A method for monitoring long-term changes in primary productivity of rivers at a continental scale
By constructing an LSTM model based on daily meteorological and hydrological elements and watershed attributes, a continuous time series of river primary productivity was generated, solving the problem of monitoring long-term changes at the continental scale and realizing low-cost and efficient river ecosystem analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INST OF GEOGRAPHY & LIMNOLOGY
- Filing Date
- 2026-01-27
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to monitor long-term changes in river primary productivity at a continental scale, limited by reliance on high-frequency monitoring data and high costs, and machine learning methods also have limitations in spatial coverage and accuracy.
A continental-scale river primary productivity simulation model was constructed by combining daily-scale meteorological and hydrological elements and watershed attributes with a long short-term memory network (LSTM) deep learning algorithm. Hyperparameters were optimized through grid iteration to generate continuous daily-scale primary productivity time series data, and the annual mean and trend were analyzed.
It enables low-cost and easily scalable monitoring of long-term changes in river primary productivity, breaking through the time and space limitations of traditional methods, providing robust monitoring support, and is applicable to river ecological monitoring in different continents or regions.
Smart Images
Figure CN122132722A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of river ecosystem monitoring, and in particular relates to a method for monitoring long-term changes in primary productivity of rivers at a continental scale. Background Technology
[0002] River primary productivity, as a key indicator for understanding energy flow, trophic status, and aquatic population succession within river ecosystems, is the sum of all matter synthesized by photosynthesis in aquatic organisms.
[0003] In recent years, with climate warming and intensified human activities, river ecosystems have undergone unprecedented changes. Frequent outbreaks of eutrophication, algal blooms, and hypoxia events can have immeasurable impacts on river primary productivity. Currently, river productivity is mainly measured in the form of carbon dioxide (CO2) and dissolved oxygen (DO). Taking dissolved oxygen as an example, river primary productivity is estimated using a "day-night method" by continuously and frequently measuring the diurnal variation of dissolved oxygen concentration. Specifically, since photosynthesis requires light, the dynamic changes in dissolved oxygen at night are used to estimate the river ecosystem's respiration, while the dynamic process of dissolved oxygen during the day is used to estimate the river's net productivity. The difference between the two yields the river's primary productivity. The difficulty in estimating river primary productivity using dissolved oxygen lies in removing nitrogen dioxide (F2). dv,t and F dv,t Contribution to the diurnal dynamic variation of dissolved oxygen.
[0004] Existing methods for calculating river primary productivity all utilize high-frequency dissolved oxygen monitoring data at the minute-to-hour level. Statistical analysis methods, such as maximum likelihood estimation, Kalman filtering, and Bayesian methods, are then used to eliminate the influence of oxygen diffusion in river water and river-atmosphere oxygen exchange, thus obtaining the river's primary productivity. Meanwhile, machine learning algorithms have recently been explored for understanding the short-term variations in the primary productivity of individual rivers.
[0005] However, estimating river primary productivity using high-frequency dissolved oxygen diurnal variation data and statistical methods generally suffers from the following drawbacks: 1) it heavily relies on high-frequency monitoring data; 2) it has high computational time costs and poor monitoring timeliness, making it unsuitable for analyzing and monitoring long-term changes in river primary productivity at the regional scale. Furthermore, the low monitoring accuracy and spatial coverage of machine learning monitoring methods limit their potential for application at the continental scale. Summary of the Invention
[0006] To address the technical problems existing in the background art described above, and to understand the response of river primary productivity to climate change at the continental scale, this invention provides a method for monitoring long-term changes in river primary productivity at the continental scale.
