0087] The formula for obtaining the slope and aspect data is:

[0088]

[0089] wherein, Slope is the slope, Aspect is the aspect, Slope x is the slope in the horizontal direction of the two-dimensional image, Slope y is the slope in the vertical direction of the two-dimensional image.

[0090] The formula for the Pearson correlation coefficient is:

[0091]

[0092] wherein, Y i and X iis a feature parameter, and is the average value of the feature parameter, n is the number of parameters, i is the variable index, PCC Y-X is the Pearson correlation coefficient; the Pearson correlation coefficient threshold is selected as 0.9, when the calculated correlation coefficient is greater than or equal to 0.9, the feature parameter is judged to be a parameter with strong correlation with the snow depth, and is used as an input feature parameter of the snow depth inversion model.

[0093] S3: Constructing a neural network model, inputting the optimized feature data set into a residual neural network, training to obtain a snow depth inversion model, and introducing a time variable to predict the snow depth.

[0094] Further, the predicted snow depth includes that the snow depth inversion model is constructed based on different neural network model frameworks, a residual neural network ResNet model is selected for training to obtain the snow depth inversion model, the model is trained, and the difference between the inversion model result and the measured value, that is, the inversion accuracy, is obtained through testing, the model parameters are adjusted according to the result, the model parameters with the highest inversion accuracy are selected, the snow depth inversion model is constructed based on the parameters, and the time variable is input into a long short-term memory network LSTM model for training to obtain a snow depth long time sequence inversion model, and the snow depth is predicted.

[0095] The backbone network of the snow depth inversion model is a Resnet model, such as Figure 2 The model is composed of a standard residual module, a down-sampling residual module and a full connection layer; in the standard residual module, the input data is subjected to feature transformation through two 3x3 convolution layers to obtain the output after feature transformation; at the same time, the input data is directly transmitted through a straight connection edge, and then added with the initial output to obtain the final output, in the down-sampling residual block, a 1x1 convolution layer is added to the straight connection edge, which is used to adjust the dimension of the input to match the dimension of the output, so as to perform addition operation, the formula is:

[0096] y=F(x,W i )+x

[0097] y=F(x,W i )+W s (x)

[0098] Wherein, W i is the weight parameter of the convolution layer in the residual block, x is the input data, y is the final output, F is the output after feature transformation, and W s is the weight parameter of the convolution layer.

[0099] The model training includes defining the model structure, inputting the optimized experimental data set, initializing the learning rate, optimizer and iteration number, capturing the nonlinear relationship between the input feature parameters and the measured snow depth, and obtaining the difference between the inversion model result and the measured value through testing, that is, the accuracy of the inversion result.

[0100] The backbone network of the snow depth prediction model is an LSTM model, as shown in Figure 3 The model is composed of a forgetting gate, an input gate and an output gate; in the forgetting gate, data is read and a value between 0 and 1 is output to each number in the cell state, 1 indicating complete retention and 0 indicating complete discard, for determining discarded information; in the input gate, the old cell state is multiplied by the forgetting gate to discard the determined discarded information, and then the new cell state generated in the input gate is added to obtain the updated cell state; in the output gate, it is determined which part of the cell state is output, the cell state is processed, and the final output content is obtained;

[0101] The forgetting gate formula is represented as:

[0102] f(t) = σ(W f ·[h(t-1), x(t)] + b f )

[0103] The input gate formula is represented as:

[0104] i(t) = σ(W i ·[h(t-1), x(t)] + b i )

[0105] C(t)' = tanh(W C ·[h(t-1), x(t)] + b C )

[0106] C(t) = f(t) · C(t-1) + i(t) · C(t)'

[0107] The output gate formula is represented as:

[0108] o(t) = σ(W o ·[h(t-1), x(t)] + b o )

[0109] h(t) = o(t) · tanh(C(t))

[0110] Wherein, f(t) is the forgetting gate function, σ is the activation function, W f is the forgetting gate weight parameter, x(t) is the current input data, and b fThe h(t-1) is an output of a hidden layer at a last time, the i(t) is an input gate function, the C(t)' is a new cell state generated by the input gate, the C(t) is an updated cell state, W i and W C The b i and b C are input gate bias parameters, the C(t-1) is an old cell state, the tanh is a hyperbolic tangent activation function, the o(t) is a control signal of an output gate, W o is an output gate weight parameter, b o is an output gate bias parameter, and the h(t) is a final output.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

[0112] Embodiment 2, which is an embodiment of the present application, provides a snow depth inversion method based on surface information fusion of SAR images. In order to verify the beneficial effects of the present application, scientific demonstration is carried out through experiments.

