Water quality data prediction device, method, and program

Abstract

Predict environmental factors that cause diseases in fish and shellfish over the long term with high accuracy. [Solution] An environmental factor prediction device comprising: a predictor that uses water quality data, including values ​​corresponding to the amount of pathogens that cause diseases in fish and shellfish, and meteorological data as explanatory variables, and outputs estimated values ​​for each item of the explanatory variables after a unit time from the time series data of the explanatory variables; and a prediction means that predicts the water quality data up to N unit times (N is an integer greater than or equal to 2) by repeating predictions using the estimated values ​​by the predictor as input again to the predictor.

Description

[Technical field]

[0001] The present invention relates to a technique for predicting environmental factors, and more particularly to a technique for predicting aquatic environmental factors related to the occurrence of red tides, blue tides, blue pollen, and diseases of fish and shellfish. [Background technology]

[0002] Red tides (abnormal proliferation of plankton and bacteria) cause great damage to the fishing industry, so attempts have been made to develop methods for predicting red tides. In recent years, in particular, with the development of computing devices, simulations, artificial intelligence (AI), and IoT-related technologies, various methods for predicting red tides have been proposed.

[0003] The method in Non-Patent Document 1 creates environmental factor suitability index models for water quality and meteorological observation data observed in Ise Bay in real time, calculates a habitat suitability index from the product of these, and predicts red tides. This method predicts red tides one day in advance, but it is generally desirable to predict their occurrence three days or more in advance. Furthermore, this method has a hit rate and prediction rate of 59.4% and 69.5%, respectively, which cannot be said to be sufficient.

[0004] The method of Non-Patent Document 2 predicts red tide biomass by machine learning combining linear and non-linear analysis, suggesting that more accurate predictions than before are now possible. In addition, Patent Document 2 performs analysis that deals only with marine information. In addition, in general, chlorophyll concentration is often the prediction target in predicting environmental factors, but it appears that the prediction accuracy of chlorophyll concentration in Patent Document 2 is insufficient. Furthermore, there is a problem in that the relationship and definition of the alternative prediction target red tide biomass and red tide occurrence are unclear.

[0005] The method of Non-Patent Document 3 predicts future chlorophyll concentrations by applying a chaos recurrent neural network to time-series data of chlorophyll concentrations. However, sufficient prediction accuracy is not obtained.

[0006] In addition to red tides, it is desirable to also predict underwater environmental issues such as blue tides, blue pollen, and diseases of fish and shellfish. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Tanaka, Y; Sugimoto, Y. Development of a phytoplankton bloom prediction system using an automated water quality observation device in Ise Bay. Journal of the Japan Society of Civil Engineers, B3 (Marine Development), 2016, 72.2: I_970-I_975. [Non-Patent Document 2] QIN, Mengjiao; LI, Zhihang; DU, Zhenhong. Red tide time series forecasting by combining ARIMA and deep belief network. Knowledge-Based Systems, 2017, 125: 39-52. [Non-Patent Document 3] Masayoshi Harada, Akifumi Douma, Kazuaki Hiramatsu, Atsushi Marui, Short-term prediction of chlorophyll-a time series using chaos recurrent neural network, Lectures at the 2012 Annual Meeting of the Japanese Society of Irrigation, Drainage and Forestry Engineering, September 2012. Summary of the Invention [Problem to be solved by the invention]

[0008] The object of the present invention is to predict, over the long term and with high accuracy, environmental factors that cause the occurrence of red tides, blue tides, blue pollen, and diseases of fish and shellfish. [Means for solving the problem]

[0009] An environmental factor prediction device according to one aspect of the present invention comprises: Water quality data including values ​​according to the amount of biopigment or bioluminescence in a plurality of layers of water, water temperature, salinity, dissolved oxygen, turbidity, and flow rate; Weather data including temperature, precipitation, and sunshine hours; a predictor that uses the explanatory variables as explanatory variables and outputs estimated values ​​of each item of the explanatory variables after a unit time from the time series data of the explanatory variables; a prediction means for repeating prediction using the estimated value by the predictor as an input to the predictor again, thereby predicting the water quality data up to N unit times (N is an integer of 2 or more); The present invention is characterized by comprising:

[0010] In this disclosure, red tide refers to a phenomenon in which the color of the water changes significantly due to the abnormal proliferation of microorganisms living in seawater, particularly photosynthetic microorganisms and chemosynthetic microorganisms such as plankton and bacteria, and includes white tide and green tide in addition to red tide in the narrow sense. In this disclosure, blue-green algae refers to a phenomenon in which microorganisms living in freshwater, particularly microalgae, abnormally proliferate. In this disclosure, blue tide refers to a phenomenon in which anoxic water masses rise to the surface of the water due to the decomposition of the dead bodies of plankton and bacteria that have proliferated in large quantities. In this disclosure, plankton includes both phytoplankton and zooplankton. Note that chemosynthetic microorganisms include heterotrophic microorganisms such as noctiluca. In this disclosure, seafood diseases are classified into bacteria such as Vibrio, viruses such as carp herpes, and other protozoa, and refer to those that can be pathogens for seafood.

[0011] Values ​​corresponding to the amount of biopigments and the amount of bioluminescence include concentration, absorbance, and fluorescence intensity. Examples of biopigments include chlorophyll, carotenes, xanthophylls (lutein, fucoxanthin, etc.), and phycobilins (phycocyanin, phycoerythrin, etc.). An example of bioluminescence is luminescence caused by the chemical reaction of luciferin and luciferase. By measuring absorbance, it is possible to qualitatively and quantitatively determine the total amount of bacteria, viruses, protozoa, etc. in the water. The amount of biopigments and the amount of bioluminescence are examples of environmental factors that cause red tides, blue tides, and blue pollen.

[0012] In addition, viruses that infect species that form red tides (e.g., HaRNAV (Heterosigma akashio) Viruses), HcRNAV, HcDNAV, etc., have been confirmed to be involved in the end of red tides and other phenomena. Therefore, the above virus is one of the factors that can cause red tides and other such outbreaks.