[0007] This invention employs the following technical solution: a method for monitoring long-term changes in primary productivity of rivers at a continental scale, comprising the following steps: Acquire daily meteorological and hydrological data, combine this with geographic information of the target river's region, and determine the effective duration of daytime each day. This integrates and forms a time series of daily meteorological and hydrological elements; Match the corresponding watershed boundaries and extract watershed attributes, while filtering out high-quality primary productivity data from the river primary productivity dataset; By fusing the daily-scale meteorological and hydrological elements' time series and watershed attributes to form daily input features and high-quality primary productivity data as output variables, a continental-scale river primary productivity simulation model is constructed; the model's hyperparameters are optimized and trained using a grid iterative optimization method. The simulated output of the first continental-scale river primary productivity simulation model was generated using the trained model. Simulated primary productivity values of the Sun River And generate daily-scale primary productivity time series data; Based on the daily-scale primary productivity time series data, the corresponding mean and annual mean of primary productivity are calculated, the correlation between the annual time series and the annual mean is analyzed, and the long-term trend of primary productivity is determined.
[0008] In a further embodiment, the daily-scale meteorological and hydrological elements include: daily minimum air temperature, daily maximum air temperature, daily minimum skin temperature, daily maximum skin temperature, shortwave radiation, subsurface runoff, and river flow. The watershed attributes include climate attributes, hydrological attributes, soil attributes, geological attributes, topographic attributes, and land use-related attributes.
[0009] In a further embodiment, the geographic information includes: the longitude of the region. ,latitude Local UCT time zone Correspondingly, the effective duration of daytime The determination process is as follows: Combining geographic information and date The sunrise time of the day was calculated separately. and the time of sunset ; Based on the time of sunrise and the time of sunset The effective duration of daytime was calculated. : .
[0010] In a further embodiment, the high-quality primary productivity data is obtained through a dual screening criterion; wherein the dual screening criterion is: Data confidence level standard: Select observation records with a confidence level of medium or above from the river primary productivity dataset; Numerical validity screening criteria: observation records with primary productivity observation values greater than 0.
[0011] In a further embodiment, the construction process for the continental-scale river primary productivity simulation is as follows: The first Hidden historical states of the Japanese model With the Daily input characteristics Input the forget gate, input gate, and output gate respectively, and obtain the forget gate output respectively. Input gate output and output gate output ; Output based on forget gate Input gate output and output gate output Calculate the candidate state and in accordance with the first Daily input characteristics and historical hidden state Determine the gate status ; Based on candidate status and gating status Output the first Simulated primary productivity values of the Sun River .
[0012] In a further embodiment, a data preprocessing step is included before optimizing the model hyperparameters during training: The following formula is used to perform a logarithmic transformation on the river flow and subsurface runoff in the daily input features to obtain the transformed data. : ,in, The data before conversion includes: river flow and subsurface runoff; The following formula is used to perform a logarithmic transformation on high-quality primary productivity data to obtain the transformed output variables. : , High-quality primary productivity data before conversion; The Z-score method was used to evaluate all daily input features and transformed output variables. The dataset is standardized by randomly dividing it into a training set and a test set according to a set ratio. The training set is used for learning model hyperparameters, and the test set is used for validating model performance.
[0013] In a further embodiment, the process of generating the daily-scale primary productivity time series data is as follows: Based on the time axis of the monitoring period, the simulated values of primary productivity of the river were... Time series data are regularized according to the time axis to form continuous and complete daily-scale primary productivity time series data. , This represents the total number of days within the monitoring period.
[0014] In a further embodiment, the correlation analysis process between the annual time series and the annual mean is as follows: Based on the aforementioned daily-scale primary productivity time series data, the average primary productivity of the target river is statistically analyzed annually. : ,in, For the monitoring period Average primary productivity in years For the first The number of valid days in a year ; For the first l The year's first p Primary productivity per effective day; Using annual time series as the independent variable, the corresponding mean primary productivity Using [variable name] as the dependent variable, construct a linear regression model, which is expressed as follows: ,in, The slope of the linear regression equation is denoted as . The intercept of the linear regression equation is denoted as .
[0015] In a further embodiment, the long-term trend of primary productivity is determined as follows: If the slope This indicates that primary productivity is on a long-term upward trend; If the slope This indicates a long-term downward trend in primary productivity.