[0113] Firstly, Sentinel-1 synthetic aperture radar data is selected. The Sentinel-1 satellite load C-band SAR sensor provides dual-polarization capability of VV and VH polarization. The first-level data product of the satellite after preliminary processing and multi-view processing is used as SAR remote sensing image data, and the resolution is 20 m. The image data is preprocessed by using a remote sensing image processing platform ENVI, including orbit correction, thermal noise removal, radiation calibration, coherent spot filtering, terrain correction, decibel processing and polarization decomposition, etc. The backscattering characteristics and polarization characteristics are extracted from the radar image data. The original elevation data used in this example is SRTMDEM 90M geographic feature data, and the resolution is 90 m. After the preprocessing operation, the elevation, slope and slope direction and other geographic features are extracted from the original elevation data. The weather data used in this example is Rp5. This data provides historical monitoring data of most weather stations in the world, and the time resolution is 2 hours. After the preprocessing operation, the environmental temperature, environmental humidity, snow temperature and snow density and other weather characteristics are extracted from the data. The characteristics of SAR remote sensing images in snow, the characteristics of SAR remote sensing images in non-snow, and the geographic features and weather characteristics are integrated to form an original experimental data set:

[0114] X = [X1, X2, X3, X4]

[0115] Wherein, X1 is the backscattering and polarization features extracted from SAR remote sensing images in snow, X2 is the backscattering and polarization features extracted from SAR remote sensing images in no snow, that is, the underlying surface features, X3 is the geographic features extracted from the original elevation data, and X4 is the meteorological features extracted from the meteorological data;

[0116] Secondly, the Pearson correlation coefficient between each feature and the snow depth is calculated respectively. The original experimental data set after feature optimization is taken as the experimental data set. According to the ratio of 8:2, the training set and the test set are randomly divided and input into the snow depth inversion model with ResNet model as the backbone network. After training, the snow depth inversion model is obtained. The root mean square error RMSE and the determination coefficient R 2 Model evaluation:

[0117]

[0118] Wherein, d i is the actual observation value of snow depth, is the snow depth inversion value, is the actual average observation value of snow depth, RMSE is the root mean square error, R 2 is the determination coefficient, n is the number of parameters, and i is the variable index.

[0119] Through experiments, the RMSE of the snow depth inversion model is 2.03 cm, R 2 is 0.87, and the inversion accuracy is 85.49%. It shows that the snow depth inversion model constructed in this example can effectively capture the relationship between the snow depth and the input features.

[0120] On the basis of the above experimental data set, the time variable is supplemented, and the long-term trend of snow depth can be further analyzed. The newly constructed experimental data set is input into the snow depth prediction model with LSTM model as the backbone network. After training, the snow depth prediction model is obtained. Similarly, the model is evaluated by RMSE and R 2 , the RMSE is 2.57 cm, R 2 is 0.72, and the prediction accuracy is 82.63%. It shows that the snow depth prediction model constructed in this example can effectively predict the snow depth and has a wide application scenario in disaster warning.

[0121] It should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and they should be included in the scope of the claims of the present application.

[0122] Embodiment 3, refer to Figure 4 As a third embodiment of the present application, the embodiment provides a snow depth inversion system based on SAR image-based underlying surface information fusion, comprising a data acquisition module 100, a terrain correction and radiation calibration module 200, a feature extraction and optimization module 300, a neural network model construction and training module 400 and a snow depth prediction module 500.

[0123] The data acquisition module 100 is used to collect satellite position, speed, orbit attitude, time data and dynamic parameters, and to perform orbit correction, thermal noise removal, speckle filtering and decibel processing to reduce errors and noise in the data.

[0124] The terrain correction and radiation calibration module 200 is used to eliminate the radiation brightness error caused by the terrain undulation by using the SCS+C correction method, to ensure the consistency of the reflectivity of different terrains, and to convert the backscattering intensity received by the sensor into dimensionless backscattering coefficient according to the radar equation.

[0125] The feature extraction and optimization module 300 is used to calculate the polarization entropy, anisotropy and scattering angle by dual-polarization ratio data and coherent matrix feature decomposition, to obtain the polarization information reflecting the characteristics of snow, to extract the slope and slope direction from the elevation data, and to extract the environmental temperature and humidity data, to screen the features using the Pearson correlation coefficient, and to retain the features strongly related to the snow depth.

