Satellite-borne GNSS-R soil humidity inversion method and system

Abstract

The invention discloses a satellite-borne GNSS-R soil humidity inversion method and system, and relates to the technical field of soil humidity data monitoring, and the satellite-borne GNSS-R soil humidity inversion method specifically comprises the following steps: obtaining a data set; preprocessing the data set to obtain a preprocessed data set; dividing the preprocessed data set to obtain a training set and a test set; constructing a soil humidity inversion model through the training set and the test set; inputting a to-be-tested sample into the finally trained soil humidity inversion model to obtain a soil moisture value; according to the satellite-borne GNSS-R soil humidity inversion method, the two-dimensional data DDM graph and the one-dimensional data characteristic parameters in the input data set are respectively input into the CNN module and the Transform module, so that multi-dimensional learning is realized, and the problem of data loss caused in the process of extracting data in the DDM graph by other models is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of soil moisture data monitoring, and in particular relates to a satellite-borne GNSS-R soil moisture inversion method and system. Background Art

[0002] Soil moisture is an important component of the earth's freshwater resources. Although it only accounts for 0.05%, it plays an important role in climate change, natural disaster prevention and agricultural management. With the advancement of science and technology, especially the development of remote sensing and sensor technology, the ability to monitor soil moisture has been significantly improved. The Global Navigation Satellite System Reflectometer (GNSS-R) uses a dual-base measurement method to efficiently invert soil moisture, with the advantages of high temporal and spatial resolution, all-weather and low cost.

[0003] GNSS-R receives direct signals from navigation satellites and reflected signals from the land surface, and uses physical models to convert signal characteristics into soil moisture information. At the same time, the rapid development of deep learning has brought new opportunities for meteorological remote sensing, improving the monitoring and prediction capabilities of key meteorological parameters such as soil moisture and sea surface rainfall, showing great application potential in extreme weather event prediction and climate research.

[0004] At present, the method of GNSS-R soil moisture inversion based on machine learning methods mainly combines the L1 product parameters of GNSS-R, such as sp_lat, sp_lon, ddm_nbrcs, ddm_snr, delay Doppler map and other parameters related to soil moisture with the soil moisture data observed by the SMAP satellite. By training the model, the mapping relationship from remote sensing data to soil moisture is realized, thereby achieving the purpose of soil moisture inversion.

[0005] However, most current studies preprocess the Delay Doppler Map (DDM) input to the model and extract the information in the DDM. During the information extraction process, some important data in the DDM may be lost, which may have a significant impact on the inversion results.

[0006] This paper combines the advantages of CNN and Transformer models and proposes a new machine learning model, CNN-Transformer model, to invert soil moisture. Summary of the invention

[0007] The purpose of the present invention is to design an improved GNSS-R soil moisture inversion model based on machine learning methods, and to design and build a soil moisture inversion model based on machine learning methods. The model combines the advantages of CNN and Transformer models. The advantages of the CNN model are: it can process two-dimensional or multi-dimensional data, such as image information, video and audio information, etc. By properly adjusting the size and structure of the convolution kernel, CNN can effectively process input data in different forms, without extracting data information in the DDM diagram, and preventing the loss of important data caused in the process of extracting information; the advantages of the Transformer model are: the Transformer architecture allows all positions of the input sequence to be processed simultaneously, has a multi-head attention mechanism, allows the model to learn different representations in parallel in different representation subspaces, increases the model's ability to express different points of interest, and improves the model's expression ability and generalization ability. The soil moisture inversion through the model has good inversion performance, and provides a new inversion method and idea for GNSS-R soil moisture inversion based on machine learning methods.

[0008] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0009] The present invention provides a method for inverting soil moisture using a satellite-borne GNSS-R, comprising the following steps:

[0010] Acquire datasets, including the CYGNSS L1 soil moisture dataset and the SMAP-assisted soil moisture dataset;

[0011] Preprocessing the data set to obtain a preprocessed data set;

[0012] Dividing the preprocessed data set into a training set and a test set;

[0013] Building a soil moisture inversion model through the training set and the test set;

[0014] This includes inputting the one-dimensional data set in the training set into the Transformer module, inputting the two-dimensional data set into the CNN module for multi-dimensional learning, and training the soil moisture inversion model;

[0015] The test set is input into the trained soil moisture inversion model to generate soil moisture inversion results, and the model is evaluated and analyzed;

[0016] The samples to be tested are input into the finally trained soil moisture inversion model to obtain the soil moisture value.

[0017] As a further optimization scheme of the present invention, a data set is obtained: including DDM (delayed Doppler map), latitude and longitude information, radar cross section and front slope; the SMAP soil moisture auxiliary data includes surface roughness, vegetation water content, and vegetation optical thickness.