[0016] In a further embodiment, the model hyperparameters include: number of iterations. Batch sample quantity B Training sequence length Hidden layer dimension and learning rate ; Correspondingly, during the hyperparameter optimization training process of the continental-scale river primary productivity simulation model, the mean squared error loss function is used to measure the deviation between the model's predicted values and the measured values. The formula is expressed as follows: ; in, Indicates the number of iterations. For the number of samples, and The first i The second iteration j Predicted and measured values for each sample. The number of weights in the model. For the first The weight value of each weight.
[0017] The beneficial effects of this invention are as follows: This invention overcomes the data dependency bottleneck by using publicly available daily meteorological and hydrological elements and constant watershed attributes as core inputs for the first time. Combined with a Long Short-Term Memory (LSTM) deep learning algorithm, it eliminates the strong dependence of traditional methods on minute- to hourly high-frequency dissolved oxygen monitoring data. It eliminates the need for additional, expensive monitoring equipment, avoids the difficulties and high costs associated with acquiring high-frequency data, significantly lowers the implementation threshold for continental-scale monitoring, and achieves a low-cost, easily scalable monitoring model.
[0018] Compared to the limitations of traditional statistical methods, which are only applicable to local rivers, and conventional machine learning algorithms, which are limited to short-term changes (seasonal differences) in a single river, this invention expands the monitoring scope from a single river to a continental scale, and extends the monitoring period from the short term to a long-term dimension of 40 years or even longer. By generating continuous and complete daily-scale primary productivity time series data, it fully captures the long-term evolution trajectory and spatial distribution pattern of river primary productivity, solving the problem of long-term trend analysis caused by the limited spatial coverage and uneven temporal distribution of existing data.
[0019] This invention also employs a dual-screening criterion to select high-quality training data, combined with preprocessing steps such as logarithmic transformation and Z-score standardization, effectively eliminating the interference of zero values, extreme values, and dimensional differences on the model. Hyperparameters are then optimized through grid-based iterative optimization, resulting in significantly higher monitoring accuracy than existing machine learning methods. The model uses a mean squared error loss function with L2 regularization to effectively avoid overfitting, providing robust and reliable technical support for large-scale long-term monitoring. Furthermore, the method is highly replicable and can be extended to river ecological monitoring scenarios across different continents or regions.
[0020] This invention eliminates the complex oxygen diffusion and atmospheric exchange flux removal process in traditional methods. Instead, it automatically mines the temporal correlations between meteorological and hydrological elements, watershed attributes, and primary productivity through the gating mechanism (forget gate, input gate, output gate) of the LSTM model, significantly reducing computational time costs. The model training and simulation process is streamlined and automated, requiring no complex manual intervention, significantly improving monitoring timeliness and enabling rapid response to the monitoring needs of long-term changes in primary productivity of rivers at continental scales. Attached Figure Description
[0021] Figure 1 This is a flowchart of a method for monitoring long-term changes in primary productivity of a continental river, as described in Example 1.
[0022] Figure 2 This is the standardized dataset from Example 1. Detailed Implementation
[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0024] Example 1 like Figure 1 As shown, this embodiment provides a method for monitoring long-term changes in primary productivity of rivers at a continental scale, including the following steps: Obtain daily meteorological and hydrological data, and combine this with geographical information about the target river's location to determine the effective duration of daytime each day. The daily-scale meteorological and hydrological elements are integrated to form a time series of meteorological and hydrological elements. It should be noted that the daily-scale meteorological and hydrological elements described in this embodiment can be obtained from publicly released meteorological and hydrological data.
[0025] Match the corresponding watershed boundaries and extract watershed attributes, while filtering out high-quality primary productivity data from the river primary productivity dataset; the river primary productivity mentioned in this embodiment is the sum of organic matter produced by all organisms in the river capable of photosynthesis.