[0013] The water quality data includes, for example, data on the above items for three layers of water: upper, middle, and lower. However, the water quality data may also include data on the above items for two layers or four or more layers of water. The water quality data may be data on seawater or freshwater, depending on the purpose of the prediction.

[0014] The predictor outputs an estimated value after a unit time for each item of the input time series data. The unit time may be appropriately determined according to the system requirements, and may be, for example, 1 hour, 6 hours, or 1 day (24 hours). The above-mentioned N may be any value as long as it is 2 or more, but may be a value such that the N unit time is, for example, 3 days or more, more preferably 7 days or more, and even more preferably 30 days or more. When there is a missing explanatory variable, the estimated value can be calculated by performing missing value interpolation or the like.

[0015] According to the above configuration, the predictor outputs an estimated value of each item after a unit time for all the explanatory variables input, so that the estimated value can be recursively fed to the predictor to obtain an estimated value after the next unit time. By repeating this recursive prediction, the environmental parameters including the chlorophyll concentration can be obtained. It is possible to predict environmental factors over the long term with high accuracy.

[0016] The environmental factor prediction device in this embodiment may further include a forecasting means for forecasting the occurrence of red tide, blue tide, or blue pollen based on the chlorophyll concentration predicted by the forecasting means. The forecasting means may forecast the start or end of a red tide, or both. The forecasting means may forecast the start or end of a blue tide, or both. The forecasting means may forecast the start or end of a blue pollen, or both. The occurrence and end of a red tide, blue tide, or blue pollen may be determined using, for example, a chlorophyll concentration judgment standard value published by a public institution such as the Tokyo Metropolitan Government Bureau of Environment, a Fisheries Research and Education Agency, or other research institute such as a university, or a private company. In addition, judgment standard values ​​according to plankton species and bacteria species are also published, so when making predictions using microbial data as described later, the occurrence and end of a red tide, etc. may be determined using a judgment standard value according to a dominant plankton species or a dominant bacteria species.

[0017] The predictor in this embodiment may further receive sampling data obtained by sampling the sample water as explanatory variables, and predict the microbial data after a unit time. An example of the sampling data is microbial data on microorganisms contained in the water. An example of the microbial data is plankton (18S rRNA Examples of sampling data include qualitative and quantitative (ratio) data of microorganisms (16S rRNA gene region, 18S rRNA gene sequence) and bacteria (16S rRNA gene region, 16SrRNA gene sequence), and qualitative and quantitative (number of individuals) data of algae at the visual level using a microscope. Another example of sampling data is qualitative or quantitative organic and inorganic data of organic and inorganic matter in the hydrosphere using NMR analysis or ICP analysis. By using microbial data in the sampling data, forecasting tools can also predict the dominant plankton species and dominant bacterial species in red tides, blue tides, or blue-green algae that will occur.

[0018] The prediction means in this embodiment may correct the estimated values ​​of weather data or water quality data obtained from the predictor based on weather forecast data or water quality forecast data obtained from a simulation different from that of the predictor, and then input the corrected values ​​to the predictor.

[0019] In this aspect, the meteorological data may include wind speed. Since wind speed affects the underwater flow speed, taking wind speed into consideration allows for more accurate prediction.

[0020] In this embodiment, the predictor may be trained by machine learning. Examples of machine learning include Simple RNN, LSTM (Long Short-Term Memory), and GRU (Gated Recurrent Unit). Examples of machine learning algorithms include RNNs (recurrent neural networks). The type of the predictor is not limited as long as it can output an estimated value of each item after a unit time from the time series data of the explanatory variables described above and predict the environmental factor intended as the effect of the present invention. Examples of the predictor other than the recurrent neural network include reinforcement learning, but are not limited to this. In addition, the trained model may be reconstructed by transfer learning. In addition, the predictor does not have to be trained by machine learning, and may be, for example, a normal simulator.

[0021] The present invention can also be understood as an environmental factor prediction method in which the above-mentioned process is executed by a computer. That is, another aspect of the present invention is an environmental factor prediction method executed by a computer, comprising: Water quality data including values ​​according to the amount of biopigment or bioluminescence in a plurality of layers of water, water temperature, salinity, dissolved oxygen, turbidity, and flow rate; Weather data including temperature, precipitation, and sunshine hours; The first step is to obtain time series data of the explanatory variables. Using a predictor that outputs an estimated value of each item of the explanatory variable after a unit time from the time series data of the explanatory variable, The second step is to obtain estimates for each of the explanatory variables; a third step of predicting the water quality data up to N unit hours by repeating the process of the second step, in which the estimated value obtained in the second step is input again to the predictor; The present invention is characterized by comprising:

[0022] The present invention can also be seen as a program for causing a computer to execute the above method. The present invention can also be seen as a trained model for performing the above method. The present invention can also be seen as a computer-readable storage medium storing the program or trained model. Effect of the Invention

[0023] According to the present invention, environmental factors that cause red tides, blue tides, blue pollen, and diseases of fish and shellfish can be predicted over the long term with high accuracy. [Brief description of the drawings]

[0024] [Figure 1] FIG. 2 is a diagram illustrating an overview of prediction in the present embodiment. [Diagram 2] FIG. 2 is a functional block diagram of a learning device for learning a predictor. [Diagram 3] FIG. 1 is a functional block diagram of a red tide prediction device according to one embodiment. [Figure 4] 1 is a flowchart of a red tide prediction process in a red tide prediction device according to one embodiment. [Diagram 5] FIG. 11 is a functional block diagram of a red tide prediction device according to another embodiment. [Figure 6] 13 is a flowchart of a red tide prediction process in a red tide prediction device according to another embodiment. [Figure 7] FIG. 13 is a diagram for explaining a learning result of a predictor according to an embodiment (analysis example 1). [Figure 8] FIG. 13 is a diagram for explaining the results of long-term prediction using a predictor according to the embodiment (analysis example 1). [Figure 9] FIG. 13 is a diagram showing another example of the learning result and the long-term prediction result of the predictor according to the embodiment (analysis example 2). [Figure 10] FIG. 13 is a diagram showing ocean sampling data with missing value interpolation (Analysis Example 3). [Figure 11] FIG. 13 is a diagram showing another example of the learning result and the long-term prediction result of the predictor according to the embodiment (analysis example 3). [Figure 12] FIG. 13 is a diagram showing the results of a comparative evaluation of missing value interpolation methods. [Figure 13] FIG. 13 is a diagram illustrating a prediction result by a predictor generated using transfer learning. [Figure 14] FIG. 13 is a diagram for explaining the results of long-term forecasting performed using data obtained by external simulation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0025] (Embodiment 1) Hereinafter, an embodiment of the present invention will be described with reference to the drawings, but the present invention is not limited thereto. The components of each embodiment described below can be combined as appropriate. In addition, in the following embodiments, seawater data is used as the target for predicting the occurrence of red tides. Seawater data is also used as the target for predicting the occurrence of blue tides. Freshwater data can be used as the target for predicting the occurrence of blue-green algae.

