CN115952421A - High-precision time-space simulation method for coupling ecological process model and machine learning algorithm - Google Patents

Abstract

The invention discloses a high-precision time-space simulation method coupling an ecological process model and a machine learning algorithm, and belongs to the research fields related to ecology, agronomy and global change. Including: use the long-term positioning test data in the study area to calibrate the relevant parameters of the ecosystem process model; randomly select a number of grid points, randomly generate management measures and climate change scenario combinations for each grid point, and use the calibrated process model to simulate each Ecosystem state variable changes under grid points and scenarios; training and evaluating the performance of machine learning, and obtaining the optimal integrated model according to the weighted evaluation method of different model weights; using the optimal machine learning integrated model as a proxy model of the process model, The agent simulates changes such as crop yield and soil organic carbon at high-resolution spatio-temporal scales. Through the invention, the spatio-temporal change rule of a certain ecosystem state variable and its response to different management measures and climate change can be predicted more quickly and efficiently at high spatio-temporal resolution.

Description

A high-precision spatio-temporal simulation method for coupling ecological process models and machine learning algorithms

technical field

The invention belongs to the related research fields of ecology, agronomy and global change, and specifically relates to a high-precision spatio-temporal simulation method coupling an ecological process model and a machine learning algorithm.

Background technique

Because the process-based ecosystem model comprehensively considers the interaction between ecosystem processes and the environment (such as climate, soil, etc.) change, food security, biodiversity conservation and natural resource management). However, the application of process models on a large scale has the problems of high computational cost, large model uncertainty, and poor data availability. It is difficult to complete fast simulations at high-precision spatio-temporal resolution within an acceptable time, which limits this type of model. Ubiquity and operability on a large scale.

The process model uses mathematical equations to simplify the ecosystem process, and it needs to input a large number of parameters to complete the calculation. These parameters are estimated based on the experimental data of field observation stations or field test sites, but on a large scale with more complex and diverse environmental conditions, Some ecological process parameters cannot be well estimated, and the uncertainty of the model is large. Machine learning can achieve higher accuracy and simulation efficiency than process models on a large scale, but it may produce results that are contrary to ecosystem processes, and sufficient and high-quality training data must be provided. In this context, the method of "Coupling Process Models and Machine Learning Integrated Simulation" came into being. The process model can be calibrated at the site scale first, and the calibrated model can produce a large amount of simulated data under enough environmental-management scenarios, which provides a large amount of high-quality data for the machine learning model. Based on the simulated data of these process models, the machine learning model can automatically learn the intrinsic relationship between complex ecosystem processes and environment-management, and can achieve fast generalization simulation. Combining the implicit prior knowledge of the process model with the learning ability of machine learning, a small amount of site observation data can be used to flexibly realize fast simulations at high-resolution spatio-temporal scales. However, due to differences in algorithms, different machine learning methods have different abilities to understand ecological processes and model generalization. In order to reduce the uncertainty of machine learning integration as much as possible, the present invention uses adaptive screening to eliminate models with poor learning ability, and builds weights by calculating the root mean square error of different models, weighted average machine learning models with strong learning ability, Thereby improving the simulation accuracy and reducing the uncertainty of the ensemble surrogate model.

The present invention combines a process-based ecosystem model with a variety of machine/deep learning, establishes an optimal integrated agent model through self-adaptive screening and weighted average, and realizes fast and accurate simulation with high-precision spatio-temporal resolution. Provide technical methods for modeling and management decision-making of ecosystems at different scales.

Contents of the invention

The purpose of the present invention is to overcome the defects in the prior art, and provide a high-precision spatio-temporal simulation method coupling the ecological process model and the machine learning algorithm. The present invention couples the process-based ecosystem model with various machine/deep learning models, establishes an optimal integrated agent model through self-adaptive screening and weighted average, and accurately and quickly realizes the high-precision spatio-temporal resolution of ecosystem state variables. simulation.

The concrete technical scheme that the present invention adopts is as follows:

The present invention provides a high-precision spatio-temporal simulation method coupling ecological process model and machine learning algorithm, specifically as follows:

S1: Calibrate and validate the relevant parameters of the ecosystem process model using long-term location test data in the target study area;

S2: In the study area, randomly select a number of grid points that can cover the spatial variation of climate types and soil properties in the study area, and randomly generate a combination of management measures and climate change scenarios for each grid point; The soil data and historical climate data of the soil data and the historical climate data, use the ecosystem process model verified by step S1 to simulate the changes of the ecosystem state variables under each grid point and scenario;

S3: Integrate the input data and output data of the scenario simulation of the ecosystem process model in step S2, and use it to train different types of machine learning models, so that the machine learning model can fully learn and simulate the state variable changes simulated by the ecosystem process model; evaluate the machine learning performance, adaptively eliminate machine learning models with poor performance, and obtain the optimal integrated model according to the weighted evaluation method of different model weights;

S4: Using the optimal integrated model as a proxy model of the ecosystem process model, based on spatial data, simulate and generate high-precision digital mapping products required for research purposes at high-resolution spatio-temporal scales.

