Method for generating a learning model, computer program, microwave radiometer, and estimation method

Abstract

A learning model generation method which enables efficiently estimating the amount of precipitable water.　This learning model generation method involves: acquiring microwave measurement data, which is data obtained by measuring emitted from the atmosphere with a microwave radiometer microwaves, and training data, which includes the amount of precipitable water obtained on the basis of the GNSS signal received by the GNSS receiver provided in the microwave radiometer; generating, on the basis of the acquired training data, a learning model that, when microwave radiometry data is inputted, outputs the amount of precipitable water; further continuing the acquisition of microwave measurement data, which is measured with the microwave radiometer, and training data, which includes the amount of precipitable water obtained on the basis of GNSS signals received by the aforementioned GNSS receiver; and repeating re-training of the learning model on the basis of the continuously acquired GNSS data.

Description

Technical Field

[0001] The present invention relates to a method for generating a learning model, a computer program, a microwave radiometer, and an estimation method.

Background Art

[0002] For the observation of precipitable water, that is, water vapor observation, it is known to use a GNSS receiver, a microwave radiometer, etc. Water vapor observation by a GNSS (Global Navigation Satellite System) receiver utilizes radio waves of multiple frequencies radiated from satellites. If radio waves radiated from four or more satellites at two or more different frequencies can be received, the delay amount of the radio waves can be captured. The delay amount of the radio waves corresponds to the amount of water vapor, and the amount of water vapor can be observed (for example, Patent Document 1). The precipitation prediction system described in Patent Document 1 calculates the precipitable water based on the zenith atmospheric delay amount calculated from the received GPS (GNSS) data and the satellite orbit information.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, since the precipitation prediction system of Patent Document 1 calculates the precipitable water based on GNSS data, it is difficult to calculate the local precipitable water.

[0005] The present invention has been made in view of such circumstances, and an object thereof is to provide a method for generating a learning model and the like that can efficiently estimate the precipitable water.

Means for Solving the Problems

[0006] The method for generating a learning model according to this embodiment involves acquiring training data that includes microwave measurement data, which is data obtained by measuring microwaves radiated from the atmosphere using a microwave radiometer, and the amount of precipitable water obtained based on GNSS signals received by a GNSS receiver installed in the microwave radiometer; generating a learning model that outputs the amount of precipitable water when microwave measurement data is input, based on the acquired training data; and further continuing to acquire training data that includes microwave measurement data measured by the microwave radiometer and the amount of precipitable water obtained from GNSS signals received by the GNSS receiver, and repeatedly retraining the learning model based on the continuously acquired training data.

[0007] The computer program according to this embodiment acquires microwave measurement data, which is data obtained by measuring microwaves radiated from the atmosphere using a microwave radiometer, and training data including the amount of precipitable water obtained based on GNSS signals received by a GNSS receiver installed in the microwave radiometer. Based on the acquired training data, it generates a learning model that outputs the amount of precipitable water when microwave measurement data is input. Furthermore, it continues to acquire training data including microwave measurement data measured by the microwave radiometer and the amount of precipitable water obtained from GNSS signals received by the GNSS receiver, and causes the computer to execute a process of repeatedly retraining the learning model based on the continuously acquired training data.

[0008] The microwave radiometer according to this embodiment includes a microwave measurement unit that measures microwave measurement data, which is data obtained by measuring microwaves radiated from the atmosphere; a GNSS receiver that receives GNSS signals; a learning model that outputs the amount of precipitable water when microwave measurement data is input; and a relearning unit that relearns the learning model. The relearning unit further continues to acquire training data, which includes microwave measurement data measured by the microwave measurement unit and the amount of precipitable water obtained from GNSS signals received by the GNSS receiver, and repeats the relearning of the learning model based on the continuously acquired training data.

[0009] The estimation method according to this embodiment involves a computer acquiring microwave measurement data, which is data obtained by measuring microwaves emitted from the atmosphere using a microwave radiometer, inputting the acquired microwave measurement data measured by the microwave radiometer into a learning model that outputs the amount of precipitable water when the microwave measurement data is input, outputting the amount of precipitable water, and further continuing to acquire training data that includes the microwave measurement data measured by the microwave radiometer and the amount of precipitable water obtained based on the GNSS signal received by the GNSS receiver installed in the microwave radiometer, and having the computer execute a process to repeatedly retrain the learning model based on the training data that is continuously acquired. [Effects of the Invention]

[0010] According to the present invention, it is possible to provide a method for generating a learning model that efficiently estimates the amount of precipitable water. [Brief explanation of the drawing]

[0011] [Figure 1] This is a block diagram showing an example of the configuration of a microwave radiometer according to Embodiment 1. [Figure 2] This is a functional block diagram illustrating the functional components (during learning) included in the control unit of a microwave radiometer. [Figure 3] This is an explanatory diagram (observation data table) showing an example of observation data from a microwave radiometer. [Figure 4] This is an explanatory diagram showing an example of microwave (sky-bb) radio wave intensity included in observational data. [Figure 5] This is an explanatory diagram (GNSS data table) showing an example of precipitable water amount derived based on GNSS signals (GNSS data). [Figure 6] This is a schematic diagram illustrating an example of a machine learning method for generating a precipitable water model (learned model). [Figure 7] This is a flowchart showing an example of the processing procedure (during learning) of the control unit of a microwave radiometer. [Figure 8]This is a functional block diagram illustrating the functional components (during relearning / operation) included in the control unit of a microwave radiometer. [Figure 9] This is an explanatory diagram (precipitable water change graph) showing an example of graphed information on changes in precipitable water. [Figure 10] This is an explanatory diagram (retraining history table) illustrating the retraining history (update history) of a precipitable water model (learned model). [Figure 11] This flowchart shows an example of the processing procedure (during operation) performed by the control unit of a microwave radiometer. [Modes for carrying out the invention]

[0012] (Embodiment 1) Embodiments of the present invention will be described below. Figure 1 is a block diagram showing an example of the configuration of a microwave radiometer 1 according to Embodiment 1. The microwave radiometer 1 is connected via an external network such as the Internet to an external server G (satellite orbit data server) where orbital information (calendar data) of GNSS satellites that transmit GNSS (global navigation satellite system) signals is stored.

[0013] The microwave radiometer 1 calculates the amount of precipitable water in the vertical upper atmosphere at a GNSS receiving point based on the GNSS signal received from the GNSS satellite, the orbital information (calendar data) of the GNSS satellite, and climate data at the GNSS signal receiving point. The microwave radiometer 1 generates training data using the amount of precipitable water calculated based on the received GNSS signal as the answer data, and the microwave signal strength measured at the receiving point as the problem data (first measurement data).