[0126] The neural network model construction and training module 400 is used to define the structure of the residual neural network ResNet model, to input the optimized feature data set, to initialize the model parameters, to capture the nonlinear relationship between the feature parameters and the snow depth by training, and to evaluate the accuracy of the model.

[0127] The snow depth prediction module 500 is used to input the SAR remote sensing image based on the ResNet model obtained by training, to obtain the snow depth by inversion, and to introduce the time variable into the LSTM model to train the prediction model of the snow depth changing with time, and to perform long-term prediction of the snow depth.

[0128] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

[0129] Embodiment 4, the fourth embodiment of the present application, which is different from the first three embodiments, is:

[0130] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the technical solutions that essentially contribute to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0131] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processor, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions, or in conjunction with these instructions execution systems, apparatuses, or devices. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport programs for use by an instruction execution system, apparatus, or device, or in conjunction with these instruction execution systems, apparatuses, or devices.

[0132] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium can even be paper or other suitable medium on which the program can be printed, as the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpreting, or otherwise processing, if necessary, in other suitable ways to be electronically obtained, and then stored in the computer memory.

[0133] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technology, known in the art, or combinations thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

Claims

1. A method for retrieving snow depth based on the fusion of underlying surface information from SAR images, characterized in that: The application relates to a method for predicting snow depth by using a neural network model. The method comprises the following steps: collecting track data, pre-processing the collected track data, and removing thermal noise and coherent speckle noise; carrying out terrain correction, obtaining backscattering characteristic parameters and polarization characteristic parameters according to a radar equation, carrying out feature extraction on the corrected data, and optimizing a feature data set; constructing a neural network model, inputting the optimized feature data set into a residual neural network, training to obtain a snow depth inversion model, and introducing a time variable to predict the snow depth; the terrain correction comprises: eliminating radiation brightness errors by using a SCS+C correction method based on a Lambert reflectivity model of DEM, so that the same ground objects have the same brightness values in the image; the SCS+C correction method is expressed by the following formula: p H = p T (cosθ p cosθ z +C λ ) / (cosγ i +C λ ) p T = a λ + b λ cosγ i C λ = a λ / b λ where ρ H is the horizontal surface reflection, ρ T is the slope surface reflection, θ p is the slope angle, θ z is the solar zenith angle, γ i is the solar incidence angle, C λ is the correction factor, a λ is the slope of the regression line, and b λ is the intercept of the regression line. the radiation calibration formula is expressed by the following formula: where P d is the backscatter intensity received by the sensor, P t is the transmit power, P n is the additional power, is the free space path loss, G is the receive antenna gain, is the current gain of the radar receiver, G p is a processor constant, R is the distance propagation loss, θ el is the antenna elevation angle, θ oc is the antenna azimuth angle, L s is the atmospheric loss, L a is the system loss, A is the scattering area, λ is the wavelength, σ o is the dimensionless backscatter coefficient; the polarization characteristic parameters are obtained by adopting dual-polarization ratio data and calculating polarization entropy, anisotropy and scattering angle through eigenvalues of three coherent matrix sub-components obtained by feature decomposition of a coherent matrix; the formula for obtaining the polarization characteristic parameters is as follows: where P i is the proportion of the i-th eigenvalue to the total eigenvalues, λ i is the i-th eigenvalue, H is the polarization entropy, A is the anisotropy, α is the scattering angle, i is the index of variable, λ1 is the volume scattering of the target, λ2 is the surface scattering of the target, and λ3 is the multiple scattering component of the target.

2. The SAR image-based land cover information fusion snow depth retrieval method of claim 1, wherein: the track data collection comprises: collecting satellite position data, satellite speed data, track attitude data, time data, track dynamics parameters and track error correction data, and pre-processing the collected track data; the satellite position data comprises longitude, latitude, height, track inclination and ascending node longitude; the satellite speed data comprises speed along the track and speed perpendicular to the track; the track attitude data comprises pitch angle, yaw angle and roll angle; the time data comprises ephemeris time and data collection time; and the track dynamics parameters comprise track semi-major axis, track eccentricity, perigee angle and mean anomaly angle.