[0026] Known species that cause red tides include plankton belonging to diatoms, raphidophytes, dinoflagellates, cryptophytes, ciliates, etc. All of these contain large amounts of chlorophyll a and chlorophyll c. In this embodiment, the concentrations of chlorophyll a and / or chlorophyll c among chlorophyll are selected as explanatory variables to forecast the start or end of a red tide.

[0027] It is known that the dominant species in red tides are often diatom species such as Skeletonema, Thalassiosira, Eucampia, Rhizosolenia, and Chaetoceros; raphidophyte species such as Heterosigma and Chattonella; dinoflagellates such as Prorocentrum, Ceratium, Noctiluca, and Karenia; cryptophyte species such as Chloomonas; and ciliate species such as Mesodinium.

[0028] In this embodiment, in addition to the concentrations of chlorophyll a and / or chlorophyll c, etc., as explanatory variables for predicting environmental factors, qualitative and quantitative (ratio) data of 18S rRNA gene sequences specific to eukaryotes such as diatoms, raphidophytes, dinoflagellates, and / or cryptophytes as microbial data are used as additional explanatory variables to forecast the dominant plankton species and dominant bacterial species in red tides that are predicted to occur.

[0029] Blue-green algae occur when blue-green algae (cyanobacteria, cyanobacteria) or green algae (Chlorella, Chlamydomonas, etc.) become dominant. When blue-green algae are dominant, it is thought that there will be a lot of chlorophyll d and chlorophyll f in the environment. Therefore, by selecting the concentration of chlorophyll d and / or chlorophyll f among chlorophyll as an explanatory variable for predicting environmental factors, it is possible to forecast the start or end of blue-green algae.

[0030] In addition to the concentrations of chlorophyll d and chlorophyll f, microbial data such as qualitative and quantitative (ratio) data of 16S rRNA gene sequences specific to cyanobacteria (prokaryotes) can be used as explanatory variables to predict the bacterial species that are likely to cause blue blooms.

[0031] Some species of cyanobacteria, such as Trichodesmium, known as colonial blue-green algae (Aiacacios), can cause red tides. Therefore, the concentrations of chlorophyll d and / or chlorophyll f may be selected as explanatory variables to forecast the onset or end of red tides.

[0032] In addition to the concentrations of chlorophyll d and chlorophyll f, the qualitative and quantitative (ratio) data of the 16S rRNA gene sequence specific to cyanobacteria can be used as explanatory variables to predict the bacterial species that will cause red tides that are predicted to occur.

[0033] In addition to the concentrations of chlorophyll d and chlorophyll f, the quantitative values ​​of viruses (e.g., HaRNAV (Heterosigma acaciovirus), HcRNAV, HcDNAV, etc.) that infect red tide-forming species and are involved in the termination phenomenon can be used as explanatory variables to forecast the end of red tides. The quantitative values ​​of these viruses are obtained by measuring absorbance, etc., and therefore correspond to values ​​that correspond to the amount of vital pigments or the amount of vital luminescence.

[0034] <Overview> Below, we will take the long-term forecast of red tide occurrence as an example. When making long-term forecasts of red tide occurrences, it is important to predict what will happen one day or one hour from now, rather than directly predicting what will happen one week or one month from now. This is because the most recent phenomenon has a strong influence on the next phenomenon, so the further away the prediction time is, the lower the prediction accuracy becomes. Therefore, it is essential to establish a method for accurately predicting the immediate future phenomenon. If this can be done, it will be possible to make accurate long-term predictions by applying the short-term predicted values ​​to the prediction model again and repeating the process of calculating the next predicted values.

[0035] In the present invention, as shown in Fig. 1, predicted values ​​of each item after a unit time are estimated by regression prediction using a predictor from observed time series data of water quality and weather, including chlorophyll-a concentration. The predictor obtains estimated values ​​of each item one day later, for example, using chlorophyll-a concentration, water temperature, air temperature, etc. for three days as input data in daily units. By recursively inputting these estimated values ​​into the predictor, long-term prediction of chlorophyll-a concentration can be made. Then, it is possible to predict the start and end of red tide occurrence from the long-term prediction of chlorophyll-a concentration.

[0036] <Predictor training> 2 shows the configuration of a learning device 10 for learning a predictor. The learning device 10 includes, as its functional units, a learning data acquisition unit 11, a preprocessing unit 12, and a learning unit 13. The learning device 10 is a computer (information processing device) including an arithmetic processor, a storage device, an input device, an output device, a communication device, etc., and these functions are realized by the arithmetic processor executing a program.

[0037] The learning data acquisition unit 11 acquires learning data used for learning the predictor 15. The explanatory variables used as the learning data are roughly divided into ocean data (water quality data) and meteorological data.