[0014] As will be described later, the microwave radiometer 1 functions as a learning model generation device that learns, for example, the parameters of a neural network using the generated training data and generates a precipitable water model 121 (learning model). Further, the microwave radiometer 1 functions as a precipitable water estimation device (water vapor observation device) that estimates the precipitable water in the vertical sky at the location where the microwave radiometer 1 is installed by inputting the radio wave intensity of the microwave measured at that location into the generated precipitable water model 121. Further, in parallel with the process of estimating the precipitable water using the precipitable water model 121, the microwave radiometer 1 continues to measure the radio wave intensity of the microwave and calculate the precipitable water based on the received GNSS signal, etc., thereby generating training data for re-learning the precipitable water model 121. The microwave radiometer 1 re-learns (re-generates) the precipitable water model 121 using the generated training data for re-learning.

[0015] Although the microwave radiometer 1 is assumed to generate training data for generating the precipitable water model 121 based on the measured radio wave intensity of the microwave, it is not limited to this. The microwave radiometer 1 may learn (generate) and re-learn (re-generate) the precipitable water model 121 with training data generated based on the temperature, humidity, atmospheric pressure, and rainfall amount at the location where the microwave is measured, in addition to the radio wave intensity of the microwave.

[0016] Although the microwave radiometer 1 is assumed to generate input data for the precipitable water model 121 based on the measured radio wave intensity of the microwave, it is not limited to this. The microwave radiometer 1 may input input data generated based on the temperature, humidity, atmospheric pressure, and rainfall amount at the location where the microwave is measured, in addition to the radio wave intensity of the microwave, into the precipitable water model 121 to obtain the precipitable water.

[0017] The external server G (satellite orbit data server) is composed of, for example, a cloud server connected to the Internet or the like. The external server G stores the orbit information (satellite ephemeris data) of the satellites that transmit GNSS signals. The microwave radiometer 1 can obtain the orbit information (satellite ephemeris data) of the satellites that transmit GNSS signals by accessing the external server G.

[0018] The microwave radiometer 1 includes a microwave measurement unit 15 that measures microwaves radiated from the atmosphere, a GNSS receiver 17, and a temperature sensor 141, a humidity sensor 142, a pressure sensor 143, and a rain sensor 144 for obtaining climate data at the location where the microwave radiometer 1 is installed. Furthermore, the microwave radiometer 1 includes a control unit 11, a storage unit 12, a communication unit 13, and an input / output I / F 14 for processing or processing various measurement data obtained by these sensor groups.

[0019] The microwave measurement unit 15 includes, for example, a radio window, a black body, a receiving horn, and an AD conversion circuit. The microwaves incident through the radio window are received by the receiving horn, and the received microwaves are converted by the AD conversion circuit into the radio wave intensity (dB) for each frequency. In the present embodiment, the microwaves to be measured are, for example, about 40 channels (frequencies) in the range of 16 GHz to 26 GHz, and the radio wave intensity (dB) of each of the 40 channels is measured. The peak of the intensity of the radio waves radiated from water vapor and cloud water in the upper atmosphere is 22 GHz. By obtaining (measuring) the radio wave intensity of about 40 channels (frequencies) in the range of 16 GHz to 26 GHz, the specific accuracy of the water vapor component and the cloud water component can be improved.

[0020] The microwave radiometer 1 periodically moves a blackbody through the receiving range of the receiving horn using a motor or the like, positioning the blackbody periodically so that it covers the receiving horn from above, thereby alternately acquiring microwaves emitted from the blackbody and microwaves emitted from above. The blackbody is composed of, for example, a vantablack and corresponds to a reference radio wave absorber. In this way, the microwave radiometer 1 measures microwaves from the reference radio wave absorber (blackbody) for, for example, 10 seconds, and then repeats the process of measuring microwaves from above for, for example, 10 seconds. For each of the approximately 40 channels (frequencies), the microwave radiometer 1 calculates the difference value obtained by subtracting the radio wave intensity of microwaves from the blackbody (bb [dB]) from the radio wave intensity of microwaves from above (sky [dB]). (sky-bb[dB]) is obtained as the measured microwave signal strength.

[0021] The microwave radiometer 1 may also automatically calibrate (detector calibrate) the linearity of analog equipment when measuring microwaves from a blackbody and the sky. The microwave radiometer 1 may also switch the attenuator connected to the receiving horn to check whether the digital value after conversion by the AD conversion circuit is linear (proportional relationship with respect to an ideal straight line) with respect to the analog value before conversion, and if the digital value after AD conversion exceeds the allowable error range, it may perform analog system linearity calibration.

[0022] The GNSS receiver 17 includes a GNSS antenna and a GNSS receiving unit that receives spread spectrum GNSS signals. The microwave radiometer 1 can acquire GNSS signals at the location (observation point) where it is installed by receiving GNSS signals using the GNSS receiver 17.

[0023] The temperature sensor 141 is composed of, for example, a thermistor and detects the temperature at the location where the microwave radiometer 1 is installed. The humidity sensor 142 is, for example, a capacitance-changing or resistance-changing electrical sensor and detects the amount of water vapor in the surrounding air at the location where the microwave radiometer 1 is installed and converts it into an electrical signal. The pressure sensor 143 is composed of, for example, a pressure-receiving element and detects the atmospheric pressure at the location where the microwave radiometer 1 is installed. The rain sensor 144 is, for example, a sensor that linearly changes the output voltage value based on the amount of rainfall and detects the amount of rainfall at the location where the microwave radiometer 1 is installed and converts it into an electrical signal.

[0024] These microwave measurement unit 15, GNSS receiver 17, temperature sensor 141, humidity sensor 142, pressure sensor 143, and rain sensor 144 are connected to the control unit 11 and storage unit 12 in a communicative manner, for example, by an input / output interface 14 or an internal bus. The measurement data measured by the microwave measurement unit 15 and temperature sensor 141, etc., is associated with time information indicating the time of measurement and is stored in the storage unit 12 as time-series data in a table format (see Figure 3). As will be described in detail later, the amount of precipitable water is calculated based on the GNSS signal received by the GNSS receiver 17 and the orbital information (calendar data) of the GNSS satellite acquired via the communication unit 13. The amount of precipitable water calculated based on the GNSS signal, etc., is stored in the storage unit 12 as time-series data in a table format (see Figure 5).

[0025] The control unit 11 has one or more CPUs (Central Processing Units), MPUs (Micro-Processing Units), GPUs (Graphics Processing Units), and other arithmetic processing units equipped with timing functions, and reads and executes the program P stored in the storage unit 12 to perform various information processing and control processing for the microwave measurement unit 15, etc.

[0026] The memory unit 12 includes volatile memory areas such as SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), and flash memory, as well as EEPROM or hard memory. This includes a non-volatile storage area such as a disk. The storage unit 12 pre-stores the program P and data referenced during processing. The program P stored in the storage unit 12 may be a program P (program product) read from a recording medium 120 that the control unit 11 can read. Alternatively, the program P (program product) may be downloaded from an external computer (not shown) connected to a communication network (not shown) and stored in the storage unit 12. The storage unit 12 also stores the actual files that constitute the learning model (precipitable water model 121). These actual files may be composed of parts of the program P.