3. The SAR image based land cover information fusion snow depth retrieval method of claim 2, wherein: the pre-processing comprises: track correction, thermal noise removal, coherent speckle filtering and decibel processing of the obtained track data; the track correction comprises: error estimation through polynomial fitting of track errors, formation of an error data set through calculation of differences between actual track data and reference data, selection of a polynomial order to fit the error data, determination of polynomial coefficients by using a least square method, i.e. the transpose of a matrix multiplied by the matrix multiplied by a coefficient vector is equal to the transpose of the matrix multiplied by an error data vector, calculation of the value of the polynomial according to a time stamp after the coefficients are determined, and addition of the value as a correction quantity into the original track data for track correction; the thermal noise removal comprises: obtaining a thermal noise vector at any given azimuth time through linear interpolation of two nearest noise vectors, deriving two noise vectors from a detection range, and removing the noise vectors from the detection range through linear interpolation; the coherent speckle filtering comprises: coherent speckle filtering of a SAR remote sensing image through a filter and updating a center pixel value by using local statistical characteristics of the image, and the filter window is 7*7; the decibel processing formula is as follows: ε = 10 log 10 σ o where σ o is the dimensionless backscatter coefficient and ε is the normalized backscatter coefficient.

4. The SAR image-based land cover information fusion snow depth retrieval method of claim 3, wherein: the optimization of the feature data set comprises: pre-processing of geographical feature data to obtain slope and aspect data, pre-processing of meteorological data, feature extraction on the corrected data, and feature optimization by using a Pearson correlation coefficient; the formula for obtaining the slope and aspect data is as follows: wherein Slope is the slope, Aspect is the aspect, Slope x is the slope in the horizontal direction in the two-dimensional image, Slope y is the slope in the vertical direction in the two-dimensional image; the formula for the Pearson correlation coefficient is as follows: where Y i and X i are characteristic parameters, and are average values of the characteristic parameters, n is the number of parameters, i is the variable index, PCC Y-X is the Pearson correlation coefficient; the Pearson correlation coefficient threshold is selected as 0.9, and when the calculated correlation coefficient is greater than or equal to 0.9, the characteristic parameter is judged to be a parameter having correlation with the snow depth and is used as an input characteristic parameter of the snow depth inversion model.

5. The SAR image-based land cover information fusion snow depth retrieval method of claim 4, wherein: The predicted snow depth comprises that a snow depth inversion model is respectively constructed based on different neural network model frameworks, a residual neural network ResNet model is selected for training, the snow depth inversion model is obtained, model training is performed, and the gap between the inversion model result and the measured value, that is, the inversion accuracy, is obtained through testing, the model parameters are adjusted according to the result, the model parameters with the highest inversion accuracy are selected, the snow depth inversion model is constructed based on the parameters, and the time variable is input into a long short-term memory network LSTM model for training to obtain a long time sequence snow depth inversion model, and the snow depth is predicted. The model training comprises defining a model structure, inputting an experimental data set optimized in features, initializing a learning rate, an optimizer and an iteration number, capturing a nonlinear relationship between input feature parameters and measured snow depth, and obtaining the gap between the inversion model result and the measured value, that is, the accuracy of the inversion result, through testing.

6. A system for retrieving snow depth using the fusion of underlying surface information based on SAR images according to any one of claims 1 to 5, characterized in that: The method comprises a data acquisition module, a terrain correction and radiation calibration module, a feature extraction and optimization module, a neural network model construction and training module, and a snow depth prediction module. The data acquisition module is used to collect satellite position, speed, orbit attitude, time data and dynamic parameters, and perform orbit correction, thermal noise removal, speckle filtering and decibel processing to reduce errors and noises in the data. The terrain correction and radiation calibration module is used to eliminate the radiation brightness error caused by terrain undulation by using the SCS+C correction method to ensure the consistency of the reflectivity of different terrains, and convert the backscattering intensity received by the sensor into a dimensionless backscattering coefficient according to the radar equation. The feature extraction and optimization module is used to calculate the polarization entropy, anisotropy and scattering angle by double polarization ratio data and coherent matrix feature decomposition, obtain the polarization information reflecting the snow characteristics, extract the slope and slope direction from the elevation data, and extract the environmental temperature and humidity data, and screen the features by using the Pearson correlation coefficient to retain the features related to the snow depth. The neural network model construction and training module is used to define the structure of the residual neural network ResNet model, input the optimized feature data set, initialize the model parameters, capture the nonlinear relationship between the feature parameters and the snow depth through training, and evaluate the accuracy of the model. The snow depth prediction module is used to input the SAR remote sensing image based on the ResNet model obtained through training, obtain the snow depth through inversion, introduce the time variable into the LSTM model, train the prediction model of the snow depth changing with time, and perform long-term prediction of the snow depth. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the snow depth inversion method based on SAR image fusion of underlying surface information according to any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the snow depth inversion method based on SAR image fusion of underlying surface information according to any one of claims 1 to 5.

Images