Phytoplankton spatial distribution prediction method based on environmental DNA and machine learning

Abstract

The invention discloses a phytoplankton space distribution prediction method based on environment DNA and machine learning, and belongs to the technical field of ecological environment evaluation and treatment. Comprising the following steps: determining a sampling site, and collecting a phytoplankton eDNA sample; carrying out DNA extraction and treatment on the collected eDNA sample, and calculating a species abundance matrix; screening a training data set for training a prediction model according to the sampling time and the sampling site of the eDNA sample; inputting the training data set as an independent variable of a training set, inputting the species abundance matrix as a dependent variable of the training set, and training a prediction model; and inputting remote sensing image data into the prediction model to obtain a species spatial distribution diagram. Compared with the prior art, the method has the advantages that efficient and high-species-resolution spatial distribution prediction of algae species such as blue-green algae can be achieved, the phytoplankton prediction model is trained through the machine learning model, automatic operation of spatial prediction of various phytoplanktons is achieved, the arrangement density of sampling points is reduced, and the monitoring efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological environment evaluation and governance, and more specifically, to a method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning. Background Art

[0002] The outbreak of harmful algal blooms is a severe challenge faced by the global ecological environment field. Such events not only disrupt the ecological balance of water bodies but also may lead to the release of algal toxins, thereby polluting drinking water sources and posing a threat to human health. There are a wide variety of phytoplankton species that can cause algal blooms, and the common Microcystis genus in Lake Taihu is just one of them. In addition, many other phytoplankton species such as Pseudanabaena and dinoflagellates also have the ability to trigger harmful algal blooms. These different phytoplankton species vary in terms of growth conditions, reproduction rates, and the types of toxins they produce. Therefore, different treatment strategies need to be adopted for the algal blooms caused by them. For example, since the Microcystis genus does not have the ability to fix nitrogen, its growth is often limited by the nitrogen and phosphorus concentrations in the water body. Therefore, when treating the algal blooms caused by the Microcystis genus, the focus should be on controlling the nitrogen and phosphorus concentrations in the water body. On the other hand, dinoflagellate red tides usually occur in water bodies with a high organic matter content, which requires us to strengthen the monitoring and management of the organic matter content in the water body. Thus, to achieve targeted treatment of different types of algal blooms, it is first necessary to accurately identify and monitor the species of phytoplankton.

[0003] However, the current monitoring of phytoplankton relies on two methods: morphological monitoring and remote sensing monitoring. Although morphological monitoring can accurately distinguish different species, its process is cumbersome, inefficient, and highly dependent on the professional knowledge and experience of the appraisers. In addition, since morphological monitoring usually requires manual sampling, it is difficult to achieve high-spatial-resolution monitoring and cannot comprehensively and timely reflect the spatial distribution of phytoplankton. In contrast, remote sensing monitoring technology has advantages such as a wide coverage area and fast data acquisition, and can obtain the distribution of phytoplankton in water bodies over a large area. However, remote sensing monitoring also faces some challenges. On the one hand, its monitoring results are easily affected by meteorological conditions (such as cloud cover, water body reflection, etc.), resulting in unstable data quality. On the other hand, remote sensing monitoring technology cannot distinguish specific species and can only provide the total distribution information of phytoplankton, unable to provide fine species monitoring data for targeted treatment. Therefore, the current monitoring system for phytoplankton distribution cannot meet the dynamic monitoring requirements for the species and distribution of algal blooms. The current monitoring system for the spatial distribution of phytoplankton has the following obvious disadvantages: (1) Low efficiency: The traditional morphological monitoring method requires a large amount of time and manpower for sampling, identification, and analysis of phytoplankton, and cannot meet the need for timely obtaining the spatial distribution data of phytoplankton. (2) Inability to balance species resolution and spatial resolution simultaneously: Although remote sensing monitoring technology can obtain data over a large area, it cannot provide specific species information; while morphological monitoring can identify species, but is limited by the number and distribution of sampling points and cannot achieve high-spatial-resolution monitoring. (3) Limited species identification: The morphological monitoring method has high requirements for the professional knowledge of appraisers and can only compare based on the morphological characteristics of some species. For species with similar morphology and difficult to distinguish, its identification ability is limited, prone to misjudgment and missed judgment.

[0004] In related technologies, such as Chinese Patent CN112763429A, a three-dimensional monitoring method for cyanobacterial blooms is provided. By multi-dimensionally monitoring biological indicators in water, the dynamics of cyanobacterial blooms are grasped, and a three-dimensional monitoring system for cyanobacterial blooms from the water surface to the air is established, including on-site patrol surveys, video monitoring, remote sensing monitoring, etc., to analyze data and reflect the situation. This solution monitors the water quality and algal bloom phenomena through real-time multi-dimensional multi-biological indicator monitoring of remote sensing, near-surface monitoring, and field monitoring, but does not solve the obvious deficiencies of existing phytoplankton monitoring methods in terms of efficiency, species resolution, and spatial resolution, and cannot meet the requirements for dynamic monitoring and precise treatment of algal bloom species and their distribution.