[0027] The communication unit 13 is a communication module or communication interface for communicating with an external server G (satellite orbit data server) by wire or wireless connection, for example, Wi-Fi (registered trademark). The module is a narrow-range wireless communication module such as Bluetooth®, or a wide-range wireless communication module such as 4G or 5G. The control unit 11 communicates with an external server G via the communication unit 13, for example, through an external network such as the Internet.

[0028] The input / output interface 14 includes, for example, a communication interface compliant with a communication standard such as USB, or a connector for connecting to an internal bus. The input / output interface 14 may also be connected to a microwave measurement unit 15, a GNSS receiver 17, a temperature sensor 141, a humidity sensor 142, a pressure sensor 143, and a rain sensor 144. A display device 16, such as a liquid crystal display, may also be connected to the input / output interface 14.

[0029] In this embodiment, the microwave radiometer 1 includes a microwave measurement unit 15, various sensors, and a control unit 11 for data processing, but is not limited to this. The control unit 11 for data processing may be included in a separate computer from the microwave radiometer 1, and the microwave radiometer 1 and the computer may be connected in a communicative manner. The computer may process the various data measured by the microwave radiometer 1 to generate a learning model and perform estimation processing using the learning model. The computer may acquire observation data (radio wave intensity and climate data), GNSS signals, etc. from each of the microwave radiometers 1 installed at multiple observation points, and generate training data based on the observation data and GNSS signals etc. for each observation point. In this way, the precipitable water model 121 learned using training data using observation data and GNSS signals etc. from multiple observation points may be applied to each microwave radiometer 1, and each microwave radiometer 1 may perform estimation processing using the precipitable water model 121.

[0030] Figure 2 is a functional block diagram illustrating the functional units (during learning) included in the control unit 11 of the microwave radiometer 1. The control unit 11 of the microwave radiometer 1 functions as an acquisition unit 111, a training data generation unit 112, and a learning unit 113 by executing the program P stored in the memory unit 12.

[0031] The acquisition unit 111 acquires radio wave intensity (first measurement data) from the microwave measurement unit 15 for each of approximately 40 channels (frequencies) in the range of 16 GHz to 26 GHz. The acquisition unit 111 acquires measured values ​​such as temperature (climate data) output from the temperature sensor 141, humidity sensor 142, pressure sensor 143, and rain sensor 144. The acquisition unit 111 acquires GNSS signals from the GNSS receiver 17. In this way, the acquisition unit 111 acquires microwave radio wave intensity (first measurement data), climate data, and GNSS signals at the location where the microwave radiometer 1 is installed (observation site), and these data acquired at the same time are temporally corresponding. The acquisition unit 111 acquires orbital information (calendar data) of the GNSS satellite corresponding to the time of transmission of the GNSS signal from an external server G (satellite orbit data server) via the communication unit 13.

[0032] Figure 3 is an explanatory diagram showing an example of observation data from the microwave radiometer 1. The acquisition unit 111 associates the acquired radio wave intensity (first measurement data) and measured values ​​such as temperature (climate data) with the time at which this data was acquired (measurement time) and stores them in the storage unit 12 in a table format (observation data table), for example. The observation data table includes time, radio wave intensity from a blackbody, radio wave intensity from the upper atmosphere, temperature, humidity, atmospheric pressure, and rainfall as management items (fields).

[0033] The time management field stores date and time information indicating the date and time of observation (acquisition) of the observation data (radio wave intensity and climate data). The management field for radio wave intensity from a blackbody stores the radio wave intensity of each of approximately 40 channels (frequencies) in the range of 16 GHz to 26 GHz, emitted from the blackbody, as calibration observation data. The management field for radio wave intensity from the upper atmosphere stores the radio wave intensity of each of approximately 40 channels (frequencies) in the range of 16 GHz to 26 GHz, emitted from the upper atmosphere, as upper-air observation data.

[0034] In this embodiment, the radio wave intensity from the blackbody and the sky are stored in a table format, but the embodiment is not limited to this, and the radio wave intensity from the sky at each frequency is also stored. It may also be possible to store the difference value (sky-bb) obtained by subtracting the blackbody's radio wave intensity (bb) from the upper atmosphere's radio wave intensity. Figure 4 is an explanatory diagram showing an example of microwave (sky-bb) radio wave intensity included in the observation data. In this explanatory diagram, the vertical axis shows the difference value (sky-bb) radio wave intensity, and the horizontal axis shows the channel (frequency). Since the blackbody (reference radio wave absorber) is at room temperature (higher temperature than the upper atmosphere), its radiation intensity is stronger than that of the upper atmosphere, so the blackbody's radio wave intensity (bb) is subtracted from the upper atmosphere's radio wave intensity (sky). The difference value (sky-bb) will be a negative value, for example, ranging from 0dB to about -5dB.

[0035] The temperature field stores the temperature at the location where the microwave radiometer 1 is installed. The humidity field stores the absolute or relative humidity at the location where the microwave radiometer 1 is installed. The atmospheric pressure field stores the atmospheric pressure at the location where the microwave radiometer 1 is installed. The rainfall field stores the voltage value indicating the amount of rainfall at the location where the microwave radiometer 1 is installed.

[0036] The training data generation unit 112 generates training data for generating (training) the precipitable water model 121 based on observation data (radio wave intensity and climate data) measured by the microwave radiometer 1, GNSS signals received by the GNSS receiver 17, and orbital information of GNSS satellites acquired from the external server G. This training data includes problem data generated based on observation data (radio wave intensity and climate data) and answer data (precipitable water) generated based on GNSS signals, orbital information, and climate data. The radio wave intensity is indicated by the difference value (sky-bb) obtained by subtracting the blackbody radio wave intensity (bb) from the upper-air radio wave intensity (sky) at each frequency. .

[0037] The training data generation unit 112 may generate problem data based on multiple observation data (radio wave intensity and climate data) measured at multiple points in time that fall within a predetermined time period corresponding to the GNSS signal reception period. The training data generation unit 112 may also generate problem data based on averaged, standardized, or normalized observation data (radio wave intensity and climate data) by averaging, standardizing, or normalizing the observation data (radio wave intensity and climate data) at the multiple points in time. In other words, the training data generation unit 112 may include a standardization unit that standardizes the observation data.

[0038] The training data generation unit 112 may perform dimensionality reduction processing on first measurement data consisting of, for example, the radio wave intensity of about 40 channels (frequencies) in the range of 16 GHz to 26 GHz, and generate problem data based on the first measurement data after dimensionality reduction processing. In other words, the training data generation unit 112 may include a dimensionality reduction unit that performs dimensionality reduction processing on the observed data. The dimensionality reduction processing may be performed on the radio wave intensity at each of the multiple frequencies, for example, by principal component analysis (PCA), and a predetermined number of principal components from the first rank onward may be used as first measurement data for generating (training) the learning model. The dimensionality reduction processing may select, for example, three predetermined principal components, and the first principal component, second principal component, and third principal component. Principal component analysis is not limited to PCA, and algorithms such as factor analysis, multifactor analysis, autoencoder, independent component analysis, or non-negative matrix factorization may be used.

