CN117687044B - Deep learning method for satellite-borne GNSS-R global sea surface effective wave height estimation - Google Patents
Deep learning method for satellite-borne GNSS-R global sea surface effective wave height estimation Download PDFInfo
- Publication number
- CN117687044B CN117687044B CN202311450231.8A CN202311450231A CN117687044B CN 117687044 B CN117687044 B CN 117687044B CN 202311450231 A CN202311450231 A CN 202311450231A CN 117687044 B CN117687044 B CN 117687044B
- Authority
- CN
- China
- Prior art keywords
- data
- wave height
- gnss
- satellite
- model
- 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
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000013135 deep learning Methods 0.000 title claims abstract description 23
- 238000012549 training Methods 0.000 claims abstract description 43
- 238000012360 testing method Methods 0.000 claims abstract description 31
- 238000013136 deep learning model Methods 0.000 claims abstract description 21
- 238000003908 quality control method Methods 0.000 claims abstract description 11
- 238000001914 filtration Methods 0.000 claims abstract description 10
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims abstract description 9
- 238000011156 evaluation Methods 0.000 claims abstract description 5
- 238000010276 construction Methods 0.000 claims abstract 2
- 238000010801 machine learning Methods 0.000 claims description 14
- 238000013527 convolutional neural network Methods 0.000 claims description 11
- 238000010200 validation analysis Methods 0.000 claims description 9
- 239000000284 extract Substances 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 230000006403 short-term memory Effects 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 230000002457 bidirectional effect Effects 0.000 claims description 3
- 238000012937 correction Methods 0.000 claims description 3
- 230000002123 temporal effect Effects 0.000 claims 1
- 238000012795 verification Methods 0.000 abstract description 13
- 238000013528 artificial neural network Methods 0.000 description 17
- 230000006872 improvement Effects 0.000 description 10
- 238000012706 support-vector machine Methods 0.000 description 9
- 238000004458 analytical method Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000012544 monitoring process Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000007787 long-term memory Effects 0.000 description 3
- 230000015654 memory Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000002779 inactivation Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 241000287127 Passeridae Species 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S19/00—Satellite radio beacon positioning systems; Determining position, velocity or attitude using signals transmitted by such systems
- G01S19/01—Satellite radio beacon positioning systems transmitting time-stamped messages, e.g. GPS [Global Positioning System], GLONASS [Global Orbiting Navigation Satellite System] or GALILEO
- G01S19/13—Receivers
- G01S19/14—Receivers specially adapted for specific applications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S19/00—Satellite radio beacon positioning systems; Determining position, velocity or attitude using signals transmitted by such systems
- G01S19/38—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system
Landscapes
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
本发明公开了一种星载GNSS‑R全球海面有效波高估计的深度学习方法，包括：获取CYGNSS L1级GNSS‑R数据，ERA5风速、风向、水深和SWH数据，IMERG降雨数据，WaveWatch III(WW3)SWH数据和AVISO卫星高度计SWH数据；从星载GNSS‑R DDM数据中提取观测变量参数、计算特征观测值、采用外部数据选择辅助变量和对所有数据集进行时空匹配；数据质量控制和数据过滤以及训练数据集、验证测试集和测试数据集划分；有效波高反演的GloWH‑Net混合深度学习模型的构建和训练；对SWH反演结果进行性能评估和对比；本发明显著提高了有效波高反演精度。
The present invention discloses a deep learning method for estimating global sea surface significant wave height using satellite-borne GNSS-R, comprising: acquiring CYGNSS L1-level GNSS-R data, ERA5 wind speed, wind direction, water depth and SWH data, IMERG rainfall data, WaveWatch III (WW3) SWH data and AVISO satellite altimeter SWH data; extracting observation variable parameters from satellite-borne GNSS-R DDM data, calculating characteristic observation values, selecting auxiliary variables using external data and performing spatiotemporal matching on all data sets; data quality control and data filtering, as well as division of training data sets, verification test sets and test data sets; construction and training of a GloWH-Net hybrid deep learning model for inversion of significant wave height; and performance evaluation and comparison of SWH inversion results. The present invention significantly improves the inversion accuracy of significant wave height.
Description
Technical Field
The invention relates to the technical field of cross research of GNSS reflected signal inversion sea wave parameters and deep learning, in particular to a deep learning method for estimating the effective wave height of a satellite-borne GNSS-R global sea surface by fusing a CNN/BiLSTM/FCN network.