[0005] As can be seen from the above, related technologies do not give any technical inspiration on how to efficiently achieve species spatial distribution monitoring with high species resolution and the drawing of spatial distribution maps. Summary of the Invention

[0006] 1. Technical Problems to be Solved Aiming at the problems existing in the prior art, such as the dependence on manual sampling and analysis for algae monitoring, the insufficient ability to integrate multi-source data, the low spatio-temporal resolution of prediction models, and the weak generalization ability, the present invention provides a method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning, which can achieve efficient and high-species-resolution spatial distribution prediction of algae species such as cyanobacteria.

[0007] 2. Technical solution The object of the present invention is achieved by the following technical solutions.

[0008] This part of the content of the present application is used to introduce the concept briefly, and these concepts will be described in detail in the following specific implementation part. This part of the content of the present application is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0009] Some embodiments of the present application propose a method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning to solve the technical problems mentioned in the above background art part.

[0010] As the first aspect of the present application, some embodiments of the present application provide a method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning, including the following steps: Determine the sampling sites and collect phytoplankton eDNA samples; Extract and process the DNA of the collected eDNA samples and calculate the species abundance matrix; According to the sampling time and sampling sites of the eDNA samples, screen the training data set for training the prediction model; Input the training data set as the independent variable of the training set and the species abundance matrix as the dependent variable of the training set to train the prediction model; Input the remote sensing image data into the prediction model to obtain the species spatial distribution map.

[0011] Furthermore, the process of collecting phytoplankton eDNA samples is specifically as follows: Use sterile containers to collect 3 bottles of 0.5L water samples at the sampling sites.

[0012] Furthermore, filter the water samples with a filter membrane with a pore size of 0.45μm to retain phytoplankton cells and released eDNA.

[0013] Furthermore, for every 10 sampling sites completed, set a group of blank controls, and each group of blank controls contains 3 repeated blank controls.

[0014] Further, the process of calculating the species abundance matrix is specifically as follows: DNA extraction is performed based on the collected eDNA samples; the extracted DNA is amplified, library constructed, and sequenced; phytoplankton species are annotated according to the sequence information obtained from the sequencing, and the corresponding species abundance matrix is calculated.

[0015] Further, for the DNA extracted from eukaryotic phytoplankton, amplification is performed using 18S rRNA V9 hypervariable region primer pairs; for the DNA extracted from Cyanophyta, amplification is performed using modified 16S rRNA V3 region primer pairs.

[0016] Further, the process of screening the training data set is specifically as follows: According to the sampling time and sampling sites of the collected eDNA, remote sensing image data with a resolution of greater than or equal to 100 meters is selected; the selected remote sensing image data is subjected to atmospheric correction and radiometric correction to obtain the training data set.

[0017] Further, cloud cover screening is performed on the selected remote sensing image data, and remote sensing image data with a cloud cover below 10% is selected to construct the training data set.

[0018] Further, the process of training the prediction model is specifically as follows: The input independent variables and dependent variables are placed in the data matrix according to the points, 70% of the data in the data matrix is used as the training set of the prediction model, and the remaining 30% of the data is used as the validation set of the prediction model; The training set of the prediction model is input into the random forest model for random forest model training; Grid tuning is performed, and the parameter with the highest R-squared is selected as the training parameter of the prediction model; The training parameters are set, and the random forest model is trained again to obtain the prediction model.

[0019] As the second aspect of the present application, some embodiments of the present application provide a system based on the above-mentioned method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning, including a collection module: determining sampling sites and collecting phytoplankton eDNA samples; A module for calculating the species abundance matrix: performing DNA extraction and processing on the collected eDNA samples and calculating the species abundance matrix; A module for screening the training data set: screening the training data set for training the prediction model according to the sampling time and sampling sites of the eDNA samples; A training module: inputting the training data set as the independent variable of the training set and the species abundance matrix as the dependent variable of the training set to train the prediction model; An output module: inputting the remote sensing image data into the prediction model to obtain the species spatial distribution map.

[0020] 3. Beneficial effects Compared with the prior art, the advantages of the present invention are as follows: By combining the high-throughput and standardized operation processes of environmental DNA technology, the present invention can greatly improve the acquisition efficiency of phytoplankton data in different regions and spaces, reduce labor input and time costs, and contribute to improving the accuracy and resolution of species discrimination. Secondly, the phytoplankton prediction model trained by the machine learning model can simultaneously realize the automatic operation of multiple phytoplankton spatial predictions through loop programming, thereby reducing the layout density of sampling points and improving the monitoring efficiency.