[0039] In this embodiment, the training data generation unit 112 generates problem data based on the radio wave intensity measured by the microwave radiometer 1 and climate data (temperature, humidity, atmospheric pressure, and rainfall), but is not limited to this. The training data generation unit 112 may generate problem data based only on the radio wave intensity measured by the microwave radiometer 1. Alternatively, the training data generation unit 112 may generate problem data based on temperature, humidity, and atmospheric pressure in addition to the radio wave intensity. In other words, the training data generation unit 112 may generate problem data based on some of the data included in the climate data in addition to the radio wave intensity.

[0040] The training data generation unit 112 calculates the amount of precipitable water using a known method based on the acquired GNSS signal, GNSS satellite orbit information, and climate data. The training data generation unit 112 calculates the zenith atmospheric delay based on the GNSS signal received by the GNSS receiver 17 and the GNSS satellite orbit information acquired from the external server G (satellite orbit data server). The training data generation unit 112 calculates the zenith hydrostatic delay from the atmospheric pressure (atmospheric pressure at the observation point) included in the climate data. The training data generation unit 112 calculates the zenith moist delay by subtracting the zenith hydrostatic delay from the zenith atmospheric delay. The training data generation unit 112 calculates the amount of precipitable water by converting the zenith moist delay using a proportionality constant determined based on the atmospheric pressure and temperature included in the climate data. This amount of precipitable water corresponds to the answer data for generating the precipitable water model 121.

[0041] Figure 5 is an explanatory diagram (GNSS data table) showing an example of precipitable water derived based on GNSS signals, etc. (GNSS data). The acquisition unit 111 stores the precipitable water calculated based on the acquired GNSS data, such as the GNSS signals and GNSS satellite orbit information, in the storage unit 12 in a table format (GNSS data table), associating it with the time of reception of the GNSS signals. The GNSS data table includes, for example, time, zenith atmospheric delay, zenith hydrostatic delay, zenith moist delay, and precipitable water as management items (fields).

[0042] The time management field stores date and time information indicating the date and time the GNSS signal was received. The zenith atmospheric delay management field stores the zenith atmospheric delay calculated based on the GNSS signal received by the GNSS receiver 17 and the orbital information of the GNSS satellite obtained from the external server G (satellite orbit data server). The zenith hydrostatic delay management field stores the zenith hydrostatic delay calculated based on climate data including atmospheric pressure obtained via the barometric pressure sensor 143, etc. The zenith moist delay management field stores the zenith moist delay calculated by subtracting the zenith hydrostatic delay from the zenith atmospheric delay. The precipitable water management field stores the precipitable water converted from the zenith moist delay using a proportionality constant determined based on the atmospheric pressure and temperature included in the climate data. In this embodiment, the GNSS data table includes time, zenith atmospheric delay, zenith hydrostatic delay, zenith moist delay, and precipitable water; however, this is merely an example, and the GNSS data table may include only time and precipitable water.

[0043] The training data generation unit 112 generates training data for the precipitable water model 121, which is used to generate (train) the precipitable water model 121, using problem data generated based on observation data (radio wave intensity and climate data) and answer data consisting of precipitable water calculated based on GNSS signals, etc. (GNSS data). The answer data (precipitable water) labeled to the problem data (radio wave intensity and climate data) are time-corresponding such that the observation time of microwaves, etc. and the reception time of GNSS signals are substantially simultaneous or within the same time period.

[0044] Figure 6 is a schematic diagram showing an example of a machine learning method for generating the precipitable water model 121 (learned model). The learning unit 113 trains a neural network using training data for the precipitable water model 121, and generates the precipitable water model 121, which takes observational data (radio wave intensity and climate data) as input and outputs precipitable water. The neural network (precipitable water model 121) trained using the training data is intended to be used as a program module that is part of artificial intelligence software. The precipitable water model 121 is used in the microwave radiometer 1 equipped with a control unit 11 (CPU, etc.) and a memory unit 12 as described above, and by being executed in the microwave radiometer 1 which has computational processing capabilities in this way, a neural network system is formed.

[0045] Precipitable water model 121 is constructed using, for example, a DNN (Deep Neural Network), and The system comprises an input layer that accepts measurement data, a hidden layer that extracts features from the observation data, and an output layer that outputs the amount of precipitable water. The input layer has multiple neurons that accept input values ​​such as radio wave intensity or temperature included in the observation data, and passes the input values ​​to the hidden layer. The hidden layer is defined using an activation function such as a ReLU function or a sigmoid function, and has multiple neurons that extract features from each of the input values, and passes the extracted features to the output layer. The parameters of the activation function, such as the weighting coefficient and bias value, are optimized using backpropagation. The output layer is composed of, for example, a fully connected layer, and outputs the amount of precipitable water based on the features output from the hidden layer.

[0046] In this embodiment, the precipitable water model 121 is a DNN, but it is not limited to this, and other neural networks such as RNN (Recurrent Neural Network), LSTM (Long-short term model), CNN, SVM (Support Vector Machine), and Bayonet are also acceptable. The learning model may be constructed using other learning algorithms such as Gian networks, linear regression, regression trees, multiple regression, random forests, or ensembles. For example, if multiple regression with multiple types of variables is used, the precipitable water model 121 may be represented by a transformation formula using polynomial regression. In this case, the transformation formula is defined by the dot product of the vectors of the principal component-analyzed observational data (radio wave intensity and climate data) and unknown coefficients (S1 to Sn) that have the same number of dimensions (n) as the principal component-analyzed observational data (radio wave intensity and climate data). These coefficients (S1 to Sn) are calculated, for example, by fitting using the least squares method.

[0047] Figure 7 is a flowchart showing an example of the processing procedure (during learning) in the control unit 11 of the microwave radiometer 1. The control unit 11 of the microwave radiometer 1 receives operator input, for example, from a keyboard connected to the input / output I / F 14, and performs the following processing based on the received operation.

[0048] The control unit 11 of the microwave radiometer 1 acquires first measurement data measured by the microwave radiometer 1 (microwave measurement unit 15) (S1). The control unit 11 of the microwave radiometer 1 acquires radio wave intensity (first measurement data) for each of approximately 40 channels (frequencies) in the range of 16 GHz to 26 GHz. For each of these channels (frequencies), the control unit 11 of the microwave radiometer 1 acquires radio wave intensity from the sky to radio waves from a blackbody. The first measurement data may be the difference value (sky-bb) obtained by subtracting the intensity (bb).

[0049] The control unit 11 of the microwave radiometer 1 acquires climate data for the location where the microwave radiometer 1 is installed (S2). The control unit 11 of the microwave radiometer 1 acquires measured values ​​(climate data) such as temperature output from the temperature sensor 141, humidity sensor 142, pressure sensor 143, and rain sensor 144. The acquisition of climate data is not limited to acquisition from sensors such as the temperature sensor 141 equipped in the microwave radiometer 1; the control unit 11 of the microwave radiometer 1 may also acquire climate data for the location where the microwave radiometer 1 is installed from a climate data server connected to the internet, for example, via the communication unit 13.