[0018] As a further optimization scheme of the present invention, data preprocessing: including quality control, spatiotemporal matching, data normalization and outlier processing;

[0019] The quality control includes obtaining screening data based on the quality flag provided by the CYGNSS L1 level data;

[0020] The time-space matching includes time-space matching of CYGNSS L1 level data and SMAP soil moisture data using a bilinear interpolation method;

[0021] The outlier processing includes removing data whose observation values ​​are NaN and negative numbers in the CYGNSS L1 level observation data.

[0022] As a further optimization scheme of the present invention, the soil moisture inversion model includes an input layer, a hidden layer, a fully connected layer and an output layer which are connected in sequence.

[0023] As a further optimization scheme of the present invention, the training data set is fed into the neural network in batches, and calculations are performed layer by layer starting from the input layer until the prediction result of the output layer is obtained;

[0024] The loss function is used to calculate the difference between the predicted value and the actual label;

[0025] Use the loss value to calculate the gradient, back-propagate from the output layer to the input layer, and obtain the back-propagation data;

[0026] Based on the back-propagation data, the gradient update rule is used to adjust the weights and biases of each layer in the model to obtain a trained soil moisture inversion model.

[0027] As a further optimization scheme of the present invention, the test set is input into the model to perform soil moisture inversion to obtain an inversion result, and the inversion result is compared with the actual soil moisture data to evaluate the performance of the model.

[0028] As a further optimization scheme of the present invention, the loss function is the root mean square error (RMSE), and the calculation formula is:

[0029]

[0030] Among them, i is the number of measurements, pi is the measured value, and ai is the actual value.

[0031] As a further optimization scheme of the present invention, the activation function of the soil moisture inversion model is the Tanh function, and the calculation formula is:

[0032]

[0033] Another aspect of the present invention further provides a spaceborne GNSS-R soil moisture inversion system, comprising:

[0034] The data acquisition module is used for data collection and feature extraction, collecting observation data of the CYGNSS L1 frequency band within a specified time range;

[0035] The data preprocessing module is used for data quality control. It performs quality control based on the quality flag provided by the CYGNSS L1 data, processes invalid observations and outliers, and removes Nan values ​​and negative values ​​of observations.

[0036] The data input module inputs the two-dimensional DDM graph and one-dimensional data into the CNN module and the Transformer module for multi-dimensional learning according to the characteristic parameters of the input CYGNSS L1 data;

[0037] In the soil moisture inversion module, after the training set is input into the CNN module and the Transformer module respectively, the two modules perform model training simultaneously; an independent test set is input into the model for soil moisture inversion and the output inversion results are compared with the actual soil moisture data to evaluate the performance of the model.

[0038] The present invention has at least the following beneficial effects: the present invention divides the data set into a training set and a test set, collects CYGNSS L1 data and SMAP soil moisture auxiliary data of the same time period, and utilizes the advantages of the CNN model to effectively learn and recognize information in the image. The Transformer model gets rid of the dependence on the sequential processing of sequence data, and inputs the two-dimensional data DDM map and the one-dimensional data feature parameters in the input data set into the CNN module and the Transformer module respectively, thereby realizing multi-dimensional learning and solving the data loss problem caused by other models in the process of extracting data in the DDM map. By introducing the attention mechanism, when processing each word in the sequence, it can simultaneously pay attention to and integrate the information of all other words in the entire sequence. This ability greatly enhances the model's ability to capture global dependencies. Therefore, when processing long sequence data, the Transformer model shows higher efficiency and faster speed than traditional recurrent neural networks (RNNs). BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1It is a flow chart of the steps of the soil moisture inversion method based on the CNN-Transformer model of the present invention;

[0040] Figure 2 It is a CNN-Transformer model architecture diagram of the present invention;

[0041] Figure 3 It is a flow chart of soil moisture inversion based on CNN-Transformer model of the present invention. DETAILED DESCRIPTION

[0042] The present application is further described in detail below in conjunction with the accompanying drawings. It is necessary to point out here that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technical personnel in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0043] like Figure 1 As shown in the figure, a space-borne GNSS-R soil moisture inversion method based on the CNN-Transformer model is shown in the figure. The specific steps are as follows:

[0044] Acquire data sets, including CYGNSS L1 soil moisture data set and SMAP-assisted soil moisture auxiliary data set; extract characteristic parameters such as DDM (delayed Doppler map), latitude and longitude information, radar cross section and front slope from CYGNSS L1 data, and at the same time, extract characteristic parameters such as surface roughness, vegetation moisture content, vegetation optical thickness from SMAP data as auxiliary characteristic parameters. These characteristic parameter information will be used as input data for soil moisture inversion model.