[0021] Through the efficient combination of environmental DNA technology and remote sensing data, the accuracy and efficiency of phytoplankton spatial distribution prediction have been significantly improved. By adopting a standardized eDNA sampling process (such as sterile sampling, blank control setting, amplification of hypervariable region primers) combined with high-throughput sequencing technology, a high-resolution species abundance matrix can be quickly obtained, overcoming the defects of low species resolution and time-consuming and laborious of traditional microscope observation; through the atmospheric correction, radiometric correction and cloud cover screening of remote sensing image data, the authenticity of the surface reflectance of the training data set is ensured, providing reliable environmental variables for model input; based on the feature importance screening and grid tuning optimization of the random forest algorithm, the sensitivity of the model to key environmental factors is significantly improved, and the interference of redundant data is reduced. In addition, by automatically generating a species spatial distribution map consistent with the resolution of remote sensing images through machine learning, large-scale waters can be covered without densely arranging sampling points, greatly reducing the monitoring cost and providing efficient technical support for algal bloom early warning and ecological management. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic flow chart of a method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning in an embodiment of the present invention; Figure 2 It is a scatter plot showing the relationship between the predicted values and the measured values of the abundances of some species in an embodiment of the present invention; Figure 3 It is a graph of evaluation indexes of models corresponding to all species at the order level in an embodiment of the present invention; Figure 4 It is a spatial distribution map of the predicted values of the abundances of some species in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The present invention will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0024] Phytoplankton are the most important basic biological groups in marine and freshwater ecosystems, having an important impact on the structure and function of the ecosystem. Accurately predicting and monitoring the spatial distribution of phytoplankton is of great significance for the protection and management of the ecosystem. In the case where traditional phytoplankton monitoring methods generally rely on means such as microscopic observation and chemical analysis, which are time-consuming and laborious and difficult to achieve real-time monitoring of large-scale areas, the purpose of predicting the spatial distribution of phytoplankton based on environmental DNA is to quickly and accurately obtain information on the species, quantity, and spatial distribution of phytoplankton, providing a scientific basis for ecosystem protection and management; revealing the relationship between phytoplankton and environmental factors and understanding the response mechanism of phytoplankton to environmental changes; and providing technical support and decision-making basis for the sustainable use of marine and freshwater resources.

[0025] Among them, eDNA, that is, environmental DNA, refers to DNA fragments present in environmental samples (such as water bodies, soil, and air, etc.). These DNA fragments may come from cell shedding, excreta, secretions, etc. of organisms and are the genetic information left by organisms in the environment. The eDNA technology can indirectly reflect the existence and distribution of biodiversity by analyzing the DNA in these environmental samples without directly observing the organisms themselves. In the prediction of the spatial distribution of phytoplankton, the eDNA technology is used to extract the DNA information of phytoplankton from environmental samples. By sequencing and analyzing this DNA information, the characteristics such as the species, quantity, and spatial distribution of phytoplankton can be determined, thus providing an important basis for phytoplankton ecological research, environmental protection, and management. As DNA fragments in environmental samples, eDNA provides the genetic information of phytoplankton.

[0026] DNA, that is, deoxyribonucleic acid, is a molecule that stores genetic information in organisms. In organisms, DNA transmits genetic information to the next generation through processes such as replication and transcription and controls processes such as the growth, development, and metabolism of organisms. In the eDNA technology, the extraction, sequencing, and analysis of DNA are the key steps to obtain biodiversity information. As a molecule that stores genetic information in organisms, DNA is the analysis object and basis of the eDNA technology.

[0027] As Figures 1 to 4 shown, a method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning of the present invention includes the following steps: Determine sampling sites and collect phytoplankton eDNA samples; Extract and process the DNA of the collected eDNA samples and calculate the species abundance matrix; According to the sampling time and sampling sites of the eDNA samples, screen the training data set for training the prediction model; Use the training dataset as the independent variable input of the training set and the species abundance matrix as the dependent variable input of the training set to train a prediction model; Input the remote sensing image data into the prediction model to obtain the species spatial distribution map.

[0028] Specifically, the specific process of the method for predicting the spatial distribution of phytoplankton based on environmental DNA is as follows: S1. Determine the sampling sites and collect phytoplankton eDNA samples.

[0029] Determine the sampling sites: Before collecting phytoplankton eDNA samples, it is necessary to select the sampling sites. The process of determining the sampling sites can ensure the representativeness and accuracy of the sampling results.