[0050] The control unit 11 of the microwave radiometer 1 acquires the GNSS signal (S3). The control unit 11 of the microwave radiometer 1 acquires the GNSS signal received by the GNSS receiver 17 installed in the microwave radiometer 1.

[0051] The control unit 11 of the microwave radiometer 1 acquires orbital information of GNSS satellites (S4). The control unit 11 of the microwave radiometer 1 accesses an external server G (satellite orbit data server) via the communication unit 13 and acquires orbital information (calendar data) of satellites (GNSS satellites) that transmit GNSS signals.

[0052] The control unit 11 of the microwave radiometer 1 generates training data for training a precipitable water model 121 (learning model) that outputs precipitation, based on the first measurement data, climate data, GNSS signals, and orbital information (S5). The control unit 11 of the microwave radiometer 1 generates problem data based on the observation data (radio wave intensity and climate data) measured by the microwave radiometer 1, generates answer data consisting of precipitable water calculated based on GNSS signals, etc., and generates training data by labeling the problem data with the answer data (precipitable water). When generating the training data, the control unit 11 of the microwave radiometer 1 may perform standardization processing or dimensionality reduction processing on the observation data (radio wave intensity and climate data) at multiple time points.

[0053] The control unit 11 of the microwave radiometer 1 generates a precipitable water model 121 (learned model) by learning using training data (S6). The control unit 11 of the microwave radiometer 1 generates the precipitable water model 121 (learned model) by learning, for example, the parameters of a neural network using the training data for the precipitable water model 121.

[0054] In this embodiment, the control unit 11 of the microwave radiometer 1 acquires radio wave intensity (first measurement data) and climate data as observation data, but is not limited to this. The control unit 11 of the microwave radiometer 1 may generate training data based only on radio wave intensity (first measurement data), or on radio wave intensity (first measurement data) plus some climate data such as temperature, humidity, and atmospheric pressure, and generate a precipitable water model 121.

[0055] Figure 8 is a functional block diagram illustrating the functional units (during relearning / operation) included in the control unit 11 of the microwave radiometer 1. The control unit 11 of the microwave radiometer 1 functions as an acquisition unit 111, an input data generation unit 114, an output unit 115, a precipitable water amount calculation unit 116, a data storage unit 117, and a relearning unit 118 by executing the program P stored in the memory unit 12.

[0056] The acquisition unit 111 acquires radio wave intensity (first measurement data) from the microwave measurement unit 15 for each of approximately 40 channels (frequencies) in the range of 16 GHz to 26 GHz. The acquisition unit 111 also acquires measured values ​​(climate data) such as temperature output from the temperature sensor 141, humidity sensor 142, pressure sensor 143, and rain sensor 144. The acquisition unit 111 outputs these acquired radio wave intensity (first measurement data) and climate data to the input data generation unit 114.

[0057] The input data generation unit 114 generates input data to be input to the precipitable water model 121 based on the observation data (radio wave intensity (first measurement data) and climate data) acquired by the acquisition unit 111. The radio wave intensity included in the input data is calculated from the radio wave intensity in the upper atmosphere (sky) to the radio wave intensity of a blackbody. The data may also be expressed as a difference value (sky-bb) obtained by subtracting the wave intensity (bb). The input data generation unit 114 may generate input data based on multiple observation data measured at multiple points in time within a predetermined time period. The input data generation unit 114 may generate input data by performing standardization processing or dimensionality reduction processing on the observation data at the multiple points in time, similar to the training data generation unit 112 described above.

[0058] The input data generation unit 114 inputs the generated input data into the precipitable water model 121. The precipitable water model 121 estimates the precipitable water based on the input data.

[0059] The output unit 115 converts the amount of precipitable water estimated by the precipitable water model 121 into, for example, graph-formatted image data, and outputs the image data to the display device 16. The output unit 115 may also convert the amount of precipitable water at multiple consecutive observation points into graph-formatted image data showing changes over time.

[0060] Figure 9 is an explanatory diagram (precipitable water change graph) showing an example of graphed information on changes in precipitable water. This precipitable water change graph is an example of a graph showing the time change of precipitable water at multiple consecutive observation points. The horizontal axis of the precipitable water change graph indicates the observation point by microwave radiometer 1. The vertical axis indicates the amount of precipitable water (PWV) in the upper atmosphere (vertical direction) at the observation point by microwave radiometer 1.

[0061] The acquisition unit 111 further acquires GNSS signals from the GNSS receiver 17 and, via the communication unit 13, acquires orbital information (calendar data) of the GNSS satellite at the time of transmission of the GNSS signal from an external server G (satellite orbit data server). The acquisition unit 111 outputs the acquired GNSS signals, orbital information, and climate data to the precipitable water calculation unit 116.

[0062] The precipitable water calculation unit 116 calculates the amount of precipitable water based on the GNSS signal, orbital information, and climate data output from the acquisition unit 111. Similar to the training data generation unit 112 described above, the precipitable water calculation unit 116 calculates the amount of precipitable water using a known method based on the acquired GNSS signal, GNSS satellite orbital information, and climate data.

[0063] The data storage unit 117 stores the observation data (radio wave intensity and climate data) output from the input data generation unit 114 in the storage unit 12, associating it with the observation time. The data storage unit 117 may store the observation data in an observation data table, for example. The data storage unit 117 stores the amount of precipitable water calculated by the precipitable water calculation unit 116 in the storage unit 12, associating it with the time of reception of the GNSS signal. In addition to the amount of precipitable water, the data storage unit 117 may also store intermediate data such as the zenith atmospheric delay, zenith hydrostatic pressure delay, and zenith moist delay, which are calculated based on the GNSS signal, etc., when calculating the amount of precipitable water, in a GNSS data table, for example. The storage of observation data and the amount of precipitable water, etc. by the data storage unit 117 (storage in the storage unit 12) is performed in parallel with the output (estimation) of the amount of precipitable water by the precipitable water model 121 described above.

[0064] The relearning unit 118 generates training data for relearning the precipitable water model 121 based on the observation data and precipitable water stored in the memory unit 12 by the data storage unit 117, and then retrains (regenerates) the precipitable water model 121 using the training data.

[0065] The relearning unit 118 may, for example, refer to a scheduled relearning date and time (relearning schedule) pre-stored in the memory unit 12, and when the scheduled relearning date and time arrives, generate training data based on the observation data and precipitable water stored in the memory unit 12, and relearn (regenerate) the precipitable water model 121. Alternatively, the relearning unit 118 may, for example, determine a scheduled relearning date and time based on a predetermined period (relearning period) pre-stored in the memory unit 12, and when the scheduled relearning date and time arrives, generate training data based on the observation data and precipitable water stored in the memory unit 12, and relearn (regenerate) the precipitable water model 121. Alternatively, the retraining unit 118 may generate training data and retrain (regenerate) the precipitable water model 121 if the difference (amount of difference) between the precipitable water amount estimated by the currently operating precipitable water model 121 and the precipitable water amount calculated (derived) based on the GNSS signal exceeds a predetermined value (above a threshold).