Background
Effective wave height (SWH) is an important element of marine power environment monitoring, and is a key parameter for characterizing sea waves, and is generally used for sea condition monitoring to ensure the safety of offshore navigation. Has important significance for ocean transportation and ocean resource protection. SWH usually adopts buoy station to observe, and the observation precision is higher, but buoy station layout cost is high, and the space coverage of observation station is limited in quantity to be easily influenced by sea complex environmental factor, lead to easily being destroyed. Satellite radar altimeters are one of the main tools to acquire SWH information, but cannot be used for large scale sea state monitoring unless many such satellites are deployed. Compared with the prior remote sensing technology means, the satellite-borne GNSS-R technology has the unique advantages of high space-time resolution, low observation cost, wide coverage, all-weather operation and the like, and can provide a new thought for estimating SWH. GNSS-R technology has proven to be an important tool for inverting the sea surface SWH. One typical SWH inversion method is an interference complex field (Interferometric Complex Field, ICF) that inverts the wave SWH using an effective correlation time function of the interference ICF of the GNSS reflected signal and the direct signal. Another method is the GNSS-R interferogram Technique (INTERFERENCE PATTERN Technique, IPT) which was proposed by humans in 2015, and shore-based test results indicate that the method estimates an effective wave with a high accuracy of up to 6cm. Since CYGNSS is intended to be satellite-based, the receiver is in motion, which increases the difficulty of data processing compared to terrestrial GNSS. IPT patterns require that signals be received from the same point for a period of time, which requires that the receiver be stationary, and hence IPT patterns are not suitable for satellite based GNSS.
The existing research for inverting the effective wave height based on the satellite-borne GNSS-R data is mainly based on an empirical model and a machine learning model, and has proved that the method has feasibility in inverting the effective wave height. Unfortunately, empirical models face great challenges in constructing a multivariate regression model in consideration of a variety of factors, resulting in low robustness and low inversion accuracy of the model, which is difficult to apply practically. Machine learning algorithms such as stepwise linear regression, support Vector Machines (SVMs), artificial Neural Networks (ANNs), sparrow search algorithms-extreme learning machines, and Bagged Trees (BT) have proven advantageous in constructing multivariate regression efficient wave height inversion models and have higher inversion accuracy than empirical models, but neglecting key feature information in DDM images in the model input data limits the accuracy of efficient wave height inversion. Recently, deep learning methods such as Convolutional Neural Networks (CNNs) have also proven to be an effective method for satellite-borne GNSS-R effective wave height inversion. CNN automatically extracts eigenvalues related to sea surface SWH from BRCSDDM and effective scattering areas, which is good at automatically extracting complex spatial features from multiple input images. However, existing deep learning methods ignore time series characteristics. Long-term memory (LSTM) models are intended to process long-term sequences and extract complex time-series features, but LSTM is prone to overfitting, cannot directly use information at future times, and can only be predicted from current and past information. However, the two-way long and short term memory network (BiLSTM) can more fully understand the sequence data in consideration of the past and future context information at the same time, and can better capture the relationship between the contexts so as to improve the accuracy and the robustness of the model. The fully connected network model (FCN) is mainly used for processing CYGNSS observed variable parameters, GNSS-R characteristic observed values and other auxiliary variable information, the FCN can improve the expression capacity of the model by increasing the number of layers and the number of nodes, and the FCN is suitable for processing complex nonlinear relations. It can be seen that the single network model has respective advantages and disadvantages, and up to the present, no related research reports to directly estimate the ocean effective wave height parameters by using the mixed model fused with the single network.
In summary, the invention firstly discloses a deep learning method for estimating the global sea surface effective wave height of a satellite-borne GNSS-R fused with a CNN/BiLSTM/FCN network.
Disclosure of Invention
In order to solve the problems that the existing satellite-borne GNSS-R sea surface wave height inversion empirical model is low in robustness, low in inversion precision and difficult to popularize in practical application, and the precision of a machine learning and deep learning method is limited, the invention provides a satellite-borne GNSS-R global sea surface effective wave height estimation deep learning method fused with a CNN/BiLSTM/FCN network. The convolution characteristic extraction module is used for extracting spatial characteristic information related to the effective wave height of the sea surface from a two-dimensional matrix of BRCSDDM and EFFECTIVE SCATTERING AREA by using a convolution neural network, and the characteristic relation reasoning module is used for carrying out a reasoning process between characteristic relations based on a BiLSTM network. The model may utilize a two-way long and short term memory network (BiLSTM) to extract time-related features and capture contextual information to better understand the input sequence. Thereby remarkably improving the inversion precision of the effective wave height.
In order to achieve the above purpose, the invention adopts the following technical scheme: a deep learning method for estimating the global sea surface effective wave height of a satellite-borne GNSS-R comprises the following steps:
Step 1, obtaining CYGNSS L-level GNSS-R data, ERA5 wind speed, wind direction, water depth and SWH data, IMERG rainfall data, WAVEWATCH III (WW 3) SWH data and AVISO satellite altimeter SWH data;
Step 2, extracting observation variable parameters from satellite-borne GNSS-R DDM data, calculating characteristic observation values, adopting external data to select auxiliary variables and carrying out space-time matching on all data sets;
Step 3, data quality control and data filtering, training data set, verification test set and test data set division;
step 4, constructing and training a GloWH-Net mixed deep learning model for effective wave height inversion;
And 5, inputting the test dataset into the trained GloWH-Net mixed deep learning model to obtain an inversion effective wave height value, and performing performance evaluation and comparison on SWH inversion results of the GloWH-Net mixed deep learning model and a previous model, namely an empirical model and a machine learning model.