[0045] Preprocessing the data set to obtain a preprocessed data set; this step covers multiple processing steps for CYGNSS L1 level data, including quality control, spatiotemporal matching, data normalization, and outlier processing;

[0046] Since the CYGNSS satellite was originally designed for the observation of ocean surface information, the ocean surface information needs to be eliminated when performing soil moisture inversion. Data quality control mainly relies on the quality flag provided by the CYGNSS L1 data to screen the data, which includes eliminating sampling points where abnormalities occur during DDMI sensor detection, such as satellite attitude offset, GPS transmit power estimation errors, GPS satellite and velocity signal errors, etc. At the same time, observation data with the following characteristics are eliminated: large spacecraft attitude errors or anomalies, high uncertainty in transmitter power or antenna gain and incident angle, invalid or abnormal DDM, and instrument data transmission or calibration problems. In addition, data with NaN or negative observation values ​​in CYGNSS L1 observation data must be excluded and the data must be normalized.

[0047] The preprocessed data set is divided into a training set and a test set; the training set is used to train the model, and the test set is used to test the model and make predictions.

[0048] Building a soil moisture inversion model through the training set and the test set;

[0049] Model building: The CNN-Transformer model combines the CNN model and the Transformer model, integrates the machine learning advantages of the two, builds the CNN-Transformer model, and realizes multi-dimensional learning;

[0050] The three elements of model training mainly include: model, algorithm, and strategy. The model is the construction model, the algorithm is the learning optimizer, and the strategy is the loss function. For the present invention, the input of the soil moisture inversion model proposed by the present invention is the characteristic parameters of the CYGNSSL1 level product, and the output is the soil moisture value. The soil moisture inversion model includes an input layer, a hidden layer, a fully connected layer, and an output layer connected in sequence, wherein the input layer accepts the characteristic parameters from CYGNSS and SMAP auxiliary data, and the number of its nodes is determined by the number of characteristic parameters in the input data set; the fully connected layer connects the hidden layer and the output layer of the model, and uses the Tanh function as the activation function of the model. The calculation formula is:

[0051]

[0052] The Adam optimizer is used to optimize the model, and the root mean square error (RMSE) is selected as the loss function in the training process. The calculation formula is:

[0053]

[0054] Among them, i is the number of measurements, pi is the measured value, and ai is the actual value, which is used to measure the difference between the predicted value and the actual soil moisture value and evaluate the model performance;

[0055] The training phase of the model has four processes, namely forward propagation, loss calculation, back propagation and weight update. The training data set is sent to the neural network in batches, starting from the input layer and calculating layer by layer until the prediction result of the output layer is obtained;

[0056] Specifically, it includes inputting one-dimensional data sets such as sp_lat, sp_lon, ddm_nbrcs, ddm_les and other data in the training set into the Transformer module, and inputting the two-dimensional data set DDM map into the CNN module to fully utilize the advantages of the CNN model and the Transformer model to achieve multi-dimensional learning and train the soil moisture inversion model.

[0057] The test set is input into the trained soil moisture inversion model to generate soil moisture inversion results, and the model is evaluated and analyzed by comparing the inverted soil moisture with the true soil moisture.

[0058] The trained model is then used to perform analysis and evaluation on the test samples, ultimately achieving prediction of soil moisture values ​​for new samples.

[0059] The present invention provides a space-borne GNSS-R soil moisture inversion system based on a CNN-Transformer model, comprising:

[0060] The data acquisition module is used for data collection and feature extraction, collecting observation data of the CYGNSS L1 frequency band within a specified time range;

[0061] The data preprocessing module is used for data quality control. It performs quality control based on the quality flag provided by the CYGNSS L1 data, processes invalid observations and outliers, and removes Nan values ​​and negative values ​​of observations.

[0062] The data input module inputs the two-dimensional DDM graph and one-dimensional data into the CNN module and the Transformer module for multi-dimensional learning according to the characteristic parameters of the input CYGNSS L1 data;

[0063] In the soil moisture inversion module, after the training set is input into the CNN module and the Transformer module respectively, the two modules perform model training simultaneously; an independent test set is input into the model for soil moisture inversion and the output inversion results are compared with the actual soil moisture data to evaluate the performance of the model.