[0030] Specifically, use a geographic information system (GIS) to plan the sampling sites to ensure the representativeness and uniformity of the sampling points; determine the sampling sites according to the comprehensive factors of the study area.

[0031] The study area includes lakes and rivers, etc. The comprehensive factors of the study area include ecological characteristics, water flow dynamics, historical data, and potential environmental impact factors, etc. The study area includes lakes and rivers. The ecological characteristics of the study area include the shape, water depth, and water flow velocity of the lake or river; the water flow dynamics include seasonal changes and water flow direction, etc. The historical data includes the research results of the previous phytoplankton community structure. The potential environmental impact factors include pollution sources and nutrient input.

[0032] Collect eDNA samples: After the sampling sites are determined, collect phytoplankton eDNA (environmental DNA) samples at the sampling sites.

[0033] Use a sterile container to collect water samples and filter the water samples with a filter membrane to intercept phytoplankton cells and the eDNA they release.

[0034] Specifically, at each sampling site, use a sterile container to collect 3 bottles of 0.5 L water samples each; filter the water samples with a 0.45 μm pore size filter membrane to intercept phytoplankton cells and the eDNA they release.

[0035] During the process of filtering the water samples with the filter membrane, ensure that the water samples are evenly distributed on the filter membrane to avoid filter membrane rupture or contamination.

[0036] More specifically, the sterile container is sterilized by high temperature and high pressure before use to ensure avoiding any contamination during the collection process. The filter membrane needs to be sterilized before use and stored under sterile conditions after collection to prevent DNA degradation. The way of storing under sterile conditions can be to seal it with a sterile centrifuge tube or plastic bag.

[0037] Set up a blank control: Set up a blank control during the process of collecting eDNA samples, so as to minimize the false positive results caused by contamination or misoperation during the sampling process to the greatest extent, and improve the accuracy and reliability of the data.

[0038] Specifically, every time after sampling 10 sampling sites, set up a group of blank controls, and each group of blank controls contains 3 repeated blank controls.

[0039] In a specific embodiment, taking Poyang Lake as an example, 52 representative sites are selected as sampling sites; at each sampling site, 3 bottles of 0.5L water samples are collected using sterile containers, and the water samples are filtered through a filter membrane with a pore size of 0.45μm to intercept phytoplankton cells and the eDNA released by them.

[0040] S2. Perform DNA extraction and processing on the collected eDNA samples, and calculate the species abundance matrix.

[0041] Based on the eDNA samples collected in step S1, perform DNA extraction; and perform amplification, library construction and sequencing on the extracted DNA; annotate phytoplankton species according to the sequence information obtained by sequencing, and calculate the corresponding species abundance matrix. Specifically, during the process of DNA extraction from the collected eDNA samples, use sterile instruments and the extraction reagents corresponding to the eDNA samples to ensure the effective extraction of phytoplankton cells and the eDNA released by them from the filter membrane.

[0042] The choice of extraction reagent depends on the properties of the target compound, and the properties of the target compound include chemical properties such as polarity, non-polarity, acid-base property and thermal stability.

[0043] In a specific embodiment, the process of DNA extraction includes: Wash the filter membrane with buffer solution to remove impurities, and use physical or chemical methods to break the cell membrane to release DNA; Add extraction reagent and incubate, and remove insoluble impurities by methods such as centrifugation, filtration or precipitation; Use methods such as silica column chromatography and electrophoresis to purify DNA and remove proteins, RNA and other impurities.

[0044] Through the process of DNA extraction, it can be ensured that high-quality DNA is efficiently and safely extracted from the eDNA samples, providing a reliable basis for subsequent analysis and research.

[0045] Specifically, after completing the extraction of DNA, perform amplification and library construction on the extracted DNA.

[0046] Using the corresponding primer pairs respectively, DNA amplification was performed on eukaryotic phytoplankton and Cyanophyta; the amplified products were purified and library construction was carried out, and finally a library suitable for high-throughput sequencing was generated.

[0047] The process of library construction includes: end repair of DNA fragments, adding A tails, and ligating sequencing adapters.