As a further improvement of the present invention, the CYGNSS L-level GNSS-R data includes BRCSDDM, effective scattering area (EFFECTIVE SCATTERING AREA), normalized Bistatic Radar Cross Section (NBRCS), leading Edge Slope (LES), signal-to-noise ratio (SNR), power DDM (power-analog), longitude and latitude of specular reflection point (sp_lon, sp_lat), receiver antenna gain (sp_rx_gain), distance between specular reflection point and GNSS emitter (tx_to_sp_range), distance between specular reflection point and receiver (rx_to_sp_range), DDM observation time (DDM _time_ utc), and angle of incidence (sp_inc_angle); ERA5 wind speed, wind direction, water depth and SWH data, IMERG rainfall data, WAVEWATCH III (WW 3) SWH data, AVISO satellite altimeter SWH data.
As a further improvement of the present invention, step 2 specifically comprises the steps of:
Step 2.1, calculating characteristic observations from the on-board GNSS-R DDM data, including Normalized Bistatic Radar Cross Section (NBRCS), leading Edge Slope (LES), trailing Edge Slope (TES), and DDM power average (DDMA). Wherein NBRCS, LES and TES observations are calculated from the BRCS image, and DDMA is calculated from the power DDM (Power_analog);
Step 2.2, selecting observation variable parameters from the satellite-borne GNSS-R data, wherein the observation variable parameters comprise signal-to-noise ratio (SNR), longitude and latitude (sp_lon, sp_lat) of a specular reflection point, receiver antenna gain (sp_rx_gain), distance correction gain (RCG) and incident angle (sp_inc_angle) variables; wherein, RCG is calculated by the following formula:
Where R T and R R are the distance between the specular reflection point and the GNSS transmitter (tx_to_sp_range) and the distance between the specular reflection point and the receiver (rx_to_sp_range), respectively; g R is receiver antenna gain (sp_rx_gain), RCG is in units of 10 27×dBi×m-4;
Step 2.3, CYGNSS GNSS-R L, DDM observation time (DDM _time_ utc) BRCSDDM, effective scattering area (EFFECTIVE SCATTERING AREA), power DDM (power_analog), GNSS-R characteristic observation values, GNSS-R observation variable parameters and other auxiliary variables are unified with ERA5 wind speed, ERA5 wind direction, ERA5 effective wave height, IMERG rainfall data, WW3 and AVISO effective wave height data to obtain a matched data set.
As a further improvement of the present invention, in step 3, the data quality control is represented as the following table:
the data sets after data filtering are divided into training sets, verification sets and test sets according to time sequences.
As a further improvement of the invention, the GloWH-Net mixed deep learning model for the effective wave height inversion in the step 4 comprises three large modules, wherein the first network module is a convolution feature extraction module, the input is BRCS DDM and effective scattering area (EFFECTIVE SCATTERING AREA), and the convolution neural network can be used for effectively extracting the spatial feature information related to the sea surface effective wave height from BRCSDDM and EFFECTIVE SCATTERING AREA; the second network module is a characteristic relation reasoning module, and can utilize a two-way long-short-Term Memory network (Bidirectional Long Short-Term Memory, biLSTM) to extract time related characteristics and capture context information, so that an input sequence can be better understood; the third network module is a fully connected network, and inputs are CYGNSS observed variable parameters, GNSS-R characteristic observed values and other auxiliary variable information.
As a further improvement of the invention, the training process of the GloWH-Net mixed deep learning model for the effective wave height inversion is characterized by comprising the following steps:
Preprocessing all data sets, dividing the data sets into a training set, a verification set and a test set after filtering low-quality data, and then using the training set for training GloWH-Net mixed models;
The preprocessing data is normalized to zero mean and unit variance according to the characteristics, the verification set is used for avoiding overfitting, and an early stop condition with six epochs (epochs) is adopted; the learning rate (LEARNING RATE) is set to 0.0001; random inactivation (dropout) was set to 0.1; the training set, the verification set and the test set are completely independent at the observation time; the training set is used for training a network model, and the verification set is used for supervising model training; batch training size (batch size) is set to 32 and epoch number is set to 100.
The invention discloses a deep learning method for estimating the global sea surface effective wave height of a satellite-borne GNSS-R fused with a CNN/BiLSTM/FCN network, wherein a GloWH-Net deep learning model for inverting the effective wave height consists of three core neural network modules, wherein the first network module is a convolution characteristic extraction module, is input into BRCS DDM and an effective scattering area (EFFECTIVE SCATTERING AREA), and can effectively extract spatial characteristic information related to the sea surface effective wave height from the BRCS DDM and EFFECTIVE SCATTERING AREA by utilizing a convolution neural network; the second network module is a characteristic relation reasoning module, and can utilize a two-way long-short-term memory network (BiLSTM) to extract time related characteristic information and capture context information, so that an input sequence can be better understood; the third network module is a fully connected network, and inputs are CYGNSS observed variable parameters, GNSS-R characteristic observed values and other auxiliary variable information. And finally, outputting the effective wave height inversion value by combining the three core networks and adopting a fully-connected network. The method comprises the following steps: acquiring CYGNSS L-level GNSS-R data, ERA5 wind speed, wind direction, water depth and SWH data, IMERG rainfall data, WAVEWATCH III (WW 3) SWH data and AVISO satellite altimeter SWH data; extracting observation variable parameters, calculation characteristic observation values and other auxiliary parameters from satellite-borne GNSS-R data, and performing space-time matching on all data sets; data quality control and filtering; establishing and training a GloWH-Net mixed deep learning model for effective wave height inversion; and inputting the test dataset into the GloWH-Net model after training to obtain an inversion effective wave height value, and performing performance evaluation and comparison on SWH inversion results of the GloWH-Net model and the previous model (namely an empirical model and a machine learning model).