Preferably, the step S1 is specifically as follows:

For specific research purposes, determine the research area that needs to be simulated and the appropriate ecosystem process model; according to the needs of the ecosystem process model, obtain the observation data of the ecosystem in situ in the study area, and use the differential evolution algorithm to model the ecosystem process model. Calibration and verification of key parameters.

Further, the ecosystem in-situ site observation data includes one or more of meteorological data, soil attribute data, farmland ecosystem data, and ecosystem state variable data.

Furthermore, the meteorological data includes daily temperature, precipitation and solar radiation, the soil attribute data includes soil organic carbon, pH and soil bulk density, and the farmland ecosystem data includes management data of sowing date, fertilization, and irrigation stages and yield data, the ecosystem state variable data includes soil organic carbon and greenhouse gas emission observation data.

Further, the key parameters include crop variety parameters, soil organic carbon and N ₂ O emission process parameters in the farmland ecosystem.

As preferably, the step S2 is specifically as follows:

In the research area, a spatial grid is generated according to the set spatial resolution, and a number of grid points that can cover the spatial variation of the climate type and soil properties in the research area are randomly selected from all the grid points, and a grid point is randomly generated for each grid point. combination of various management measures and climate change scenarios; collect and prepare soil data and historical climate data corresponding to selected grid points to drive the ecosystem process model that has been verified in step 1, and simulate historical and future scenarios State variables of interest.

Further, the management measure scenario refers to various possible human activities management. In the farmland ecosystem, the management measure scenario includes crop system scenario, crop sowing date scenario, fertilization scenario, irrigation scenario and straw return scenario; The climate change scenario refers to the future climate change scenario, which is selected and determined from the global climate model and the emission scenarios of the shared economic path.

Preferably, in the step S3, the machine learning model includes LASSO model, support vector machine, random forest, XGBoost, convolutional neural network and long short-term memory network.

Preferably, in the step S3, the root mean square error RMSE and the coefficient of determination ^R2 are respectively calculated to evaluate the performance of each machine learning model to predict the output of the process model, and by setting the screening criteria, ^R2 ranks models lower than the average level of each model It will be discarded, and several models whose performance is above the average level are selected; according to the model performance, the weight is calculated for each selected model, and the weight is 1/RMSE, and a weighted average integration model is constructed to obtain the optimal integration Model.

As preferably, the step S4 is specifically as follows:

Collect and prepare high-precision spatial data, drive the agent model that has been constructed, realize fast simulation through parallel computing, and generate high-precision digital mapping products required for research purposes.

Compared with the prior art, the present invention has the following beneficial effects:

1) By obtaining site-scale ecosystem observation data, calibrating the ecosystem process model, and then carrying out a large number of management-environmental scenario simulations through the process model. Use a variety of machine learning models to simulate the scenario simulation results of the process model, fully learn the internal relationship in the process model, and obtain the optimal integrated model through adaptive screening and weighted average, and finally use it for higher resolution spatio-temporal scale Fast simulation of ecosystem state variables.

2) The innovative use of process model and machine learning hybrid modeling not only ensures the mechanism of the model but also improves the model accuracy and simulation efficiency of the model at a fine scale. The weight method is used to integrate the model to further reduce system errors and improve data quality. Accuracy and reliability will help us understand and predict the impact of simulated future climate change and different management measures on ecosystems.

3) The present invention is applicable to the state attribute simulation of any region, any time, and any ecosystem, such as ecosystem productivity, crop yield, and leaf area index on a global scale, a national scale, or a regional scale, historical or future scenarios , soil organic carbon, N ₂ O emissions and other indicators can be simulated more efficiently and accurately to predict the simulation of ecosystem state attributes and its response to different human activities management and climate change at high-precision spatio-temporal resolution. The wide application of this technology will provide scientific and technical support for ecosystem simulation on high-precision spatio-temporal scales.

Description of drawings

Fig. 1 is a schematic flow chart of the method of the present invention.

Detailed ways

The present invention will be further elaborated and illustrated below in conjunction with the accompanying drawings and specific embodiments. The technical features of the various implementations in the present invention can be combined accordingly on the premise that there is no conflict with each other.