[0048] Specifically, for the DNA extracted from eukaryotic phytoplankton, amplification was carried out using the 18S rRNA V9 hypervariable region primer pair. This region has high species specificity and is suitable for distinguishing different eukaryotic phytoplankton species. The 18S rRNA V9 hypervariable region primer pair is: 1380F: TCCCTGCCHTTTGTACACAC; 1510R: CCTTCYGCAGGTTCACCTAC. For the DNA extracted from Cyanophyta, amplification was carried out using a modified 16S rRNA V3 region primer pair. These primers can specifically amplify the DNA fragments of Cyanophyta, thereby achieving an accurate assessment of this important phytoplankton group. The 16S rRNA V3 region primer pair is: 341F: ACCTACGGGRSGCWGCAG; 518R: GGTDTTACCGCGGCKGCTG. Among them, eukaryotic phytoplankton refers to phytoplankton whose cells have obvious cell nuclei and other complex organelles. Eukaryotic phytoplankton mainly includes multiple species such as diatoms, calcareous algae, dinoflagellates, and euglenoids. The growth cycle of eukaryotic phytoplankton is relatively short, non-reproductive cells do not differentiate, and there are no true leaves, roots, and stems. The distribution and abundance of eukaryotic phytoplankton are affected by various environmental factors, including water temperature, pH value, dissolved oxygen, water depth, and nutrients (such as nitrate nitrogen), etc. Cyanophyta, also known as Cyanobacteria, is the most primitive group among algae plants. Cyanophyta is mostly blue-green in color, and its living habits include aquatic, aerophilic, endophytic, parasitic, symbiotic, etc.; it is mainly distributed in fresh water with more organic matter, and some live in wet soil, on rocks, tree trunks, and in the ocean, form lichens in symbiosis with fungi, or live inside plants to form endophytes; a few species can live in hot springs above 85°C or in the polar regions with perennial snow.

[0049] Specifically, the constructed library was sequenced with 150 bp paired ends to obtain the raw sequencing data; bioinformatics tools were used to perform sequence alignment and species annotation on the raw sequencing data, and phytoplankton species were annotated according to the raw sequencing data; at the same time, all species signals generated in the blank control were removed to obtain the species abundance matrix.

[0050] The species abundance matrix reflects the relative abundance and diversity of phytoplankton species at different sampling sites, providing an important basis for further ecological analysis and the formulation of conservation strategies.

[0051] Specifically, the sequencing platform can be an Illumina series sequencer, and the quality control includes removing low-quality sequences and adapter sequences.

[0052] S3. According to the sampling time and sampling sites of the eDNA samples, screen the training data set for training the prediction model.

[0053] According to the sampling time and sampling sites for collecting eDNA, select remote sensing image data; perform atmospheric correction and radiometric correction on the selected remote sensing image data to obtain the training data set.

[0054] Specifically, according to the sampling time and sampling sites during the collection of eDNA, select remote sensing image data with a resolution higher than or equal to 100 meters within the sampling area to form an initial training data set; perform atmospheric correction and radiometric correction on the initial training data set to obtain the training data set for training the prediction model. For the obtained species abundance matrix and remote sensing image data set, perform data cleaning and normalization processing.

[0055] In a specific embodiment, according to the sampling time during the collection of eDNA, including the specific date or time period, screen out the remote sensing image data within the corresponding sampling time range from the remote sensing image data; then, according to the sampling sites during the collection of eDNA, such as the longitude and latitude range, further select the remote sensing image data covering the sampling area from the screened remote sensing image data; then, further select the remote sensing image data with a resolution greater than or equal to 100 meters from the screened remote sensing image data to meet the requirement of spatial resolution. Finally, use the screened remote sensing image data to form the training data set for training the prediction model.

[0056] More specifically, perform cloud cover screening on the screened remote sensing image data, and select the remote sensing image data with a cloud cover lower than 10% as the remote sensing image data in the training data set. To reduce the influence of clouds on the surface reflectance and improve the accuracy and reliability of the remote sensing image data.

[0057] In a specific embodiment, remote sensing image data of Sentinel-2 within the sampling time period is acquired, and cloud cover screening is performed to delete remote sensing image data with cloud cover higher than 10% within the time period. The specific process includes: in Google Earth Engine, screening the remote sensing image data of Sentinel-2 according to the sampling time and sampling sites; determining the proportion of cloud cover by dividing the number of pixels with the first digit of the QA60 band data being 1 in the screened data by the total number of pixels, and then determining whether the cloud cover of this data set is greater than 10%. Finally, the data sets that do not meet the conditions are removed.

[0058] Atmospheric correction and radiometric correction are performed on the remote sensing image data in the training data set to eliminate the influence of atmospheric and illumination conditions on the remote sensing data, and obtain data closer to the true surface reflectance.

[0059] Specifically, radiometric correction includes absolute radiometric correction and relative radiometric correction. Absolute radiometric correction is to convert the DN (digital number) value of the remote sensing image data into radiance value or reflectance value; relative radiometric correction is to correct the remote sensing image data acquired by the same remote sensor at different times and different detection angles to reduce the error caused by the change of instrument performance.

[0060] The remote sensing image data after atmospheric correction and radiometric correction is converted into the true surface reflectance, and the surface reflectance is an important parameter reflecting the surface material composition and characteristics.