The beneficial effects of the invention are as follows:
By adopting the technical scheme of the invention, the GloWH-Net model can show the capability of inverting the sea wave height with high precision and high resolution in the global scope, thereby strongly promoting the great potential of the deep learning method in the satellite-borne GNSS-R sea wave height monitoring. In addition, the method provided by the invention opens a window for performing satellite-borne GNSS-R SWH inversion by using a deep learning method.
Drawings
FIG. 1 is a flow chart of a deep learning method for estimating the global sea surface effective wave height of a satellite-borne GNSS-R fused with a CNN/BiLSTM/FCN network in an embodiment of the invention;
FIG. 2 is a block diagram of a GloWH-Net deep learning model for satellite-borne GNSS-R global sea surface effective wave height inversion in an embodiment of the invention;
FIG. 3 is a plot of the speckle densities of the inversion of the effective wave height and the ECMWF ERA5 effective wave height for different models in an embodiment of the present invention;
FIG. 4 is a global distribution diagram of ERA5 effective wave height and different model inversion effective wave height deviations in an embodiment of the present invention;
FIG. 5 is a histogram of the distribution of the variation of ERA5 effective wave height and the inversion of the effective wave height of different models in an embodiment of the invention;
FIG. 6 is a plot of the speckle density for different models inverting the effective wave height and WW3 effective wave height in an embodiment of the invention;
FIG. 7 is a histogram of the deviation distribution of WW3 effective wave height and inversion effective wave height of different models in an embodiment of the present invention;
FIG. 8 is a plot of the speckle density for different models inverting the effective wave height and AVISO effective wave heights in an embodiment of the invention;
FIG. 9 is a histogram of the distribution of deviations of AVISO effective wave heights from the inversion of different models in an embodiment of the present invention;
FIG. 10 is a graph of inversion performance of different models over different SWH intervals in an embodiment of the present invention.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
Examples
Example 1
To verify the effectiveness of the proposed method, CYGNSS GNSS-R observations, ECMWF ERA5 re-analysis data, IMERG rainfall data, WAVEWATCH III effective wave height data and AVISO effective wave height data were obtained for all days of 1 year from 2019.04 to 2020.04 years, and the experimental results of the present invention were compared with the results of inversion of effective wave heights of existing models such as DDMA, LEWS, combined experience model, bag Tree (BT), support Vector Machine (SVM) and Artificial Neural Network (ANN). The basic configuration of the experimental platform for constructing GloWH-Net model is as follows:
The experimental environment is Intel Core i7-10700@2.90GHz eight-Core CPU, NVIDIA GeForce RTX 2080 SUPER GPU,32GB (DDR 4 3000MHz in Intel) running memory, windows 10 operating system, the experimental platform is Matlab 2019b and Python 3.7, and the deep learning API is tensorflow.3.1 and Keras.2.4.3.
The implementation flow of the technical scheme of the deep learning method for estimating the global sea surface effective wave height of the satellite-borne GNSS-R fused with the CNN/BiLSTM/FCN network is shown as a figure 1, and the method comprises the following steps:
Step S1, obtaining CYGNSS L-level GNSS-R data, ERA5 wind speed, wind direction, water depth and SWH data, IMERG rainfall data, WAVEWATCH III (WW 3) SWH data and AVISO satellite altimeter SWH data;
s2, extracting observation variable parameters from satellite-borne GNSS-R DDM data, calculating characteristic observation values, adopting external data to select auxiliary variables and carrying out space-time matching on all data sets;
Step S3, data quality control and data filtering, training data set, verification test set and test data set division;
S4, constructing and training a GloWH-Net mixed deep learning model for effective wave height inversion; the structure diagram of GloWH-Net deep learning model of effective wave height inversion is shown in figure 2;
And S5, inputting the test dataset into the trained GloWH-Net mixed model to obtain an inversion effective wave height value, and performing performance evaluation and comparison on SWH inversion results of the GloWH-Net model and the previous model (namely an empirical model and a machine learning model).
As an implementation manner of this embodiment, in step S1, the CYGNSS satellite-borne GNSS-R L1 observation data includes BRCSDDM, an effective scattering area (EFFECTIVE SCATTERING AREA), a normalized double-based radar scattering cross section (NBRCS), a Leading Edge Slope (LES), a signal-to-noise ratio (SNR), a power DDM (power_analog), a longitude and latitude (sp_lon, sp_lat) of a specular reflection point, a receiver antenna gain (sp_rx_gain), a distance (tx_to_sp_range) between the specular reflection point and a GNSS emitter, a distance (rx_to_sp_range) between the specular reflection point and a receiver, a DDM observation time (DDM _time_ utc), and an incident angle (sp_inc_angle); ERA5 wind speed, wind direction, water depth and SWH data, IMERG rainfall data, WAVEWATCH III (WW 3) SWH data, AVISO satellite altimeter SWH data.