[0038] The oceanographic data includes chlorophyll a concentration, water temperature, salinity, dissolved oxygen, turbidity, and flow velocity for the upper, middle, and lower layers of the ocean. The oceanographic data may further include pH. Here, the chlorophyll a concentration is used as the chlorophyll concentration, but other chlorophyll concentrations such as chlorophyll b concentration, chlorophyll c concentration, chlorophyll d concentration, chlorophyll e concentration, and chlorophyll f concentration may be used instead of or in addition to this. Concentrations of bacteriochlorophyll a, b, c, d, e, f, g, etc. may also be used. Chlorophyll is an example of a biological pigment, and the concentrations of other biological pigments, such as carotenes, xanthophylls (such as lutein and fucoxanthin), and phycobilins (such as phycocyanin and phycoerythrin), may also be used. Also, absorbance or fluorescence may be used as an explanatory variable instead of chlorophyll concentration. The same applies to other biological pigments. Also, bioluminescence may be used as an explanatory variable instead of or in addition to chlorophyll concentration. For example, the amount of light emitted by the luciferin-luciferase chemical reaction of the Noctiluca spp. can be obtained by direct measurement using seawater samples or sensors in seawater, or by remote measurement using hyperspectral sensors using satellites. In this case, the horizontal distance upstream or downstream relative to the direction of the water flow can also be a predictor.

[0039] The weather data includes temperature, precipitation, and sunshine hours. The weather data may also include air pressure, wind speed, humidity, and cloud cover.

[0040] These oceanographic and meteorological data are periodically observed and published by public institutions such as the Japan Meteorological Agency and the Tokyo Bay Environmental Information Center, and the like, so it is sufficient for the learning data acquisition unit 11 to acquire these observed values. However, data published by other institutions such as private companies, or data observed independently may also be used.

[0041] In addition to the above, microbial data from experiments using marine sampled samples (seawater) may be added to the training data. For example, plankton (18S rRNA gene region, 18S rRNA gene sequence) and bacteria (16S rRNA gene region, 16S rRNA gene sequence qualitative and quantitative (ratio)) data from PCR amplicon sequencing technology of seawater may be used. ) data may be added. Qualitative and quantitative data on marine organic and inorganic matter from NMR analysis or ICP analysis of seawater may also be added. Qualitative and quantitative (number of individuals) data on algae at the visual level using a microscope may also be added.

[0042] For ocean data and meteorological data, forecast data obtained by meteorological and oceanic simulations may be used. For example, atmospheric pressure, wind speed, wind volume, wind direction, temperature, humidity, precipitation, cloud cover, solar radiation (downward shortwave radiation), infrared radiation (downward longwave radiation), etc., obtained using a Global Numerical Weather Forecast Model (GSM) and targeting the atmosphere of the entire earth may be used. The above data, which are obtained using a Mesoscale Numerical Weather Forecast Model (MSM) and targeting the atmosphere of the forecast target area, instead of a balloon-based numerical weather forecast model, may be used. In addition, ocean currents, water temperature, salinity, etc., obtained using an ocean general circulation model (e.g., Regional Ocean Modeling System (ROMS) and the like) and targeting the seawater of the entire earth may be used. Forecast data obtained using a model that couples the atmosphere and ocean may be used.

[0043] In addition, chlorophyll concentration, water temperature, salinity concentration, dissolved oxygen amount, turbidity, flow velocity, etc. obtained using a river inflow model may also be used.

[0044] The pre-processing unit 12 performs outlier processing, missing value interpolation, data integration, and normalization processing. The pre-processing unit 12 removes or interpolates outliers and missing values ​​as necessary. For the interpolation, values ​​using the k-nearest neighbor method, MissForest, median, average value, etc. may be adopted. Missing value interpolation is particularly effective when the data has periodicity or little fluctuation. Normalization is a process that sets the maximum value to 1 and the minimum value to 0. The pre-processing unit 12 may perform discretization processing that converts continuous data into discrete data, and data compression processing (eigenvalue decomposition) as necessary.

[0045] The learning unit 13 learns a predictor (prediction model) 15 that obtains an estimated value of each explanatory variable included in the input data one day later from the observed time series data for a predetermined number of days (for example, three days). Since prediction is made based on the time series data, a recurrent neural network (RNN), specifically, a simple RNN, a long short-term memory (LSTM), a gated recurrent unit (GRU), or the like, can be used for the predictor 15. The type of machine learning algorithm is not limited as long as it is capable of outputting an estimated value of each item after a unit time from the time series data of the explanatory variables and predicting the environmental factors intended as the effect of this embodiment. Examples of the machine learning algorithm other than the recurrent neural network include reinforcement learning, but are not limited to this. The trained model may be reconstructed by transfer learning. The predictor does not have to be trained by machine learning, and may be, for example, a normal simulator.

[0046] <Red tide occurrence forecast> 3 shows the configuration of one embodiment of a red tide prediction device 20 for predicting the occurrence of a red tide. The red tide prediction device 20 includes, as its functional units, an input data acquisition unit 21, a long-term prediction unit 22, a predictor 23, and a forecast unit 24. The red tide prediction device 20 is a computer (information processing device) including an arithmetic processor, a storage device, an input device, an output device, a communication device, etc., and these functions are realized by the arithmetic processor executing a program.

[0047] Fig. 4 is a flowchart showing the flow of the red tide prediction process performed by the red tide prediction device 20. The red tide prediction device 20 will be described below with reference to Fig. 3 and Fig. 4. In this configuration example, it is assumed that the predictor 23 has been trained based on marine observation data and meteorological observation data.

[0048] First, in step S101, the input data acquisition unit 21 acquires observation data for the most recent A days (predetermined number of days) at the location where red tide occurrence prediction is to be performed. Here, three days (A=3) of marine observation data 25 and meteorological observation data 26 are acquired.

[0049] In loop L1, the long-term prediction unit 22 inputs data for the most recent A days to the predictor 23. The data for the next day is predicted by using A, and this is then passed back to the predictor 23 repeatedly to predict data up to the Nth day. Specifically, with A=3, data for the Tth day is predicted from observed data for days T-1 to T-3, data for the T+1th day is predicted from the predicted data for the Tth day and the observed data for days T-1 and T-2, and data for the T+2th day is predicted from the predicted data for days T+1 and T and the observed data for day T-1. By repeating this process, data predictions for up to the Nth day can be performed. Here, N may be any value as long as it is an integer equal to or greater than 2, but for example, N can be set to 30.