[0066] Retraining using the precipitable water model 121 is not limited to using the training data stored (memorized) by the data storage unit 117 to fine-tune or transfer-learn an already generated learning model. The storage unit 12 stores all of the previously used training data, and the retraining unit 118 may use the previously used training data and the training data added in the current retraining to learn the parameters of the untrained neural network and regenerate the precipitable water model 121.

[0067] The retraining unit 118 replaces (updates) the previously used precipitable water model 121 (the current precipitable water model 121) with the retrained (regenerated) precipitable water model 121, thereby starting the application of the retrained precipitable water model 121. From this point onward, the output (estimation) of precipitable water is performed using the latest precipitable water model 121, which has been retrained with the latest training data.

[0068] The relearning unit 118 may generate and output relearning history information when the precipitable water model 121 is relearned, and store it in the memory unit 12. The relearning unit 118 may also store the relearning history information in a table format (relearning history table) in the memory unit 12.

[0069] Figure 10 is an explanatory diagram (retraining history table) illustrating the retraining history (update history) of the precipitable water model 121 (learning model). The retraining history table includes management items (fields) such as the cumulative number of retrainings, the date and time of retraining (date and time of update application), and the observation period of the original data.

[0070] The cumulative number of retraining attempts field stores the cumulative number of times each retraining attempt has been performed. This cumulative number allows you to understand the total number of retraining attempts performed to date. This cumulative number corresponds to the version number of the retrained precipitable water model 121.

[0071] The "Retraining Date and Time (Update Application Date and Time)" management item (field) stores the date and time when retraining was performed, i.e., the date and time when the retrained precipitable water model 121 was applied. The "Observation Period of the Original Data to be Stored" management item (field) stores the observation period (reception period) of the original data, such as observation data and GNSS signals, used to generate the training data added for retraining. By storing and managing information about the retraining history in the retraining history table in this way, the traceability of the retrained precipitable water model 121 can be ensured.

[0072] Figure 11 is a flowchart showing an example of the processing procedure (during operation) by the control unit 11 of the microwave radiometer 1. The control unit 11 of the microwave radiometer 1 receives operator input, for example, via a keyboard connected to the input / output I / F 14, at the observation site where the microwave radiometer 1 is installed, and continuously performs the following processing based on the received input.

[0073] The control unit 11 of the microwave radiometer 1 acquires first measurement data measured by the microwave radiometer 1 (microwave measurement unit 15) (S101). The control unit 11 of the microwave radiometer 1 acquires climate data for the location where the microwave radiometer 1 is installed (S102). The control unit 11 of the microwave radiometer 1 performs the processing from S101 to S102 in the same way as processing S1 to S2. The control unit 11 of the microwave radiometer 1 may also associate the acquired first measurement data (radio wave intensity) and climate data with the observation time of these data and store them in the observation data table stored in the storage unit 12.

[0074] The control unit 11 of the microwave radiometer 1 generates input data based on the first measurement data and climate data (S103). The radio wave intensity included in the input data may be shown as the difference value (sky-bb) obtained by subtracting the radio wave intensity of a blackbody (bb) from the radio wave intensity in the upper atmosphere (sky). The control unit 11 of the microwave radiometer 1 generates input data based on the first measurement data and climate data (S103). The input data may be generated based on multiple observation data measured at multiple time points within the inter-duration. The control unit 11 of the microwave radiometer 1 may generate the input data by performing standardization processing or dimensionality reduction processing on the observation data at the multiple time points.

[0075] The control unit 11 of the microwave radiometer 1 inputs the input data into the precipitable water model 121 and obtains the precipitable water amount (S104). The control unit 11 of the microwave radiometer 1 inputs the input data into the precipitable water model 121 and obtains the precipitable water amount estimated by the precipitable water model 121.

[0076] The control unit 11 of the microwave radiometer 1 outputs the amount of precipitable water (S105). The control unit 11 of the microwave radiometer 1 outputs the information regarding the amount of precipitable water to the display device 16, for example, by graphing it.

[0077] The control unit 11 of the microwave radiometer 1 performs a loop process to execute the processing from S101 again after processing S105. By repeating the processing from S101 to S105 in this way, information on precipitable water that changes over time can be added or updated and continuously output based on the observation results continuously performed by the microwave radiometer 1. The control unit 11 of the microwave radiometer 1 may also continuously graph the information on precipitable water that changes over time and output it to the display device 16, thereby periodically updating the graph displayed on the display device 16. In other words, the control unit 11 of the microwave radiometer 1 can acquire precipitable water at multiple points in time (multiple time-series data) by sequentially inputting multiple input data generated from multiple time-series observation data (first measurement data and climate data) into the precipitable water model 121.

[0078] In this embodiment, the control unit 11 of the microwave radiometer 1 acquires radio wave intensity (first measurement data) and climate data when generating input data, but is not limited to this. The control unit 11 of the microwave radiometer 1 may generate input data based only on radio wave intensity (first measurement data), or on radio wave intensity (first measurement data) plus some climate data such as temperature, humidity, and atmospheric pressure, and input this input data to the precipitable water model 121 to acquire precipitable water.

[0079] The control unit 11 of the microwave radiometer 1 acquires a GNSS signal (S111). The control unit 11 of the microwave radiometer 1 acquires orbital information of the GNSS satellite (S112). The control unit 11 of the microwave radiometer 1 acquires a GNSS signal via the GNSS receiver 17 and also acquires orbital information of the GNSS satellite at the time the GNSS signal was transmitted from the external server G (satellite orbit data server) via the communication unit 13.

[0080] The control unit 11 of the microwave radiometer 1 calculates the amount of precipitable water based on the acquired GNSS signal, etc. (S113). The control unit 11 of the microwave radiometer 1 calculates the amount of precipitable water based on the acquired GNSS signal, orbital information, and climate data including atmospheric pressure acquired in S102. The control unit 11 of the microwave radiometer 1 may store the calculated amount of precipitable water in the GNSS data table stored in the storage unit 12, in association with the reception time of the GNSS signal. After processing in S113, the control unit 11 of the microwave radiometer 1 performs loop processing to execute the processing from S111 again.

[0081] The control unit 11 of the microwave radiometer 1 determines whether or not to retrain the precipitable water model 121 (learned model) (S121). For example, if the control unit 11 of the microwave radiometer 1 retrains the precipitable water model 121 at a predetermined period, it may determine whether or not to retrain based on whether a predetermined period corresponding to the period has elapsed since the last retraining.