As an implementation of this embodiment, step S2 includes the following sub-steps:
Step S2.1, calculating characteristic observations from the on-board GNSS-R DDM data, including Normalized Bistatic Radar Cross Section (NBRCS), leading Edge Slope (LES), trailing Edge Slope (TES), and DDM power average (DDMA). Where NBRCS, LES and TES observations are calculated from the BRCS image, and DDMA is calculated from the power DDM (Power_analog).
In step S2.2, the observation variable parameters are selected from the satellite-borne GNSS-R data, including signal-to-noise ratio (SNR), longitude and latitude of specular reflection points (sp_lon, sp_lat), receiver antenna gain (sp_rx_gain), distance correction gain (RCG), and incident angle (sp_inc_angle) variables. Wherein, RCG is calculated by the following formula:
Where R T and R R are the distance between the specular reflection point and the GNSS transmitter (tx_to_sp_range) and the distance between the specular reflection point and the receiver (rx_to_sp_range), respectively; g R is the receiver antenna gain (sp_rx_gain), the unit of RCG is 10 27×dBi×m-4.
Step S2.3, CYGNSS GNSS-R L, wherein DDM observation time (DDM _timestamp_ utc) BRCSDDM, effective scattering area (EFFECTIVE SCATTERING AREA), power DDM (power_analog), GNSS-R characteristic observation values, GNSS-R observation variable parameters and other auxiliary variables are unified with ERA5 wind speed, ERA5 wind direction, ERA5 effective wave height, IMERG rainfall data, WW3 and AVISO effective wave height data to obtain a matched data set.
As an implementation manner of this embodiment, the filtering of the matched data set in step S3 is critical to the quality control of data according to the quality identifiers listed in the following table 1.
Table 1 data quality control identifier (quality_ flags _1)
For more information on quality control identity (quality_ flags), please refer to the CYGNSS L1V3.1 user manual and data description on the CYGNSS official website (access: https:// cygnss. End. Umiche. Edu/data-products /). The data-filtered data set in step S3 is divided into a training set, a verification set and a test set according to a time sequence, wherein the data set from 4 months in 2019 to 8 months is used as the training set, the data set from 8 months in 2019 to 10 months is used as the verification set, and the data set from 10 months in 2020 to 4 months is used as the test set.
The training process of the deep learning model for estimating the global sea surface effective wave height of the satellite-borne GNSS-R fused with the CNN/BiLSTM/FCN network in the step S4 specifically comprises the following steps:
All data sets are preprocessed, quality control and filtering are carried out on sampled data, a filtered high-quality data set is obtained, the data sets are divided into a training set, a verification set and a test set according to observation time, and the training set, the verification set and the test set respectively account for 30%, 15% and 55% of the filtered data set.
The pre-processed data is normalized to zero mean and unit variance by feature, and the validation set is used to avoid overfitting and employs an early stop condition with six epochs (epochs). The learning rate (LEARNING RATE) was set to 0.0001. To further prevent the over-fitting problem, the random inactivation (dropout) was set to 0.1;
By adopting Relu functions as the optimal activation functions (shown by formulas) of all the neural network models, the average absolute percentage error between the predicted effective wave height and the reference effective wave height is used as a loss function (shown by formulas)), the invention selects an Adam optimization algorithm to train small random and disordered batches.
Where x is the input value of the neuron of the previous layer. Wherein m is the number of samples,For predicting the effective wave height, u i is the reference effective wave height.
The training set and the validation set are completely independent in observation time. The training set is used to train the network model, and the verification set is used to supervise model training. The batch training size (batch size) in the present invention is set to 32, and the epoch number is set to 100 in order to ensure good training performance.
After training is completed, the test dataset is input into a GloWH-Net effective wave height inversion model after training to obtain inversion effective wave height values, and the results are evaluated by analyzing effective wave heights, WAVEWATCHIII and AVISO effective wave height data respectively by ECMWF ERA 5. Root Mean Square Error (RMSE), bias, correlation Coefficient (CC) and Mean Absolute Percentage Error (MAPE) are used as indicators for evaluating the inversion performance of the model.
To demonstrate the improvement of the invented efficient wave height inversion model based on the deep learning method, the inversion results of the invention are compared with the inversion results of two existing models, namely, an empirical model (comprising DDMA, LEWS and a combined model), a machine learning (comprising a Support Vector Machine (SVM), a bag-in-tree (BT) and an Artificial Neural Network (ANN) model). In the empirical model method, ERA5 wind speed, wind direction, water depth, and IMERG rainfall data were not used to invert the effective wave height. In the machine learning method, the input parameters in the ANN model include only GNSS-R observations, CYGNSS observation variables, and other assistance parameters. However, the GloWH-Net hybrid deep learning model invented fuses BRCS DDM, effective scattering area (EFFECTIVE SCATTERING AREA), CYGNSS observed variable parameters, GNSS-R characteristic observations, and other auxiliary variable information. The information of the input parameters of the different models is shown in table 2. The models are trained directly from the training set, the performance of the different model training is supervised using the validation set, and the generalization ability of the methods is tested using a test set that is completely independent of the training data set. The reference SWH data in the test set includes ERA5, WW3, and AVISO SWH data, where WW3 and AVISO SWH data are not used for model training, so both data sets can better evaluate the robustness and versatility of the inventive model.