[0050] When the prediction up to the Nth day is completed, in step S103, the forecasting unit 24 predicts the start and end of the red tide based on the chlorophyll-a concentration up to N days later. The occurrence and end of a red tide can be determined using, for example, a judgment standard value based on the chlorophyll-a concentration published by a public institution such as the Tokyo Metropolitan Government's Bureau of Environment.

[0051] Fig. 5 shows the configuration of a red tide prediction device 30 according to another embodiment. Fig. 6 is a flowchart showing the flow of red tide prediction processing performed by the red tide prediction device 30 in this configuration example. The same components and processing as those described above are given the same reference numerals.

[0052] The difference from the above-mentioned configuration (FIG. 3) is that red tide occurrence prediction is performed using microbial sampling data 27 and weather forecast data 28 as well. In this configuration example, it is assumed that the predictor 23 has been trained based on oceanographic observation data, meteorological observation data, microbial sampling data, and weather forecast data.

[0053] In step S101, the input data acquisition unit 21 acquires A days' worth of oceanographic observation data 25, meteorological observation data 26, and microorganism sampling data 27. In step S201, the input data acquisition unit 21 also acquires weather forecast data for up to N days in the future. The weather forecast data is, for example, simulation data of atmospheric pressure, wind speed, wind volume, wind direction, temperature, humidity, precipitation, cloud cover, solar radiation, etc. for the atmosphere of the entire earth using a Global Numerical Weather Forecast Model (GSM).

[0054] In the processing of loop L1, the long-term prediction unit 22 inputs the ocean observation data, meteorological observation data, and sampling data to the predictor 23 to predict data one day later, and then recursively inputs the data to the predictor 23. However, in step S202, the long-term prediction unit 22 corrects the meteorological data predicted by the predictor 23 based on the acquired meteorological forecast data, and then recursively inputs the corrected data to the predictor 23. The specific correction method is arbitrary, and it is assumed that, for example, a simple average or a weighted average is taken. In this way, the predicted value of the meteorological data is corrected and then recursively input to the predictor 23, thereby enabling more accurate prediction. Note that, for the ocean observation data (water quality data) 25, the ocean data (water quality data) predicted by the predictor 23 may be corrected using prediction data based on a simulation, and then recursively input to the predictor 23.

[0055] In step S103 for N days from now, the forecasting unit 24 predicts the start and end of the red tide based on the chlorophyll-a concentration for N days from now. Furthermore, the forecasting unit 24 predicts the dominant plankton species and dominant bacterial species in the red tide that will occur based on the predicted values ​​of the microbial sampling data for the time when the red tide is predicted to occur. The forecasting unit 24 may predict the start and end of the red tide using a criterion according to the dominant plankton species and dominant bacterial species (for example, 500 cells / ml or more when Karenia species is the dominant species). In this configuration example, since the microbial sampling data is also used, it is possible to predict the dominant plankton species and dominant bacterial species and predict the occurrence of a red tide using a criterion according to the dominant plankton species and dominant bacterial species in this way.

[0056] <Analysis example 1> As an example of red tide prediction, we attempted to predict chlorophyll-a concentration as an indicator of red tide occurrence in the inner Tokyo Bay (off the coast of Urayasu). For real-time meteorological observation data, we obtained hourly sunshine hours and precipitation at the Edogawa critical level from the Japan Meteorological Agency's website, and calculated the daily average and total amount. For real-time oceanographic observation data, we obtained hourly chlorophyll-a concentration, water temperature, salinity, dissolved oxygen, turbidity, flow rate (current speed) data for the upper, middle and lower ocean layers off the coast of Urayasu from the Tokyo Bay Environmental Information Center's website, as well as wind volume (wind speed) data, and calculated the daily average and total amount.

[0057] Specifically, the explanatory variables used in this embodiment are 52 items as follows: [Ocean Data] Chlorophyll a concentration (upper, middle and lower layers: daily average) Water temperature (upper, middle and lower layers: daily average) Salinity (upper, middle and lower layers: daily average value) Dissolved oxygen (upper, middle and lower layers: daily average) Turbidity (upper, middle and lower layers: daily average value) -Current velocity (upper, middle and lower layers: daily average and daily total for each component, east-west, north-south) [Weather data] Temperature (daily average) - Precipitation (daily average, daily total) Sunshine hours (daily average, daily total) Wind speed (daily average and daily total for each component of east-west-north-south)

[0058] The predictor uses the 52 items from 1 to 3 days ago as explanatory variables, and the 52 items from day 0 as the objective variables. In other words, the number of nodes in the input layer and output layer is the same, and a prediction model for all variables is created. The hidden layer has 100 nodes.

[0059] To train the predictor, we used three recurrent neural network algorithms: Simple When the predictions were made using a predictor trained using RNN, LSTM, and GRU, GRU showed the best predictions. The results are shown in Figure 7. Figure 7A shows the fluctuations in chlorophyll a concentration from the observed data. Seven years of data were used as learning data (training data), and one year of data was used as test data (validation data). Figure 7B shows the prediction results one day later by a predictor trained using GRU. The figure shows the data superimposed on the measured data. A comparison of predictions using the training data and the observed values ​​showed that they were in close agreement, indicating that training had been carried out satisfactorily. Furthermore, the mean learning error (RMSE) was 11.97 μg / L. To check for over-learning and evaluate versatility, approximately one year's worth of test data that had not been used in training was applied to the prediction model, and a prediction was made one day later. The prediction error was 15.29 μg / L, indicating that the prediction is highly versatile.

[0060] Furthermore, the predicted values ​​were recursively input into the prediction model for one day later created in this way to perform long-term prediction up to 30 days later. The results are shown in Figure 8. Figure 8A shows the fluctuations in the observation data of chlorophyll-a concentration, as in Figure 7A, but the period from the latter half of 2017 to mid-January 2018 is shown in an enlarged view. Figure 8B shows the values ​​predicted for 30 days up to the end of January 2017 by inputting three days of observation data at the end of December 2017 to perform a prediction for one day later, and then recursively inputting the predicted value for one day later into the prediction model. As shown in the figure, the predicted values ​​and observed values ​​showed good agreement for about 15 days up to mid-January 2018. This shows that this method is useful for long-term prediction.