[0082] Information regarding a predetermined period (predetermined time) is stored in the memory unit 12. For example, if the predetermined period (predetermined time) is set to one week, the control unit 11 of the microwave radiometer 1 refers to the relearning history table, obtains the date and time of the previous (most recent) relearning, and determines whether relearning is necessary based on whether one week (predetermined time) has elapsed since that relearning date and time. If the predetermined period has elapsed, the control unit 11 of the microwave radiometer 1 determines that relearning is necessary. If the predetermined period has not elapsed, the control unit 11 of the microwave radiometer 1 determines that relearning is unnecessary. The predetermined period (predetermined time) may be stored in the memory unit 12 as information that can be changed by the operator of the microwave radiometer 1, etc. The control unit 11 of the microwave radiometer 1 may change the predetermined period (predetermined time) based on operator input, such as from a keyboard connected to the input / output I / F 14, or based on change instruction data received via the communication unit 13.

[0083] Alternatively, the control unit 11 of the microwave radiometer 1 may determine to retrain the precipitable water model 121 if, for example, the difference (difference amount) between the precipitable water amount estimated by the precipitable water model 121 and the precipitable water amount derived based on the GNSS signal exceeds a predetermined value. As described above, the control unit 11 of the microwave radiometer 1 continuously calculates and stores the precipitable water amount based on the received GNSS signal and orbital information in parallel with the process of outputting (estimating) the precipitable water amount using the precipitable water model 121, and continuously compares the precipitable water amount from the precipitable water model 121 with the precipitable water amount from the GNSS signal, etc. If the absolute value of the difference (difference amount) resulting from this comparison exceeds a predetermined value, the control unit 11 of the microwave radiometer 1 may determine to retrain, and if the absolute value of the difference (difference amount) is less than the predetermined value, the control unit 11 of the microwave radiometer 1 may determine not to retrain.

[0084] If relearning is not performed (S121: NO), the control unit 11 of the microwave radiometer 1 performs a loop process to execute the process of S121 again. If relearning is performed (S121: YES), the control unit 11 of the microwave radiometer 1 generates training data for relearning the precipitable water model 121 that outputs precipitation, based on the first measurement data, climate data, GNSS signal and orbital information (S122). The control unit 11 of the microwave radiometer 1, for example, refers to the observation data table and GNSS data table stored in the memory unit 12 and generates training data for relearning using problem data generated based on observation data (radio wave intensity and climate data) and answer data consisting of precipitable water calculated based on GNSS signals etc. (GNSS data).

[0085] The control unit 11 of the microwave radiometer 1 regenerates the precipitable water model 121 by relearning it using training data (S123). The control unit 11 of the microwave radiometer 1 starts applying the relearned (regenerated) precipitable water model 121 (S124). The control unit 11 of the microwave radiometer 1 starts applying the relearned (regenerated) precipitable water model 121 by replacing (updating) the precipitable water model 121 that has been used up to the present (current precipitable water model 121) with the relearned (regenerated) precipitable water model 121. From this point onward, the output (estimation) of precipitable water is performed using the latest precipitable water model 121 that has been relearned with the latest training data.

[0086] The control unit 11 of the microwave radiometer 1 outputs relearning history information (S125). The control unit 11 of the microwave radiometer 1 generates and outputs relearning history information and stores it in the relearning history table, thereby storing the history information in the storage unit 12.

[0087] The control unit 11 of the microwave radiometer 1 may retain the precipitable water model 121 as a backup when the precipitable water model 121 is retrained. The control unit 11 of the microwave radiometer 1 periodically retrains the precipitable water model 121, so that the precipitable water model 121 is updated (version upgraded) according to the number of retrainings. When the precipitable water model 121 is retrained (updated), the control unit 11 of the microwave radiometer 1 may, for example, assign a version number or retraining (update) date and time to each of these precipitable water model 121, and save not only the latest precipitable water model 121 but also the precipitable water model 121 used in the past. In this way, the control unit 11 of the microwave radiometer 1 may manage the versions of multiple precipitable water models 121 that are generated (updated) by repeated retraining. The control unit 11 of the microwave radiometer 1 may retrain the precipitable water model 121 and, upon generating the latest precipitable water model 121, perform a validity check of the latest precipitable water model 121. If the validity check is unfavorable, it may perform a recovery process to revert to the pre-retrained precipitable water model 121, which is saved as a backup.

[0088] The control unit 11 of the microwave radiometer 1 performs a series of processes from S101 to S105, a series of processes from S111 to S112, and a series of processes from S121 to S125 in parallel, for example, using multiprocessing or multithreading. By performing each of these series of processes in parallel, the reception of observation data and GNSS signals at the observation site where the microwave radiometer 1 is installed can be continued in parallel with the process of outputting (estimating) the amount of precipitable water by the precipitable water model 121. This makes it possible to generate training data based on the observation site in real time and to continuously retrain the precipitable water model 121 using this training data (repeat). By using the continuously retrained precipitable water model 121, the amount of local precipitable water at the observation site can be efficiently estimated.

[0089] According to this embodiment, the precipitable water model 121 (learning model) generated by the training data is implemented in the microwave radiometer 1 and retrained using training data that includes first measurement data (radio wave intensity: sky-bb) measured by the microwave radiometer 1 at the location where the microwave radiometer 1 is installed, and precipitable water based on the GNSS signal received by the GNSS receiver 17 included in the microwave radiometer 1. Since the training data for this retraining consists of first measurement data measured by the microwave radiometer 1 on which the precipitable water model 121 (learning model) is implemented, and data based on the GNSS signal received by the GNSS receiver 17 provided in the microwave radiometer 1, retraining can be performed in a manner suitable for the installation environment of the microwave radiometer 1. The control unit 11 of the microwave radiometer 1 repeats the retraining of the precipitable water model 121 (learning model) so that this retraining is carried out continuously at predetermined intervals of, for example, one week or one month, enabling flexible signal processing that can be applied to the latest data. This eliminates the need for blackbody calibration using liquid nitrogen, allowing the microwave radiometer 1 to always be operated with the latest calibration results.

[0090] According to this embodiment, the control unit 11 of the microwave radiometer 1, in parallel with the process of outputting (estimating) the amount of precipitable water using the precipitable water model 121, continuously compares the outputted (estimated) amount of precipitable water with the amount of precipitable water obtained from the GNSS signal corresponding to the measurement time of the first measurement data input to the learning model. If, as a result of this comparison, for example, the difference (amount of difference) between the amount of precipitable water estimated by the precipitable water model 121 and the amount of precipitable water derived based on the GNSS signal exceeds a predetermined value, the control unit 11 of the microwave radiometer 1 retrains the precipitable water model 121. This prevents excessive retraining and an unnecessary increase in the computational load on the control unit 11 of the microwave radiometer 1.

[0091] According to this embodiment, the control unit 11 of the microwave radiometer 1 periodically retrains the precipitable water model 121, generates history information of the retraining, stores it in the memory unit 12, and outputs (transmits) it to a cloud server or the like (observation data aggregation server) connected to the internet via the communication unit 13. The retraining history information includes, for example, the cumulative number of retrainings, the date and time the retraining was performed, and the acquisition period (observation period) of the first measurement data and GNSS signals included in the training data used for retraining. By outputting (transmitting) this retraining history information to a cloud server or the like (observation data aggregation server), the state of the learning model applied to the microwave radiometer 1 installed at the observation site can be notified to the administrator of the microwave radiometer 1 in a timely manner.