TABLE 2 input parameter information for different models
Table 3 gives the inversion accuracy statistics for different models for inversion of effective wave height versus ERA5 data on the test dataset. From the table it can be derived that:
(1) The inversion accuracy of the empirical model method is poor, compared with the ERA5 re-analysis effective wave height, the empirical model method is remarkably improved in the aspects of RMSE, CC and MAPE, and is respectively improved by 53.87%, 42.09% and 60.91% compared with the DDMA observation value model. Compared with LEWS observed value model, 53.47%, 40.80% and 60.40% are respectively improved. Compared with the combined model, the method improves by 53.45%, 40.70% and 60.50% respectively. Furthermore, 41.54%, 19.32% and 44.60% were improved, respectively, compared to the SVM model.
(2) The GloWH-Net model of the invention is improved by 21.92%, 6.05% and 19.51% in terms of RMSE, CC and MAPE, respectively, compared with the BT model.
Table 3 different models invert the accuracy of the effective wave height versus ERA5 data on the test dataset
In order to compare the outstanding improvement of the inversion effective wave height of the model provided by the invention in the correlation with ERA5 effective wave height products compared with other inversion methods, a scatter density chart of inversion effective wave height and ERA5 effective wave height data of different models is shown in figure 3. The red dashed line in the figure represents a 1:1 reference line, the magenta solid line represents a linear fitting result (fitting equation (y=ax+b) is shown in the figure) of inverting the effective wave height to the ERA5 effective wave height, and CC represents a correlation coefficient between the model inverting the effective wave height and the ERA5 effective wave height. As can be seen from the figures:
(1) The correlation between the inversion result of the GloWH-Net effective wave height model and the ERA5 effective wave height is good and is superior to an empirical model and a machine learning method. Wherein both the empirical model and the machine learning model are presented with significant underestimation at significant wave height values greater than 3 m. The inversion performance of the ANN model and the GloWH-Net model of the invention is superior to that of other models under the condition of large effective wave height, and the inversion effective wave height values are relatively gathered along the 1:1 reference line distribution.
(2) DDMA, LEWS, combined model and SVM models exhibit deviations from 1:1 above the reference line for effective wave heights in the range of 0-4 m. However, the BT, ANN and GloWH-Net models do not have this problem significantly. Wherein the GloWH-Net model of the invention performs optimally in inversion, which shows that after adding the DDM image to the input layer of the GloWH-Net model, the architecture comprising the convolutional layer has better performance than the architecture with only fully connected layers (such as ANN). In addition, the network BiLSTM is added to better capture the contextual information, so that the effective wave height inversion performance is remarkably improved.
(3) Compared with the existing experience model for satellite-borne GNSS-R effective wave height inversion and the traditional neural network model, the GloWH-Net model architecture has better inversion performance under the condition of high effective wave height, namely the GloWH-Net model obviously improves the underestimation phenomenon under the sea condition of large effective wave height.
To further evaluate the global performance of the GloWH-Net model in inverting SWH, fig. 4 shows the global distribution of ERA5 effective wave heights from 2 months 2020 to 3 months 2020 with the GloWH-Net model inversion effective wave heights. Fig. 5 shows a GloWH-Net model and a distribution histogram of the effective wave height deviation of the inversion of the existing model from ERA5 (average deviation (μ), standard deviation (σ), average absolute error (MAE) and 80% quantile of deviation (Qua) are shown in the figure, blue bar represents the error distribution, red solid represents the probability density function fitted curve of the error, and green dashed represents the deviation of 0 m). It can be seen from fig. 4 and 5 that GloWH-Net performs better than other models in inverting the global effective wave height. The significant wave height inversion results of the GloWH-Net model are very concentrated (80% significant wave height deviation less than 0.38 m) and approach the 0m bias line with respect to ERA5 data, whereas the BT model 80% significant wave height deviation is less than 0.48m and exhibits a greater SWH bias between 20 ° -40 ° south latitude worldwide than the GloWH-Net model. From the effective wave height deviation histograms of the seven models, the global effective wave height result obtained by inversion of LEWS method is worst, and the 80% effective wave height deviation is 0.98m. The analysis further shows that the GloWH-Net model has better performance when being used for global ocean effective wave height inversion.
Table 4 gives the accuracy statistics by comparing the effective wave height results of the different model inversions with WW3 data. As can be seen from the table, the RMSE of the GloWH-Net model of the invention is minimal (0.393 m), followed by the BT model and the ANN model, with RMSE of 0.483m and 0.502m, respectively. The comparison of the inverted effective wave height with the WW3 effective wave height is similar to ERA 5. In addition, the performance of the GloWH-Net model of the invention is significantly better than that of the model based on the MVE method combined with the DDMA and LEWS results, with respect to four indexes (RMSE, bias, CC and MAPE), the RMSE, CC and MAPE are respectively improved by 48.06%, 41.29% and 56.29%; compared with a DDMA observation model, 48.15%, 41.29% and 55.38% are respectively improved; 48.16%, 41.30% and 57.27% improvement compared to LEWS% observation model, respectively; the BT model is improved by 18.72%, 6.35% and 14.20% respectively. This shows that optimizing the network structure and fusing DDM and other auxiliary parametric modeling is very important for improving the performance of the effective wave height model.