[0064] By dividing the data set into a training set and a test set, collecting CYGNSS L1 data, SMAP soil moisture data, and SMAP auxiliary data from the same time period, and taking advantage of the CNN model's ability to effectively learn and recognize information in images, the two-dimensional data DDM graph in the CYGNSS L1 data is preprocessed and input into the CNN module for convolution. The significant advantage of the Transformer model is that it innovatively breaks away from the reliance on sequential data processing and instead adopts a new architecture. By introducing the attention mechanism, the model can simultaneously focus on and integrate the information of all other words in the entire sequence when processing each word in the sequence. This ability greatly enhances the model's ability to capture global dependencies. Therefore, when processing long sequence data, the Transformer model shows higher efficiency and faster speed than traditional recurrent neural networks (RNNs). Taking advantage of this advantage of the Transformer model, the one-dimensional feature parameters of CYGNSS L1 data and SMAP auxiliary data, such as sp_lat, sp_lon, ddm_nbrcs, ddm_les, etc., are preprocessed and input into the Transformer module to train the CNN-Transformer model.

[0065] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A spaceborne GNSS-R soil moisture inversion method, characterized in that: The following steps are involved: Acquire a data set, and preprocess the data set; Perform feature extraction based on the preprocessed data to obtain a one-dimensional data set and a two-dimensional data set; Divide the data set based on the one-dimensional data set and the two-dimensional data set to obtain a training set and a test set; Building a model and optimizing the model using the training set and the test set; Combine the CNN model and the Transformer model to build a CNN-Transformer model; The one-dimensional data set in the training set is input into the Transformer module, and the two-dimensional data set is input into the CNN module for multi-dimensional learning to train the soil moisture inversion model; The test set is input into the trained soil moisture inversion model to generate soil moisture inversion results, and the model is evaluated; based on the evaluation results, it is determined whether the model should be retrained until the optimal model is obtained.

2. The method for inverting soil moisture using satellite-based GNSS-R according to claim 1, characterized in that: Acquired data sets: including DDM (delayed Doppler map), latitude and longitude information, radar cross section and front slope; the SMAP soil moisture auxiliary data includes surface roughness, vegetation water content, and vegetation optical thickness.

3. The spaceborne GNSS-R soil moisture inversion method according to claim 1, characterized in that: Data preprocessing: including quality control, spatiotemporal matching, data normalization and outlier processing; The quality control includes obtaining screening data based on the quality flag provided by the CYGNSS L1 level data; The time-space matching includes time-space matching of CYGNSS L1 level data and SMAP soil moisture data using a bilinear interpolation method; The outlier processing includes removing data whose observation values ​​are NaN and negative numbers in the CYGNSS L1 level observation data.

4. The method for inverting soil moisture using satellite-based GNSS-R according to claim 1, characterized in that: The soil moisture inversion model includes an input layer, a hidden layer, a fully connected layer and an output layer which are connected in sequence.

5. The method for inverting soil moisture using satellite-based GNSS-R according to claim 4, characterized in that: The training data set is fed into the neural network in batches, and calculations are performed layer by layer starting from the input layer until the prediction results of the output layer are obtained; The loss function is used to calculate the difference between the predicted value and the actual label; Use the loss value to calculate the gradient, back-propagate from the output layer to the input layer, and obtain the back-propagation data; Based on the back-propagation data, the gradient update rule is used to adjust the weights and biases of each layer in the model to obtain a trained soil moisture inversion model.

6. The method for inverting soil moisture using satellite-based GNSS-R according to claim 5, characterized in that: The test set is input into the model to perform soil moisture inversion to obtain an inversion result, the inversion result is compared with the actual soil moisture data, the performance of the model is evaluated, and based on the evaluation result, it is determined whether to retrain the model until the optimal model is obtained.

7. The method for inverting soil moisture using satellite-based GNSS-R according to claim 5, characterized in that: The loss function is the root mean square error (RMSE), and the calculation formula is: Among them, i is the number of measurements, pi is the measured value, and ai is the actual value.

8. The method for inverting soil moisture using satellite-based GNSS-R according to claim 5, characterized in that: The activation function of the soil moisture inversion model is the Tanh function, and the calculation formula is:

9. The method for inverting soil moisture using satellite-based GNSS-R according to claim 8, characterized in that: The sample to be tested is input into the trained soil moisture inversion model to obtain the soil moisture value.

10. A satellite-borne GNSS-R soil moisture inversion system, characterized in that: include: The data acquisition module is used for data collection and feature extraction, collecting observation data of the CYGNSS L1 frequency band within a specified time range; The data preprocessing module is used for data quality control. It performs quality control based on the quality flag provided by the CYGNSS L1 data, processes invalid observations and outliers, and removes Nan values ​​and negative values ​​of observations. The data input module inputs the two-dimensional DDM graph and one-dimensional data into the CNN module and the Transformer module for multi-dimensional learning according to the characteristic parameters of the input CYGNSS L1 data; In the soil moisture inversion module, after the training set is input into the CNN module and the Transformer module respectively, the two modules perform model training simultaneously; an independent test set is input into the model for soil moisture inversion and the output inversion results are compared with the actual soil moisture data to evaluate the performance of the model.

Images