[0092] According to this embodiment, the multiple frequencies included in one measurement data set are, for example, about 40 channels (frequencies) in the range of 16 GHz to 26 GHz, and the radio wave intensity (dB) of each of the 40 channels is stored in the storage unit 12. The control unit 11 performs dimensionality reduction processing on the acquired and stored radio wave intensity of each of the 40 channels, thereby reducing the number of dimensions of the first measurement data (problem data) included in the training data and reducing computational costs.

[0093] According to this embodiment, the problem data included in the training data includes not only the first measurement data measured by the microwave radiometer 1, but also climate data at the location where the microwave radiometer 1 is installed. This climate data includes at least one of temperature, humidity (absolute humidity or relative humidity), atmospheric pressure, and rainfall, and may include all of these physical quantities. By training using such problem data, the estimation accuracy of the precipitable water model 121 can be improved.

[0094] The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The technical features described in each embodiment can be combined with each other, and the scope of the present invention is intended to include all modifications within the claims and equivalents thereof. [Explanation of symbols]

[0095] G External server (satellite orbit data server) 1. Microwave radiometer 11 Control Unit 111 Acquisition Department 112 Training Data Generation Unit 113 Learning Department 114 Input Data Generation Unit 115 Output section 116 Precipitable amount calculation section 117 Data Storage Unit 118 Relearning Section 12 Storage section 120 recording media 121 Precipitable water model (learning model) P Program (Program Product) 13 Communications Department 14 Input / Output Interfaces 141 Temperature Sensor 142 Humidity Sensor 143 Barometric pressure sensor 144 Rain sensor 15 Microwave Measurement Unit 16 Display device 17 GNSS receiver

Claims

1. Microwave measurement data, which is data obtained by measuring microwaves emitted from the atmosphere using a microwave radiometer, and training data including precipitable water amount obtained based on the GNSS signal received by the GNSS receiver installed on the microwave radiometer are acquired. Based on the acquired training data, a learning model is generated that outputs the amount of precipitable water when microwave measurement data is input. After generating the aforementioned learning model, the acquisition of training data including microwave measurement data measured by the microwave radiometer and precipitable water amount obtained from the GNSS signal received by the GNSS receiver is continued. Based on the continuously acquired training data, the learning model is repeatedly retrained. The difference between the amount of precipitable water output by the learning model and the amount of precipitable water obtained based on the GNSS signal received by the GNSS receiver is derived. If the derived difference exceeds a predetermined value, the learning model is retrained based on the continuously acquired training data. Methods for generating learning models.

2. The amount of precipitable water included in the training data is derived based on the GNSS signal received by the GNSS receiver, the orbital information of the satellite transmitting the GNSS signal, and the climate data of the location where the microwave radiometer is installed. A method for generating a learning model according to claim 1.

3. The retraining of the aforementioned learning model is performed by repeatedly retraining in parallel with the process of outputting the amount of precipitable water using the aforementioned learning model. A method for generating a learning model according to claim 1 or claim 2.

4. The retraining of the learning model based on the continuously acquired training data is performed at predetermined intervals. A method for generating a learning model according to any one of claims 1 to 3.

5. The training data further includes climate data, which includes at least one of temperature, humidity, atmospheric pressure, and rainfall at the location where the microwave radiometer is installed. Based on the acquired training data, a learning model is generated that outputs precipitable water when microwave measurement data and climate data are input. A method for generating a learning model according to any one of claims 1 to 4.

6. The aforementioned microwave measurement data represents the radio wave intensity at each of multiple frequencies. A method for generating a learning model according to any one of claims 1 to 5.

7. The radio wave intensity at each of the multiple frequencies included in the microwave measurement data has undergone dimensionality reduction processing. A method for generating a learning model according to claim 6.

8. The microwave measurement data includes the difference between the microwave measurement value from above and the microwave measurement value from the reference radio wave absorber. A method for generating a learning model according to claim 6 or claim 7.

9. Microwave measurement data, which is data obtained by measuring microwaves emitted from the atmosphere using a microwave radiometer, and training data including precipitable water amount obtained based on the GNSS signal received by the GNSS receiver installed on the microwave radiometer are acquired. Based on the acquired training data, a learning model is generated that outputs the amount of precipitable water when microwave measurement data is input. After generating the aforementioned learning model, the acquisition of training data including microwave measurement data measured by the microwave radiometer and precipitable water amount obtained from the GNSS signal received by the GNSS receiver is continued. Based on the continuously acquired training data, the learning model is repeatedly retrained. The difference between the amount of precipitable water output by the learning model and the amount of precipitable water obtained based on the GNSS signal received by the GNSS receiver is derived. If the derived difference exceeds a predetermined value, the learning model is retrained based on the continuously acquired training data. A program that instructs a computer to perform a process.

10. The retraining history information of the aforementioned learning model is generated, The generated history information is output. The program according to claim 9.

11. Multiple microwave measurement data measured at multiple time points are acquired by the aforementioned microwave radiometer. Based on the amount of precipitable water at each of the multiple time points output by the aforementioned learning model, information on changes in the amount of precipitable water is generated. The generated change information for the amount of precipitable water will be output. The program according to claim 9 or claim 10.

12. A microwave measurement unit that measures microwave measurement data, which is data obtained by measuring microwaves emitted from the atmosphere, A GNSS receiver that receives GNSS signals, A learning model that outputs precipitable water when microwave measurement data is input, The system includes a relearning unit that relearns the aforementioned learning model. The aforementioned relearning unit, After generating the learning model, the acquisition of training data including microwave measurement data measured by the microwave measurement unit and precipitable water amount obtained from the GNSS signal received by the GNSS receiver is continued. Based on the continuously acquired training data, the learning model is repeatedly retrained. The difference between the amount of precipitable water output by the learning model and the amount of precipitable water obtained based on the GNSS signal received by the GNSS receiver is derived. If the derived difference exceeds a predetermined value, the learning model is retrained based on the continuously acquired training data. Microwave radiometer.

13. to the computer Microwave measurement data, which is data obtained by measuring microwaves emitted from the atmosphere using a microwave radiometer, A learning model that outputs the amount of precipitable water when microwave measurement data is input is used, and the microwave measurement data measured by the acquired microwave radiometer is input, and the amount of precipitable water is output. After generating the aforementioned learning model, the acquisition of training data including microwave measurement data measured by the microwave radiometer and precipitable water amount obtained based on the GNSS signal received by the GNSS receiver installed on the microwave radiometer is continued. Based on the continuously acquired training data, the learning model is repeatedly retrained. The difference between the amount of precipitable water output by the learning model and the amount of precipitable water obtained based on the GNSS signal received by the GNSS receiver is derived. If the derived difference exceeds a predetermined value, the learning model is retrained based on the continuously acquired training data. An estimation method for having a computer perform a process.

Images