Table 4 different models invert the accuracy of the effective wave height versus WW3 data on the test dataset
FIG. 6 shows a plot of the speckle densities for different models inverting the effective wave height and WW3 effective wave height. It can be seen that the comparison of the effective wave height and WW3 for the different model inversion is consistent with ERA5 data. The GloWH-Net model is reliable, has higher generalization capability and has good practical application value. Fig. 7 also shows the distribution histograms of the deviations between the effective wave height inversion results of the 7 models and the WW3 effective wave height. It can also be seen from FIG. 7 that the GloWH-Net model still outperforms the other 6 existing models. The result of the GloWH-Net model effective wave height inversion is very concentrated (80% of effective wave height deviation is less than 0.43 m) with the deviation of WW3 data and concentrated by 0m of deviation line, while the 80% of effective wave height deviation of the empirical model is less than 0.98m. Although the global effective wave high performance of the machine learning method inversion is better than that of the empirical model, the method is obviously worse than that of the GloWH-Net deep learning model. The analysis further proves that the GloWH-Net model has better generalization capability in inverting the global ocean effective wave height.
To test the performance of 7 model inversion SWH results in comparison to AVISO data, table 5 gives statistics of the accuracy of the comparison of SWH results from different model inversion to AVISO data. As can be seen from the table, the RMSE of the GloWH-Net model of the invention is minimal (0.433 m), followed by the BT model and the ANN model, and is 0.453m and 0.469m, respectively. The comparison result of inversion effective wave height and AVISO effective wave height is similar to ERA5 and WW3 data comparison result. Furthermore, compared to the BT model, there were 4.47%, 1.93% and 8.19% increases in RMSE, CC and MAPE, respectively; compared with the combined model, 40.63%, 36.77% and 55.32% are respectively improved; 21.34%, 16.49% and 30.29% improvement compared to the SVM model, respectively; compared with the ANN model, the method improves by 7.63%, 4.40% and 12.69% respectively. The above analysis also shows that the GloWH-Net deep learning model of the invention has higher SWH inversion accuracy than the empirical model and the machine learning model.
Table 5 different models invert the accuracy of the effective wave height versus AVISO data on the test dataset
FIG. 8 shows a plot of the speckle density for different models inverting the effective wave height and AVISO effective wave height. It can be seen that the effective wave heights for the different model inversions have a high correlation with AVISO data. Fig. 9 also shows a histogram of the distribution of deviations between the inversion result of the effective wave heights of the 7 models and AVISO effective wave heights. It can also be seen from FIG. 9 that the GloWH-Net model still outperforms the other 6 existing models. The GloWH-Net model effective wave height inversion results are very concentrated (80% of effective wave height deviation is less than 0.48 m) with AVISO data and concentrated by 0m of deviation line, while the empirical model 80% of effective wave height deviation is less than 1.05m. The use of AVISO effective wave heights as reference values also further proves that the accuracy of the GloWH-Net deep learning model inversion SWH is obviously superior to that of an empirical model and a machine learning model.
To verify the inversion performance of different models under different SWH sea conditions, FIG. 10 shows the statistics of the RMSE and Bias for the different models estimating SWH over different SWH intervals. Compared with other six existing models, the GloWH-Net model disclosed by the invention has excellent inversion performance, and is particularly good in improvement of inversion performance under the condition of high SWH.
The foregoing examples merely illustrate specific embodiments of the invention, which are described in greater detail and are not to be construed as limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention.