[0026] By fusing the daily-scale meteorological and hydrological elements' time series and watershed attributes to form daily input features and high-quality primary productivity data as output variables, a continental-scale river primary productivity simulation model is constructed; the model's hyperparameters are optimized and trained using a grid iterative optimization method. The simulated output of the first continental-scale river primary productivity simulation model was generated using the trained model. Simulated primary productivity values of the Sun River And generate daily-scale primary productivity time series data; Based on the daily-scale primary productivity time series data, the corresponding mean and annual mean of primary productivity are calculated, the correlation between the annual time series and the annual mean is analyzed, and the long-term trend of primary productivity is determined.
[0027] In a further embodiment, daily-scale meteorological and hydrological elements include: daily minimum temperature, daily maximum temperature, daily minimum skin temperature, daily maximum skin temperature, shortwave radiation, subsurface runoff, and river flow. These data can be obtained from publicly available meteorological and hydrological observation datasets and long-term monitoring network data, such as the ERA5-Landdaily dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF) (providing daily minimum temperature, daily maximum temperature, daily minimum skin temperature, daily maximum skin temperature, shortwave radiation, subsurface runoff, etc.), the Global Runoff and Discharge Center (GRDC) long-term monitoring dataset (supplementing river flow data), and the USGS hydrological monitoring network data (providing long-term observation records of river flow and related hydrological elements in key areas). The data spans from 1980 to 2020, and daily element data for the corresponding time period of the target river can be extracted through latitude and longitude spatiotemporal matching.
[0028] Correspondingly, the watershed attributes described in this embodiment include climate attributes, hydrological attributes, soil attributes, geological attributes, topographic attributes, and land use-related attributes. These can be extracted from the GAGES II (Geospatial Attributes of Gages for Evaluating Streamflow) dataset by matching the latitude and longitude of the target river to the corresponding watershed boundary, ensuring accurate correspondence between the data and the study area. For ease of understanding, this embodiment provides examples of the above data. Climate attributes include: average watershed temperature, average watershed precipitation, relative humidity of the watershed, annual average potential evapotranspiration, snow-to-precipitation ratio, and precipitation seasonality index, comprehensively reflecting the long-term climate characteristics of the watershed and the spatiotemporal distribution patterns of precipitation and temperature, providing climatic background support for primary productivity simulation.
[0029] The hydrological attributes include baseflow index, topographic humidity index, river-subsurface water exchange time index, and river density, which characterize the watershed's hydrological cycle process, water resource replenishment characteristics, and river network distribution pattern, reflecting the regulatory role of hydrological conditions on primary productivity.
[0030] Soil properties can include soil particle composition (ratio of sand, clay, and silt), soil bulk density, soil organic carbon content, groundwater depth, rock layer depth, soil erosion factors, soil precipitation runoff factors, soil permeability, and soil moisture content. These properties encompass core physical, chemical, and hydrological attributes of the soil, reflecting its ability to retain and supply water and nutrients, and indirectly affecting the photosynthetic efficiency of aquatic organisms.
[0031] Geological attributes are represented by the area proportion of dominant geological types, clarifying the distribution of major geological structures and rock types within the watershed, and relating to soil formation, hydrological processes, and nutrient release patterns, thus providing geological background parameters for primary productivity simulation.
[0032] Topographic attributes include river elevation, average elevation of the basin, topographic relief, basin slope, basin aspect, latitude and longitude, and basin area. These attributes visually present the topographic features and spatial location information of the basin, influencing the spatial distribution of sunlight, temperature, and precipitation, as well as river morphology and hydrological connectivity.
[0033] Land use-related attributes are the proportions of farmland, water bodies, urban land, and forest area, reflecting the impact of human activities and natural cover on the watershed ecosystem. They indirectly affect the primary productivity of rivers by altering surface runoff, nutrient input, and light conditions.
[0034] Furthermore, the geographical information described in this embodiment includes: the longitude of the region. ,latitude Local UCT time zone Correspondingly, the effective daytime duration The determination process is as follows: Combining geographic information and date The sunrise time of the day was calculated separately. and the time of sunset ; Based on the time of sunrise and the time of sunset The effective duration of daytime was calculated. : .