[0061] <Analysis example 2> Next, learning and prediction were attempted without using wind speed data, using a method similar to that used in Analysis Example 1. The data used in the analysis was basically the same as that used in Analysis Example 1, but the number of explanatory variables was the same as that of the 52 The total number of items is 44, excluding eight items related to wind speed.

[0062] FIG. 9A shows the results of a prediction made by a predictor trained by a GRU algorithm using seven years of observation data as training data and the remaining one year of test data, as in the first analysis example. The figure shows the results superimposed on the data. The prediction error was 18.11 μg / L, which is lower than in Analysis Example 1, but shows that sufficient accuracy was obtained.

[0063] Figure 9B shows the results of a long-term forecast up to 30 days into the future, where predicted values ​​were recursively input into the created one-day-later forecast model. As in analysis example 1 (Figure 8B), three days' worth of observation data from the end of December 2017 was input to make a one-day-later forecast, and the one-day-later predicted value was then recursively input into the forecast model, resulting in a forecast for the 30 days up to the end of January 2017. In this way, even without using wind speed data, the increase in chlorophyll-a concentration that occurred in mid-January 2018 was predicted.

[0064] <Analysis example 3> Next, learning and prediction were performed using the method in Analysis Example 1 with additional seawater sampling data. Here, the sampling data used included seven items of microbial data contained in the sample water, such as Karenia brevis, 214 items of organic data, such as amino acids and sugars, and 20 items of inorganic data, such as nitrogen, phosphorus, and silicon. Of the 52 items of oceanographic and meteorological data, 19 items of daily total values ​​were not used, and only 33 items were used.

[0065] This analysis was conducted using three years' worth of data from Kawasaki Artificial Island. There are many missing values ​​in each piece of data, so KNN (k-nearest neighbor method) was used for interpolation. Figures 10A to 10C show, as examples, the observed data and interpolated data for dinoflagellates (marine microorganisms), glycine (organic matter), and silicon (inorganic matter). In the figures, the data marked with circles is the observed data, and the rest is the data obtained by interpolation. In this way, by performing the missing value interpolation process, the discrete data is converted into continuous data (data on a daily basis), and the oceanographic data, meteorological data, and sampling data are integrated.

[0066] Figure 11A shows the fluctuations in chlorophyll-a concentration in observational data from April 2015 to November 2018, approximately three and a half years. Of this observational data, the first three years were used as learning data (training data), and the remaining data was used as test data (validation data). Figure 11B shows the prediction results for one day later by a predictor trained using a GRU, superimposed on the observational data. The learning error was 5.93 μg / L and the prediction error was 8.21 μg / L, which shows that adding sampling data can achieve more accurate predictions.

[0067] <Consideration of missing value interpolation method> Since observational data, especially sampling data, often contain missing values, we conducted a comparative evaluation of missing value imputation methods. Here, we performed LOOCV (leave-one-out cross validation) on the daily sampling data for the 72 days mentioned above (7 microorganisms, 214 metabolites, and 20 elements).

[0068] Figure 12 is a box plot showing the error between the interpolated value and the observed value obtained by each interpolation method. The number after the underscore in the interpolation method indicates how many data points are used to calculate the nearest neighbor distance in KNN. For MatFac (Matrix Factorization), SVD, and SoftImpute, it indicates how many components the eigenvalues ​​are decomposed into to generate the interpolation model.

[0069] In Fig. 12, the methods are listed in order of average error (RMSD). The matrix interpolation methods SVD and SoftImpute have many items with large errors, while K Machine learning interpolation methods such as NN and MissForest tend to reduce errors in all items. Therefore, it is preferable to adopt a machine learning interpolation method, and we adopted KNN with K = 15 or 30. It is more preferable to use a method other than the machine learning interpolation method. It is possible to perform an interpolation with less error if it is a method between KNN_30 and random shown in FIG.

[0070] <Modification> In the above description, predictions are made using data at one-day intervals, but the time interval of the data may be longer or shorter. For example, learning and predictions may be made using data at hourly intervals.

[0071] In the above explanation, the predictor predicts data one day later using data for three days (three unit times), but data for a longer period may be used to predict data one day later. By increasing the number of days, improvement in prediction accuracy can be expected, and especially if the periodicity of the data is reflected in the learning, this contributes greatly to improvement in prediction accuracy. On the other hand, by increasing the number of days, the possibility of encountering missing value data also increases, so it is advisable to determine the number of days taking this point into consideration.

[0072] In the above embodiment, an example in which two types of input data, namely, marine observation data 25 and meteorological observation data 26, are used (FIG. 3), and an example in which four types of input data, namely, marine observation data 25, meteorological observation data 26, sampling data 27, and weather forecast data 28, are used (FIG. 5) have been described. However, three types of input data, namely, marine observation data 25, meteorological observation data 26, and sampling data 27, may be used, or three types of input data, namely, marine observation data 25, meteorological observation data 26, and weather forecast data 28, may be used. Furthermore, the sampling data 27 may include data on only one or two of microorganisms, organic matter, and inorganic matter.

[0073] In the above configuration, since regional features are not used algorithmically, it is possible to predict environmental factors of a certain region using a trained model that uses learning data (training model) from another region. In this way, by obtaining a trained model that does not use regional features, predictions for any region are possible. Note that topographical elements and characteristic climates of the region may also be added to the explanatory variables, in which case it is expected that the prediction accuracy in the target region will improve.

[0074] In the above explanation, the occurrence of red tides is predicted, but the occurrence of blue tides can also be predicted using the same configuration. Also, the method and device for predicting environmental factors may perform processing up to predicting chlorophyll-a concentration without predicting the occurrence of red tides or blue tides. The occurrence of blue-green algae may also be predicted using the same configuration as above, in which case the sample water is freshwater rather than seawater.

[0075] (Embodiment 2) This embodiment is basically similar to the first embodiment, but differs in that a predictor trained using learning data from a certain region is used to perform transfer learning and then applied to another region.