Claims (5)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311450231.8A CN117687044B (en) | 2023-11-02 | 2023-11-02 | Deep learning method for satellite-borne GNSS-R global sea surface effective wave height estimation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311450231.8A CN117687044B (en) | 2023-11-02 | 2023-11-02 | Deep learning method for satellite-borne GNSS-R global sea surface effective wave height estimation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117687044A CN117687044A (en) | 2024-03-12 |
| CN117687044B true CN117687044B (en) | 2024-08-30 |
Family
ID=90125256
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311450231.8A Active CN117687044B (en) | 2023-11-02 | 2023-11-02 | Deep learning method for satellite-borne GNSS-R global sea surface effective wave height estimation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117687044B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119578201B (en) * | 2024-09-19 | 2025-07-15 | 崂山国家实验室 | Method and system for inverting SLA (service level agreement) based on airborne altimeter along-track data |
| CN119355787B (en) * | 2024-10-19 | 2025-06-10 | 昆明理工大学 | Global lake wind speed estimation method for Chinese Tianmu No. one satellite-borne BDS/GPS/GLONASS/Galileo reflection signals |
| CN119474739B (en) * | 2024-11-05 | 2025-05-02 | 昆明理工大学 | A hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion |
Family Cites Families (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP4089660A4 (en) * | 2020-01-09 | 2023-10-11 | Seadronix Corp. | METHOD AND DEVICE FOR PORT AND SHIP MONITORING IN CONSIDERATION OF SEA LEVEL |
| KR20210150776A (en) * | 2020-06-04 | 2021-12-13 | 대우조선해양 주식회사 | System for predicting wave information through machine learning of sea level image |
| CN111950438B (en) * | 2020-08-10 | 2021-06-01 | 中国人民解放军国防科技大学 | Depth learning-based effective wave height inversion method for Tiangong No. two imaging altimeter |
| CN112014842B (en) * | 2020-09-16 | 2021-03-19 | 国家海洋技术中心 | Effective wave height estimation method and system for high-resolution three-signal full-polarization SAR data |
| CN112734723B (en) * | 2021-01-08 | 2023-06-30 | 温州医科大学 | Multi-source data-oriented breast tumor image classification prediction method and device |
| CN113156439B (en) * | 2021-05-08 | 2022-08-30 | 中国石油大学(华东) | SAR wind field and sea wave joint inversion method and system based on data driving |
| CN113688773B (en) * | 2021-09-03 | 2023-09-26 | 重庆大学 | A method and device for repairing tank dome displacement data based on deep learning |
| CN114861537A (en) * | 2022-04-29 | 2022-08-05 | 武汉大学 | GNSS-R sea surface wind speed inversion method and system based on CNN multi-information fusion |
| CN115660855A (en) * | 2022-11-09 | 2023-01-31 | 山东大学 | A Stock Closing Price Prediction Method Fused with News Data |
| CN116148863A (en) * | 2022-12-20 | 2023-05-23 | 昆明理工大学 | Multi-mode deep learning method for satellite-borne GNSS-R sea surface wind speed inversion based on CNN |
| CN116105697B (en) * | 2023-01-17 | 2025-04-11 | 昆明理工大学 | A deep learning-based method for inverting global sea surface significant wave height using spaceborne GNSS-R |
| CN116794652A (en) * | 2023-06-11 | 2023-09-22 | 昆明理工大学 | Satellite-borne GNSS-R global ocean surge height inversion method based on CNN-ConvLSTM model |
-
2023
- 2023-11-02 CN CN202311450231.8A patent/CN117687044B/en active Active
Non-Patent Citations (1)
| Title |
|---|
| 融合星载GNSS-R数据和多变量参数全球海洋有效波高深度学习反演法;布金伟 等;测绘学报;20240513;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117687044A (en) | 2024-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN117687044B (en) | Deep learning method for satellite-borne GNSS-R global sea surface effective wave height estimation | |
| CN113936142B (en) | Precipitation proximity forecasting method and device based on deep learning | |
| CN109886217B (en) | A method for detecting wave heights from videos of near-shore ocean waves based on convolutional neural networks | |
| CN113156439B (en) | SAR wind field and sea wave joint inversion method and system based on data driving | |
| CN116105697B (en) | A deep learning-based method for inverting global sea surface significant wave height using spaceborne GNSS-R | |
| CN116953707B (en) | A tidal level monitoring radar device and monitoring method and system | |
| CN116148863A (en) | Multi-mode deep learning method for satellite-borne GNSS-R sea surface wind speed inversion based on CNN | |
| CN113486705A (en) | Flood monitoring information extraction method, device, equipment and storage medium | |
| CN113960625B (en) | Water depth inversion method based on satellite-borne single-photon laser active and passive remote sensing fusion | |
| CN117710508A (en) | Near-surface temperature inversion method and device based on improved conditional generative adversarial network | |
| CN119474739B (en) | A hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion | |
| CN116794652A (en) | Satellite-borne GNSS-R global ocean surge height inversion method based on CNN-ConvLSTM model | |
| Zhou et al. | An interpretable attention-based deep learning method for landslide prediction based on multi-temporal InSAR time series: A case study of Xinpu landslide in the TGRA | |
| Hu et al. | Deep learning-based homogeneous pixel selection for multitemporal SAR interferometry | |
| Fu et al. | CKMDiff: A generative diffusion model for CKM construction via inverse problems with learned priors | |
| CN118311576B (en) | Dual-polarized SAR shallow sea water depth inversion construction method based on physical constraint neural network | |
| CN117969881B (en) | Method, device and medium for inverting typhoon sea conditions and sea surface wind speed based on remote sensing data | |
| KR102726421B1 (en) | Method and system for measuring water depth | |
| CN114924265A (en) | Lake water level determination method and device, storage medium and electronic equipment | |
| Li et al. | Self-attention Guided Multi-indicator Retrieval for Ocean Surface Wind Field with Multi-modal Data Augmentation and Fusion | |
| Chen et al. | Improving NLOS/LOS Classification Accuracy in Urban Canyon Based on Channel-Independent Patch Transformer with Temporal Information | |
| CN119620243B (en) | A method and system for forecasting short-term rainfall and floods based on deep learning | |
| Champenois et al. | Real-time reconstruction of 3D ocean temperature fields from reanalysis data and satellite and buoy surface measurements | |
| Wang | Improving Radar Quantitative Precipitation Estimation Through Optimizing Radar Scan Strategy and Deep Learning | |
| CN118865147B (en) | ConvLSTM-based differential GNSS scene identification method and system |
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 |