[0035] The sunrise time described in this embodiment and the time of sunset The formula for determining it is as follows: ; .
[0036] To address the issues of inconsistent data reliability and the presence of invalid outliers in existing river primary productivity datasets, the high-quality primary productivity data in this embodiment is obtained through a dual screening criterion. The dual screening criterion is as follows: Data confidence level standards: From the river primary productivity dataset, observation records with a confidence level of medium or higher are selected. The confidence level is comprehensively assessed based on the proportion of outliers in the observation data, the range of variation of the standardized gas exchange coefficient, and the standard deviations of the gas exchange coefficient and process error. Specifically, in this embodiment, a confidence level of medium or higher is defined as follows: negative primary productivity records below 50%, the range of variation of the standardized gas exchange coefficient below 50, and the maximum standard deviation of the gas exchange coefficient and process error below 1.2.
[0037] Numerical validity screening criteria: Observational records with primary productivity values greater than 0. Invalid records with negative values due to observational equipment malfunction, sensor drift, abnormal data transmission, or environmental interference were screened out to ensure that the selected data conformed to the ecological essential attribute of primary productivity as the total amount of photosynthetic substances synthesized by aquatic organisms.
[0038] Furthermore, the construction process for the continental-scale river primary productivity simulation in this embodiment is as follows: The first Hidden historical states of the Japanese model With the Daily input characteristics Input the forget gate, input gate, and output gate respectively, and obtain the forget gate output respectively. Input gate output and output gate output Among them, the forget gate output Hidden in history The time-series information is effective for simulating current primary productivity; the input gate output is... The contribution weights of input features to the primary productivity simulation.
[0039] The formula for calculating the forgetting gate is as follows: ; in, W f This is the unit weight matrix of the forget gate. C t-1 for t The cell state update value at time -1. and r t These represent updating the door and resetting the door state, respectively. It is the activation value of the Forgotten Gate.
[0040] The formula for calculating the input gate is: ;in, for sigmoid function, and These represent the unit weight matrix and activation value of the input gate, respectively.
[0041] The formula for calculating the output gate is: ; It is the hyperbolic tangent function. and These are the unit weight matrix and activation value of the output gate, respectively.
[0042] Output based on forget gate Input gate output and output gate output Calculate the candidate state The specific calculation formula is as follows: .
[0043] Furthermore, gating status The formula for determining it is: .
[0044] in ，w hg , b hg and These are the unit weight matrix, activation value, and hidden gate output value of the hidden node output gate, respectively.
[0045] Finally, based on the candidate state and gating status Output the first Simulated primary productivity values of the Sun River : ,in, The output function of the LSTM model iteration, i.e., the final simulated value is the nth... t Daily Iteration Candidate Status d t and gating status o t The dot product.
[0046] To ensure model training accuracy and avoid interference from zero values, extreme values, and dimensional differences in the original data on the long short-term memory network model's learning of the correlation between meteorological and hydrological elements, watershed attributes, and primary productivity, data preprocessing steps are included before optimizing the model's hyperparameters during training. The following formula is used to perform a logarithmic transformation on the river flow and subsurface runoff in the daily input features to obtain the transformed data. : ,in, The original data includes river flow and subsurface runoff; the transformed data distribution is made closer to a Gaussian distribution to eliminate the influence of zero values. The following formula is used to perform a logarithmic transformation on high-quality primary productivity data to obtain the transformed output variables. : , To ensure high-quality primary productivity data before conversion, and to avoid the disproportionate impact of extreme values.
[0047] The Z-score method was used to evaluate all daily input features and transformed output variables. The dataset is standardized by randomly dividing it into training and test sets in a ratio such as 7:3. The training set is used for learning model hyperparameters, and the test set is used for validating model performance.