[0076] Since the first embodiment does not use geographical features as explanatory variables, a predictor trained using learning data from one region can be applied to another region. This is effective in that it is not necessary to obtain learning data at each mounting site, but applying the predictor as is may result in a decrease in prediction accuracy. Therefore, in this embodiment, a predictor suitable for the mounting site is generated using transfer learning.

[0077] In this embodiment, first, a predictor (learning model) is generated in an area where a sufficient amount of learning data can be obtained (for example, Tokyo Bay) by using the method of the first embodiment. This predictor (learning model) can also be called a pre-trained predictor (learning model). Next, learning data is prepared for the implementation site where the predictor is to be applied. The content of the learning data itself is the same as that of the first embodiment. The predictor is further trained using the learning data from the implementation site. At this time, The weights of the layers of the network close to the input may be fixed, and only the weights of the layers close to the output may be adjusted. Alternatively, only the weights of the nodes related to the geographic parameters, specifically, the weights of the nodes that are (presumably) greatly influenced by the geographic specificity, may be adjusted. For example, learning may be performed in advance using learning data from multiple locations, and only the nodes whose importance varies depending on the location may be adjusted. In this embodiment, the weight information is obtained by calculating the importance using a random forest, but it may also be obtained using other machine learning algorithms. In this embodiment, a predictor obtained by such transfer learning is used to perform on-site prediction.

[0078] Figure 13 compares predictions made when predictor A, trained using training data from Tokyo Bay without transfer learning, is applied directly to the implementation site (a tuna farm in Imari Bay), with predictions made when predictor B, trained using training data from the implementation site through transfer learning, is applied.

[0079] Graph 1300 shows the actual measured values ​​of chlorophyll-a concentration at the implementation site. Graph 1301 shows the predicted values ​​one day later using predictor A, and graph 1302 shows the predicted values ​​three days later using predictor A. For comparison, the actual measured values ​​are also shown in thin lines. As can be seen from the figure, there is a large error between the predicted values ​​and the actual measured values ​​in this example. Note that because the error in the prediction three days later is large, a prediction seven days later using predictor A was not performed.

[0080] Graphs 1311 to 1313 are graphs plotting predicted values ​​1 day later, 3 days later, and 7 days later, respectively, using predictor B trained by transfer learning. It can be seen that prediction accuracy improves by performing transfer learning. Even for predictions 7 days later, it is possible to predict chlorophyll a concentration, i.e., the occurrence and termination of red tides.

[0081] Thus, according to this embodiment, transfer learning is performed based on a predictor that has been trained in advance for an area with a large amount of training data, so that a highly accurate predictor can be easily generated even if there is little training data at the implementation site. In the first embodiment, parameters directly related to geographical factors are not used as explanatory variables, but the influence of each parameter on the occurrence of red tides may change depending on geographical influences. Transfer learning adjusts the weights between each parameter to reflect the specificity of the implementation site, improving prediction accuracy.

[0082] Furthermore, according to this embodiment, important factors for each mounting site can be extracted. The contribution rate of each parameter in the transfer-learned learning model and the pre-learned learning model may be calculated and the parameters with the highest contribution rate may be output. For example, the parameters with the highest contribution rate in Tokyo Bay are K, S, Ca, Tp, Sr, B, and NH for organic matter / inorganic matter. 3The order of contribution is -N, Na, and in the environmental physical parameters it is chlorophyll-a concentration, sunshine hours, pH, water temperature, westerly wind, easterly wind, dissolved oxygen amount, and southerly wind. On the other hand, the parameters with high contribution rates in Imari Bay are Sr, Mg, K, B, Na, S, Ca, and Li in the organic / inorganic matter category, and northerly wind, southerly wind, sunshine hours, westerly wind, chlorophyll-a concentration, easterly wind, precipitation, and water temperature in the environmental physical parameters. In this way, according to this embodiment, it is possible to determine important factors at the implementation site.

[0083] (Embodiment 3) The present embodiment is characterized in that meteorological simulation data, ocean simulation data, and river inflow simulation data are used for red tide occurrence prediction. The configuration of the red tide occurrence device is the same as that of FIG.

[0084] In this embodiment, oceanographic data and meteorological data obtained by observation and data obtained by simulation are used as explanatory variables, and oceanographic data and meteorological data one time step later are used as objective variables to make a prediction.

[0085] In the following, the terms oceanographic data and meteorological data refer to data obtained by observation or predicted by the predictor 23, and data obtained by a simulation other than the predictor 23 is called simulation data.

[0086] First, the input data acquisition unit 21 acquires the most recent predetermined time step's worth of ocean observation data 25 and meteorological observation data 26. The input data acquisition unit 21 also acquires simulation data for the most recent predetermined time step and for the time step that is the target of the long-term forecast.

[0087] The predictor 23 receives the marine data and meteorological data for the most recent predetermined time step, and the simulation data for the most recent predetermined time step and the time step targeted by the long-term forecast, and predicts the marine data and meteorological data one time step later. The long-term prediction unit 22 uses the predicted marine data and meteorological data to predict the marine data and meteorological data one time step at a time up to the next predetermined time step. Note that the input to the predictor 23 may be the marine data, meteorological data, and simulation data for the most recent predetermined time step.

[0088] The predictor 23 may be trained by the same method as in the first embodiment, except that simulation data is used as training data in addition to observed data.

[0089] Below, we will explain using a specific example of a prediction made for Tokyo Bay.

[0090] In this example, the oceanographic data used are chlorophyll a concentration, water temperature, salinity, dissolved oxygen (DO), turbidity, and current speed (east-west component, north-south component) for the upper, middle, and lower layers (21 items).The meteorological data used are temperature, wind speed (east-west component, north-south component), precipitation, and sunshine hours (5 items).

[0091] The ocean simulation data used are phytoplankton amount, water temperature, salinity, east-west current velocity, north-south current velocity, zooplankton amount, ammonia nitrogen, and nitrate nitrogen (8 items). The meteorological simulation data used are air temperature (2m above sea level), precipitation (accumulated precipitation for 1 hour from that time), downward shortwave radiation, and downward longwave radiation (4 items). The river inflow simulation data used are inflows from the Koito River, Obito River, Tama River, Tsurumi River, Arakawa River, Sumida River, Edogawa River, and Hanamigawa River (8 items). ROMS was used for the ocean simulation, and MSM for the meteorological simulation.