The invention provides a high-precision spatio-temporal simulation method coupling an ecological process model and a machine learning algorithm. The steps include: 1) using the long-term positioning test data in the research area to calibrate the relevant parameters of the ecological system process model; In the region, a number of grid points are randomly selected to cover the spatial variation of climate types and soil properties in the study area. Each grid point randomly generates a combination of management measures and climate change scenarios, and the calibrated process model is used to simulate the conditions under each grid point and scenario. Changes in ecosystem state variables (such as crop yields in farmland ecosystems, soil organic carbon, etc.); 3) Integrate the input and output data of process model scenario simulations for training different types of machine learning models, so that the machine learning models can fully learn and the state variable changes simulated by the simulation process model; evaluate the performance of machine learning, adaptively eliminate poorly performing machine learning models, and obtain the optimal integrated model according to the weighted evaluation method of different model weights; 4) combine the optimal machine learning The integrated model is used as a proxy model of the process model, and the proxy simulates changes in crop yield and soil organic carbon at high-resolution spatio-temporal scales.

Each step will be described in detail below.

Step 1: Calibration and verification of the ecosystem process model

First, for the specific research purpose, determine the research area to be simulated and the suitable ecosystem process simulation, and obtain the in-situ site observation data of the ecosystem in the research area according to the needs of the model, generally including daily temperature, precipitation, solar radiation, etc. Meteorological data, soil attribute data such as soil organic carbon, pH, soil bulk density, etc., for farmland ecosystems may also include management data such as sowing date, fertilization, irrigation, and yield data, as well as other interesting ecosystem state variable data (such as soil organic Carbon and greenhouse gas emission observation data, etc.). The key parameters of the model (such as parameters of crop varieties in the farmland ecosystem, parameters of soil organic carbon and N ₂ O emission process, etc.) were corrected and verified by differential evolution algorithm.

Step 2: Scenario generation and simulation based on the ecosystem process model

In order for the machine learning model to fully learn and simulate the internal relationship of the process model, it is necessary to simulate enough data with the process model calibrated at the site level. This step is divided into three sub-steps, namely scenario generation, input data preparation, process model simulation and output.

The first is scenario generation. Spatial grids are generated according to the set spatial resolution, and a number of grid points are randomly selected from all grids (covering all climate types and spatial variations of soil properties in the study area as much as possible). Stochastic generation of management-climate change scenarios. Management scenarios refer to the management of various possible human activities. In farmland ecosystems, management scenarios include crop system scenarios, crop sowing date scenarios, fertilization scenarios, irrigation scenarios (such as irrigation methods, irrigation depths, and irrigation periods, etc.), straw returning to the field, etc. scenarios etc. Climate scenarios refer to future climate change scenarios, which can be selected and determined from Global Climate Models (GCMs) and Shared Economy Pathway Emission Scenarios (SSP). The second is to collect and prepare the soil data and historical climate data corresponding to the selected grid points, to drive the ecosystem process model that has been verified in step 1, and to simulate the state variables of interest in historical and future scenarios, such as farmland ecology results in crop yields, soil organic carbon content, etc. in the system.

Step 3: Machine/Deep Learning Proxy Model

This step is to screen and integrate the machine/deep learning model that can act as a proxy process model. By integrating the input and output results of the grid-management-climate scenario combination in step 2, it is used to train the machine learning model (such as LASSO model, support vector Machine learning, random forest, XGBoost and other machine learning and deep learning models such as convolutional neural network or long short-term memory network), let the machine learning model learn and simulate the internal relationship of the process model. Data is grouped before model training, 80% of the data (that is, the output of the process model in step 2) trains the model, and 20% of the data verifies the model. Calculate the root mean square error (RMSE) and coefficient of determination (R ² ) to evaluate the performance of each machine learning model to predict the output of the process model. By setting the screening criteria, the model whose R ² ranking is lower than the average level of the six models will be discarded. Filter out a few models where the model performs above average. According to the model performance, the weight is calculated for each selected model, and the weight is 1/RMSE, and a weighted average integrated model is constructed, which is the proxy model of the process model.

Step 4: Fast and high-precision spatio-temporal simulation based on surrogate model

Collect and prepare high-precision spatial data, drive the constructed proxy model, realize fast simulation through parallel computing, and generate high-precision digital mapping products required for research purposes.

Example 1

The research content of this example is the simulation of crop yield and soil organic carbon changes in the winter wheat/summer corn rotation system at a spatial scale of 1 km in the North China Plain under different management modes, as shown in Figure 1, specifically including the following steps:

1) Taking the winter wheat-summer maize rotation system in the North China Plain as the research object, the soil, climate, management measures and yield, soil organic carbon, N ₂ O Observation data such as emissions are used to calibrate the parameters of the agricultural ecosystem model (take the APSIM model as an example) through the differential evolution algorithm.