[0061] Specifically, the corrected remote sensing image data, that is, pixel data, is divided by 10000 to obtain the true surface reflectance, and the processed training data set is obtained.

[0062] S4. Input the training data set as the independent variable of the training set and the species abundance matrix as the dependent variable of the training set to train the prediction model.

[0063] Specifically, data cleaning and normalization processing are performed on the species abundance matrix obtained in step S2 and the training data set obtained in step S3 to remove missing values and outliers.

[0064] By performing data cleaning to remove missing values and outliers, the missing values can be filled by interpolation methods such as linear interpolation and nearest neighbor interpolation; the outliers can be identified and processed by statistical methods such as Z-score and IQR, and options include deletion or replacement with normal values.

[0065] Specifically, the min-max normalization method is used to normalize the remote sensing data so that the data varies between 0 and 1. Using the min-max normalization method means calculating the normalized value of each pixel according to the minimum and maximum values in the data set, eliminating the dimensional difference between different data, and improving the training efficiency and prediction accuracy of the model.

[0066] In a specific embodiment, the process of data cleaning and normalization for the obtained species abundance matrix and training dataset specifically includes: Replace all missing values in the species abundance matrix and training dataset with 0 to ensure normal operation of the calculation; Use the R language to calculate the min-max normalization result of the remote sensing data, and the calculation expression is:

[0067] where, is the data to be calculated, is the maximum value of the sample data, is the minimum value of the sample data.

[0068] Use the training dataset that has completed data cleaning and normalization as the independent variable input of the training set of the random forest factor screening model, and use the species abundance matrix that has completed data cleaning and normalization as the dependent variable input of the training set of the random forest factor screening model; furthermore, based on the importance of random forest features, obtain the independent variable matrix input for constructing the prediction model training set by screening factors.

[0069] In this embodiment, use random forest for pre-calculation to obtain the feature importance of different factors, so as to improve the total variance explanation rate of the input independent variable factors.

[0070] Specifically, the random forest calculation tool uses the scikit-learn package of python 3.11. For each species, first calculate the feature importance of the input factors, and then sort according to the factor importance, select the 6 factors with the highest importance, and extract the data of these 6 factors into a data matrix.

[0071] In a specific embodiment, the process of obtaining the screened factors based on the importance of random forest features includes: Input the matrix containing all factors into the random forest model, and calculate the feature importance of all factors in the random forest based on the Gini index (Gini index), that is, the statistic , represents the th variable's average change in node split impurity across all trees in the random forest. Among them, calculate the Gini index The expression is:

[0072] where, is the number of categories in the training dataset, is the node The sample belongs to the Probability estimation value of the class.

[0073] Variable At the node The importance of, that is, the node The change in the Gini index before and after branching The expression is: ; Among them, And Respectively represent the Gini indices of the two new nodes split by the node The Gini index of the two new nodes split by the node

[0074] If the variable Appears in the th tree times, then the variable In the th tree The expression of the importance of is: ; Finally, obtain the importance of the variable In the random forest , and its expression is: ; Among them, Is the number of classification trees in the random forest, Is a natural number.

[0075] Specifically, for each input factor, based on the corresponding Sort in descending order, and the 6 factors with the highest importance are used to construct the independent variable matrix for the prediction model training set.

[0076] More specifically, based on the independent variable matrix containing the selected factors, perform machine learning model training.

[0077] In a specific embodiment, the process of training the prediction model specifically includes: Place the input independent variables (training data set) and dependent variables (species abundance matrix) in the data matrix according to the point positions, use 70% of the data in the data matrix as the prediction model training set, and use the remaining 30% of the data as the prediction model validation set; Input the prediction model training set into the random forest model for random forest model training; Perform grid tuning and select the parameter with the highest R-square as the training parameter of the prediction model; After the training parameters are set, perform random forest model training again to obtain the random forest model of the target species, that is, the prediction model.

[0078] In a specific embodiment, for different species, the species abundance matrix data of different species is used and the above steps of training the prediction model are repeated, and finally the prediction models for each species are trained.

[0079] S5. Input the remote sensing image data into the prediction model to obtain the species spatial distribution map.

[0080] According to the prediction models trained for each species, the remote sensing image data of the corresponding band is used as the independent variable and input into the prediction model for calculation to obtain the species spatial distribution map with the same resolution as the remote sensing image data.

[0081] Specifically, according to the different selected satellites, the remote sensing image data contains surface reflection data of different wavelengths. The data of different wavelengths in the image data are extracted to form a training data set, which is input into the prediction model of the corresponding predicted species for calculation, and the calculated result value is the number of reads of the eDNA of the predicted species.

[0082] Perform similar calculations on all pixels in the remote sensing image data to obtain the species eDNA abundance distribution map with the same resolution as the remote sensing image data.