[0092] In this example, six hours is one time step, and observation data and simulation data are acquired every six hours, and the predictor 23 predicts the ocean data and meteorological data of the next time step from the most recent six time steps (36 hours). The predictor 23 uses a CNN-QRNN algorithm for learning. The long-term prediction unit 22 repeats the prediction by the predictor 23 six times to predict the ocean data and meteorological data up to six time steps later (36 hours later).

[0093] Fig. 14 shows the actual measured values ​​of chlorophyll a concentration in Tokyo Bay and the predicted values ​​using the predictor trained in this example. Graph 1400 shows the actual measured values ​​of chlorophyll a concentration. Graphs 1401 to 1406 show the predicted values ​​after 6, 12, 18, 24, 30, and 36 hours, respectively. For comparison, the actual measured values ​​are also shown as thin lines.

[0094] (Embodiment 4) In the first to third embodiments, plankton and bacteria are used to predict the occurrence of red tides, blue tides, and blue pollen. The technology predicts the amount of photosynthetic and chemosynthetic microorganisms such as microbial tides. The above technology can be used to predict not only red tides, blue tides, and blue pollen, but also the occurrence of diseases in fish and shellfish. In this case, it is necessary to include parameters that represent the amount of pathogens (causative substances) that cause the disease in the ocean data (water quality data).

[0095] An example of a disease is Vibriosis. Vibriosis is a general term for infectious diseases caused by Vibrio bacteria. The amount of Vibrio bacteria can be measured by host-specific DNA sequences. If the pathogen is a protist, the amount of the pathogen can be measured by its specific DNA sequence. If the pathogen is a virus, the amount of the pathogen can be measured by a specific RNA or DNA sequence.

[0096] Furthermore, by measuring absorbance, it is possible to qualitatively and quantitatively determine biological samples such as nucleic acids composed of DNA and RNA, and proteins. This makes it possible to qualitatively and quantitatively determine the total amount of bacteria, viruses, protozoa, etc. in water. Specifically, environmental water such as seawater or lake water can be collected and purified using a filter, and then absorbance can be measured. The qualitative and quantitative values ​​of viruses obtained by measuring absorbance correspond to values ​​according to the amount of biological pigment or the amount of biological luminescence in the present invention.

[0097] This embodiment may be combined with any of the first to third embodiments. [Explanation of symbols]

[0098] 10: Learning device 11: Learning data acquisition unit 12: Preprocessing unit 13: Learning unit 20: Red tide prediction device 21: Input data acquisition unit 22: Long-term prediction unit 23: Predictor 24: Forecast unit

Claims

1. Water quality data including the amount of pathogen-specific DNA sequences in multiple layers of water, water temperature, salinity, dissolved oxygen, turbidity, and flow rate; Weather data including temperature, precipitation, and sunshine hours; are the explanatory variables, a predictor that outputs an estimated value of each item of the explanatory variable after a unit time from the time series data of the explanatory variable; a prediction means for repeating prediction by inputting the estimated value by the predictor into the predictor again, thereby predicting the water quality data up to N unit times (N is an integer of 2 or more); A water quality data prediction device comprising:

2. the predictor is obtained by training a learning model trained using first training data based on observed values ​​of water quality data and meteorological data in a first region by transfer learning using second training data based on observed values ​​of water quality data and meteorological data in a second region; The transfer learning involves adjusting weights of nodes with respect to geographic parameters. The water quality data prediction device according to claim 1.

3. the predictor outputs estimated values ​​of each item of the water quality data and the weather data after a unit time from time series data of the water quality data, the weather data, and forecast data obtained from a simulation different from that of the predictor for at least one of the water quality and the weather; The prediction means includes: Obtaining the forecast data up to N unit hours later; by repeating prediction using the estimated values ​​of the water quality data and the meteorological data by the predictor and at least a part of the forecast data as input, the water quality data is predicted for the N unit time period; The water quality data prediction device according to claim 1 or 2.

4. the prediction means corrects the estimated values ​​of the weather data or water quality data obtained from the predictor based on the weather forecast data or water quality forecast data obtained from a simulation different from that of the predictor, and then inputs the corrected values ​​to the predictor; The water quality data prediction device according to any one of claims 1 to 3.

5. The predictor is obtained by training a learning model trained using first training data based on observed values ​​of water quality data and meteorological data in a first region by transfer learning using second training data based on observed values ​​of water quality data and meteorological data in a second region. The water quality data prediction device according to any one of claims 1 to 4.

6. Water quality data including the amount of pathogen-specific DNA sequences in multiple layers of water, water temperature, salinity, dissolved oxygen, turbidity, and flow rate; Meteorological data including temperature, precipitation, and sunshine hours are used as explanatory variables. A first step of acquiring time series data of the explanatory variables; a second step of acquiring an estimated value of each item of the explanatory variable after a unit time of the time series data acquired in the first step using a predictor that outputs an estimated value of each item of the explanatory variable after a unit time from the time series data of the explanatory variable; a third step of predicting the water quality data up to N unit times (N is an integer of 2 or more) by repeating the process of the second step in which the estimated value obtained in the second step is input to the predictor again; A method for predicting water quality data, comprising:

7. A program for causing a computer to execute each step of the method according to claim 6.

8. water quality data including the amount of DNA or RNA sequences specific to a viral pathogen, water temperature, salinity, dissolved oxygen, turbidity, and flow rate in multiple layers of the water; Weather data including temperature, precipitation, and sunshine hours; are the explanatory variables, a predictor that outputs an estimated value of each item of the explanatory variable after a unit time from the time series data of the explanatory variable; a prediction means for repeating prediction by inputting the estimated value by the predictor into the predictor again, thereby predicting the water quality data up to N unit times (N is an integer of 2 or more); A water quality data prediction device comprising:

Images