[0048] The above technical solutions achieve denoising, normalization, and dimensional unification of the original data, significantly improving the quality and consistency of the model training data. This lays a solid data foundation for subsequent hyperparameter optimization and accurate model simulation of primary productivity of rivers at the continental scale, ensuring that the model can stably learn the intrinsic relationship between environmental factors and primary productivity over long time series, and ultimately improving the reliability of daily-scale simulation results and long-term trend analysis.
[0049] Therefore, the model hyperparameters described in this embodiment include: number of iterations. Batch sample quantity B Training sequence length Hidden layer dimension and learning rate .
[0050] Combination Figure 2 Taking the continental rivers of the United States as an example, all input and output data were randomly divided into 70% and 30%, respectively, to serve as the training and testing datasets for the Long Short-Term Memory (LSTM) network model. The LTM network model consisted of one hidden layer and a fully connected layer, with the sigmoid function as the activation function. The optimal hyperparameters were obtained through iterative optimization: 100 iterations, 50 batch samples, a training sequence length of 50, a hidden layer dimension of 128, and a learning rate of 0.15.
[0051] Correspondingly, during the hyperparameter optimization training process of the continental-scale river primary productivity simulation model, the mean squared error loss function is used to measure the deviation between the model's predicted values and the measured values. The formula is expressed as follows: ; in, Indicates the number of iterations. For the number of samples, and The firsti The second iteration j Predicted and measured values for each sample. The number of weights in the model. For the first The weight value of each weight.
[0052] To address the limitations of existing technologies in obtaining continuous and complete time-series data on river primary productivity and in accurately quantifying long-term trends, the process for generating daily-scale primary productivity time-series data is as follows: Using the monitoring period (1980-2020) as a benchmark, the simulated values of river primary productivity were... Time series data are regularized according to the time axis to form continuous and complete daily-scale primary productivity time series data. , This represents the total number of days within the monitoring period. It breaks through the limitations of traditional monitoring data, which suffers from uneven temporal distribution and limited spatial coverage, providing continuous and complete data support for subsequent long-term trend analysis.
[0053] Based on the aforementioned continuous and complete daily time-series data, the correlation analysis process between the annual time series and the annual mean in this embodiment is as follows: Based on the aforementioned daily-scale primary productivity time series data, the average primary productivity of the target river is statistically analyzed annually. : ,in, For the monitoring period Average primary productivity in years For the first The number of valid days in a year ; For the first l The year's first p Primary productivity per effective day; Using annual time series as the independent variable, the corresponding mean primary productivity Using [variable name] as the dependent variable, construct a linear regression model, which is expressed as follows: ,in, The slope of the linear regression equation is denoted as . The intercept of the linear regression equation is denoted as .
[0054] This model provides a mathematical basis for the quantitative analysis of long-term trends by quantifying the linear relationship between annual time and average productivity.
[0055] Furthermore, the method for determining the long-term trend of primary productivity described in this embodiment is as follows: If the slope This indicates that primary productivity is on a long-term upward trend; If the slope This indicates a long-term downward trend in primary productivity.
[0056] The above technical solution realizes a complete analysis chain from daily continuous simulation data to annual average statistics and then to long-term trend quantification, effectively solving the problem of difficulty in long-term trend analysis caused by the discontinuity and limited coverage of traditional methods.
Claims
1. A method for monitoring long-term changes in primary productivity of rivers at a continental scale, characterized in that, Includes the following steps: Acquire daily meteorological and hydrological data, combine this with geographic information of the target river's region, and determine the effective duration of daytime each day. This integrates and forms a time series of daily meteorological and hydrological elements; Match the corresponding watershed boundaries and extract watershed attributes, while selecting high-quality primary productivity data from the river ecosystem primary productivity dataset; By fusing the daily-scale meteorological and hydrological elements' time series and watershed attributes to form daily input features and high-quality primary productivity data as output variables, a continental-scale river primary productivity simulation model is constructed; the model's hyperparameters are optimized and trained using a grid iterative optimization method. The simulated output of the first continental-scale river primary productivity simulation model was generated using the trained model. Simulated primary productivity values of the Sun River And generate daily-scale primary productivity time series data; Based on the daily-scale primary productivity time series data, the corresponding mean and annual mean of primary productivity are calculated, the correlation between the annual time series and the annual mean is analyzed, and the long-term trend of primary productivity is determined.
2. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The daily-scale meteorological and hydrological elements include: daily minimum temperature, daily maximum temperature, daily minimum skin temperature, daily maximum skin temperature, shortwave radiation, subsurface runoff, and river flow. The watershed attributes include climate attributes, hydrological attributes, soil attributes, geological attributes, topographic attributes, and land use-related attributes.
3. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The geographic information includes: the longitude of the region. ,latitude Local UCT time zone Correspondingly, the effective duration of daytime The determination process is as follows: Combining geographic information and date The sunrise time of the day was calculated separately. and the time of sunset ; Based on the time of sunrise and the time of sunset The effective duration of daytime was calculated. : 。 4. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The high-quality primary productivity data was obtained through a dual screening criterion; wherein the dual screening criterion is: Data confidence level standard: Select observation records with a confidence level of medium or above from the river primary productivity dataset; Numerical validity screening criteria: observation records with primary productivity observation values greater than 0.
5. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The construction process for the continental-scale river primary productivity simulation is as follows: The first Hidden historical states of the Japanese model With the Daily input characteristics Input the forget gate, input gate, and output gate respectively, and obtain the forget gate output respectively. Input gate output and output gate output ; Output based on forget gate Input gate output and output gate output Calculate the candidate state and in accordance with the first Daily input characteristics and historical hidden state Determine the gate status ; Based on candidate status and gating status Output the first Simulated primary productivity values of the Sun River .
6. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, Before optimizing the hyperparameters of the model, a data preprocessing step is also included: The following formula is used to perform a logarithmic transformation on the river flow and subsurface runoff in the daily input features to obtain the transformed data. : ,in, The data before conversion includes: river flow and subsurface runoff; The following formula is used to perform a logarithmic transformation on high-quality primary productivity data to obtain the transformed output variables. : , High-quality primary productivity data before conversion; The Z-score method was used to evaluate all daily input features and transformed output variables. The dataset is standardized by randomly dividing it into a training set and a test set according to a set ratio. The training set is used for learning model hyperparameters, and the test set is used for validating model performance.
7. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The process of generating the daily-scale primary productivity time series data is as follows: Based on the time axis of the monitoring period, the simulated values of primary productivity of the river were... Time series data are regularized according to the time axis to form continuous and complete daily-scale primary productivity time series data. , This represents the total number of days within the monitoring period.
8. The method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The correlation analysis process between the annual time series and the annual mean is as follows: Based on the aforementioned daily-scale primary productivity time series data, the average primary productivity of the target river is statistically analyzed annually. : ,in, For the monitoring period Average primary productivity in years For the first The number of valid days in a year ; For the first l The year's first p Primary productivity per effective day; Using annual time series as the independent variable, the corresponding mean primary productivity Using [variable name] as the dependent variable, construct a linear regression model, which is expressed as follows: ,in, The slope of the linear regression equation is denoted as . The intercept of the linear regression equation is denoted as .
9. A method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 8, characterized in that, The method for determining the long-term trend of primary productivity is as follows: If the slope This indicates that primary productivity is on a long-term upward trend; If the slope This indicates a long-term downward trend in primary productivity.
10. A method for monitoring long-term changes in primary productivity of rivers at a continental scale according to claim 1, characterized in that, The model hyperparameters include: number of iterations. Batch sample quantity B Training sequence length Hidden layer dimension and learning rate ; Correspondingly, during the hyperparameter optimization training process of the continental-scale river primary productivity simulation model, the mean squared error loss function is used to measure the deviation between the model's predicted values and the measured values. The formula is expressed as follows: ； in, Indicates the number of iterations. For the number of samples, and The first i The second iteration j Predicted and measured values for each sample. The number of weights in the model. For the first The weight value of each weight.