[0083] Specifically, in the calculation process, the numpy package of Python is used for calculation.

[0084] In a specific embodiment, taking the Poyang Lake Basin as an example, the sampling time is May 2023, and based on the phytoplankton spatial distribution prediction method of the present invention, a prediction analysis is carried out for Poyang Lake.

[0085] Collect eDNA samples for Poyang Lake; perform DNA extraction and processing, and calculate the species abundance matrix; according to the sampling time and sampling sites of the eDNA samples, screen the training data set for training the prediction model.

[0086] For each species, input the species abundance matrix as the dependent variable, obtain the Sentinel-2 L2A remote sensing image data with a cloud cover lower than 10% on the date of the eDNA collection sample, extract the data of all bands therein as the training data set, and input this training data set as the independent variable; screen the appropriate independent variable input factors, and use the random forest model to train the prediction models of the abundances of each species.

[0087] As shown in Table 1, the factor importance determined by the prediction model is presented taking Pseudanabaenales, Stephanodiscales, and Nostocales as examples. Sorted from high to low, the six factors with the highest importance are selected as input factors. For the prediction model of Nostocales, the B9, B1, B12, B3, B8, and B4 bands are used as the factor inputs of the prediction model; for the prediction model of Pseudanabaenales, the B9, B8, B10, B3, B6, and B7 bands are used as the factor inputs of the prediction model; for the prediction model of Stephanodiscales, the B8, B1, B11, B7, B12, and B6 bands are used as the factor inputs of the prediction model.

[0088] Table 1 Factor importance of the random forest models for Pseudanabaenales, Stephanodiscales, and Nostocales

[0089] Based on the above factor selection, the independent variable and dependent variable data matrices are re-input into the random forest model, and the optimal model parameters are trained by the grid search method to finally obtain the prediction model.

[0090] As Figure 2 and Figure 3 shown, the prediction accuracy of the model is presented. As Figure 2 shown, it is the evaluation index diagram of the models corresponding to all species at the order level. Among them, Synechococcales is the order of Synechococcus; Chlamydomonadales is the order of Chlamydomonas; Stephanodiscales is diatoms; Sphaeropleales is the order of Chlorococcum; Melosirales is the order of Melosira; Thalassiosirales is the order of Thalassiosira; Nostocales is the order of Nostoc; Bacillariales is the order of Bacillariophyceae; Euglenales is the order of Euglena; Surirellales is the order of Surirella; Pseudanabaenales is the order of Pseudanabaena; Chlorosarcinales is the order of Chlorosarcina. It can be seen that the true values and predicted values of most species have a good corresponding relationship. Except for a few points, there is a good corresponding relationship between the corresponding species abundances and predicted values at most points. As Figure 3As shown, it is a scatter plot of the relationship between the predicted values and the measured values of the abundances of some species. Among them, RMSE represents the root mean square error; MAPE represents the mean absolute percentage error; R square represents the R-squared value. It can be seen that for the prediction accuracy of the prediction models for different species on the test set, the root mean square error (RMSE) of most models is within 10, and the root mean square error (RMSE) of most models is effectively controlled within 10, which means that the deviation between the predicted values and the true values is generally small. At the same time, the mean absolute percentage error (MAPE) of more than three-quarters of the species prediction models does not exceed 30%, further verifying the stability and reliability of the model prediction. More importantly, the R-square values of most models are greater than 0.8, proving that the prediction models have high prediction accuracy and generalization ability. The data of these prediction models fully show that by applying the method for predicting the spatial distribution of phytoplankton based on environmental DNA of the present invention, a high-precision prediction model for multiple phytoplankton species can be obtained. The prediction results of the prediction model maintain a good correspondence with the true values, and the prediction errors are all maintained within an acceptable range.

[0091] As Figure 4 shown, by inputting remote sensing image data into the prediction models of the corresponding species, a new data matrix can be calculated, and the position of each element in the matrix corresponds to its position in the remote sensing image. With the help of this data matrix, the probability density spatial distribution of different species can be displayed by calculating the kernel density probability.

[0092] Generally speaking, the method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning of the present invention can more accurately and with high spatial resolution predict the spatial distribution of different phytoplankton species, and can intuitively draw the spatial distribution map of different phytoplankton in a large area of water bodies. At the same time, due to the use of the species information provided by eDNA, the identification of species is more accurate and efficient. By effectively combining the advantages of eDNA and remote sensing technology, the prediction of phytoplankton with high species resolution and high spatial resolution is realized, so that the distribution of phytoplankton in large lakes can be monitored and mastered more accurately and intuitively, providing scientific and effective data support for the prevention and control of water blooms and meeting the needs of ecological monitoring.

[0093] In a specific embodiment, a system for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning includes a collection module: determining sampling sites and collecting phytoplankton eDNA samples; a module for calculating the species abundance matrix: extracting and processing DNA from the collected eDNA samples and calculating the species abundance matrix; a module for screening the training data set: screening the training data set for training the prediction model according to the sampling time and sampling sites of the eDNA samples; Training module: Using the training data set as the independent variable input of the training set and the species abundance matrix as the dependent variable input of the training set to train a prediction model; Output module: Inputting remote sensing image data into the prediction model to obtain a species spatial distribution map. The system for predicting the spatial distribution of phytoplankton based on environmental DNA of the present invention can greatly improve the acquisition efficiency of phytoplankton data in different regions and different spaces by combining the high-throughput and standardized operation process of environmental DNA technology, and helps to improve the accuracy and resolution of species discrimination. The phytoplankton prediction model trained by a machine learning model can simultaneously achieve automatic calculation of multiple phytoplankton spatial predictions through loop programming, thereby reducing the layout density of sampling points and improving the monitoring efficiency.

[0094] The above schematically describes the present invention and its implementation manners. This description is not restrictive. Without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Any reference signs in the claims should not limit the claims involved. Therefore, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments to this technical solution without creative efforts without departing from the purpose of this creation, they should all fall within the protection scope of this patent. In addition, the term "comprising" does not exclude other elements or steps, and the word "a" before an element does not exclude including "a plurality of" such elements. The multiple elements stated in the product claims can also be implemented by one element through software or hardware. First, second, etc. are used to indicate names and do not indicate any specific order.

Claims

1. A method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning, characterized in that: It includes the following steps: Determine the sampling sites and collect eDNA samples of phytoplankton; Extract and process the DNA of the collected eDNA samples and calculate the species abundance matrix; According to the sampling time and sampling sites of the eDNA samples, screen the training data set for training the prediction model; Use the training data set as the independent variable input of the training set and the species abundance matrix as the dependent variable input of the training set to train the prediction model; Input the remote sensing image data into the prediction model to obtain the species spatial distribution map.

2. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 1, wherein: The process of collecting eDNA samples of phytoplankton is specifically: collect 3 bottles of 0.5L water samples each at the sampling sites using sterile containers.

3. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 2, wherein: Filter the water samples with a filter membrane with a pore size of 0.45μm to intercept phytoplankton cells and released eDNA.

4. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 2, wherein: For every 10 sampling sites completed, set a group of blank controls, and each group of blank controls contains 3 repeated blank controls.

5. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 1, wherein: The process of calculating the species abundance matrix is specifically: extract DNA based on the collected eDNA samples; amplify, construct a library and sequence the extracted DNA; annotate phytoplankton species according to the sequence information obtained by sequencing and calculate the corresponding species abundance matrix.

6. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 5, wherein: For the DNA extracted from eukaryotic phytoplankton, use 18S rRNA V9 hypervariable region primer pairs for amplification; for the DNA extracted from Cyanophyta, use modified 16S rRNA V3 region primer pairs for amplification.

7. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 1, wherein: The process of screening the training data set is specifically: according to the sampling time and sampling sites of the collected eDNA, select remote sensing image data with a resolution of greater than or equal to 100 meters; perform atmospheric correction and radiometric correction on the selected remote sensing image data to obtain the training data set.

8. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 7, wherein: Perform cloud cover screening on the selected remote sensing image data, and select remote sensing image data with a cloud cover below 10% to construct the training data set.

9. The method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to claim 1, wherein: The process of training the prediction model is specifically as follows: The input independent variables and dependent variables are placed in the data matrix according to the corresponding positions, 70% of the data in the data matrix is used as the training set of the prediction model, and the remaining 30% of the data is used as the validation set of the prediction model; The training set of the prediction model is input into the random forest model for training the random forest model; Grid tuning is performed, and the parameter with the highest R-square is selected as the training parameter of the prediction model; The training parameter is set, and the random forest model is trained again to obtain the prediction model.

10. A system for the method for predicting the spatial distribution of phytoplankton based on environmental DNA and machine learning according to any one of claims 1-9, characterized in that: It includes a collection module: determining sampling sites and collecting phytoplankton eDNA samples; A module for calculating the species abundance matrix: extracting and processing DNA from the collected eDNA samples and calculating the species abundance matrix; A module for screening the training data set: screening the training data set for training the prediction model according to the sampling time and sampling sites of the eDNA samples; A training module: inputting the training data set as the independent variable of the training set and the species abundance matrix as the dependent variable of the training set to train the prediction model; An output module: inputting remote sensing image data into the prediction model to obtain a species spatial distribution map.

Images