2) According to the 1km spatial resolution of the North China Plain, 450,000 farmland farmland grids were generated, and 6,000 grids were randomly selected from them. At the same time, corresponding management scenarios and climate scenarios were randomly generated for each grid. days as the interval, from 30 days before to 30 days after the reference sowing date), nitrogen fertilizer scenario (the amount of nitrogen fertilization is 1kg ha ^-1 yr ^-1 , increased from 0kgha ^-1 yr ^-1 to 400kg ha ^-1 yr ^{-1 1.} Due to little change in regional fertilization methods, nitrogen topdressing period, basal topdressing ratio is set to the traditional mode), organic fertilizer scenario (organic fertilizer is represented by farm manure, and the unit is 10kg, increased from 0kg ha ^-1 to 4000kgha ^-1 ), straw returning scenarios (the amount of straw returning to the field is increased from 0% to 100% in units of 10%, a total of 11 straw returning scenarios), irrigation scenarios [the irrigation depth is 50cm, and the wheat season irrigation setting is: 0 times, 1 time (joint joint), 2 times (joint joint, grouting), 3 times (returning green, jointing, grouting), 4 times (returning green, jointing, booting and filling), the irrigation settings for the corn season are 0 time, 1 time (joint joint) , 2 times (jointing, tasseling), 3 times (jointing, tasseling, grouting), 4 times (jointing, bell-mouth stage, tassing and grouting) and automatic irrigation (when the field water capacity is lower than 30%, 40%, 50%, 60%, 70% and 80% start irrigation)] etc. Climate scenarios were randomly selected from more than 30 global climate models (GCMs) and shared economy pathway emissions scenarios (SSPs) in CMIP6, generating a total of 6000 grid-management-climate scenario combinations.

3) Simulate 6000 grid-management-climate scenario combinations using the calibrated agro-ecosystem model to simulate wheat and maize yields, soil organic carbon.

4) Integrate 6000 combined simulation results of grid-management-climate scenarios, 80% as training set and 20% as verification set. Six different types of machine learning models are trained on the training set, and the optimal model for training is obtained through parameter optimization and hyperparameter adjustment. The trained model is used to simulate the data of the verification set to evaluate the performance of the model.

5) Based on the simulated values of the process model and machine learning on the verification set, calculate the statistical indicators ( ^R2 and RMSE) for evaluating the performance of the machine learning simulation process model, and eliminate the machine learning models below the average level according to the level of ^R2 . Then calculate the weight according to RMSE, and obtain the optimal integrated model by weighted average. In this example, the R ² values of LASSO model and SVM model for simulating wheat yield, corn yield and soil organic carbon changes were 0.76, 0.59-0.62, and 0.68, respectively, which were lower than the average level and were discarded. The R ² of the four indicators of random forest, XGBoost, convolutional neural network and long short-term memory network simulation are all above 0.93 and are reserved for use. The four models respectively calculate the RMSE of the verification set, and 1/RMSE is used as the weight to construct the optimal integrated model for the next fine-scale simulation. For example, in this example, the RMSE of wheat yield simulated by random forest, XGBoost, convolutional neural network and long short-term memory network are respectively 0.43Mg ha ^-1 , 0.29Mg ha ^-1 , 0.35Mg ha ^-1 , 0.35Mgha ^-1 , four The corresponding weights of the model are 2.33, 3.45, 2.86 and 2.86 respectively, and the final result is obtained according to the weighted average.

玉米产量是

SOC是

未来小麦产量基本保持不变，玉米产量和土壤有机碳的变化呈现下降趋势。6) Prepare historical (1995-2014) and future (2030s, 2060s) 1km-scale grid input data in the North China Plain. The format of the input data is consistent with the data of the machine learning model, and the prepared data will drive optimal machine learning. Integrating the model and setting different management levels at the same time, the spatial distribution maps of wheat and corn yields, soil organic carbon and N ₂ O emissions in the North China Plain were simulated under historical and future scenarios. For this example, under the historical optimal nitrogen fertilizer, irrigation and straw returning management, the wheat yield in the North China Plain in the historical period is

corn yield is

SOC is

In the future, wheat production will remain basically unchanged, and changes in maize production and soil organic carbon will show a downward trend.

It can be seen that the method of the present invention can more quickly and efficiently predict the spatiotemporal variation of certain ecosystem state variables (such as crop yields in farmland ecosystems, soil organic carbon, etc.) Different management measures and responses to climate change.

The above-mentioned embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. Various changes and modifications can be made by those skilled in the relevant technical fields without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims (10)

1. A high-precision space-time simulation method for coupling an ecological process model and a machine learning algorithm is characterized by comprising the following steps:

s1: calibrating and verifying relevant parameters of the ecosystem process model by using long-term positioning test data in a target research area;

s2: randomly selecting a plurality of grid points capable of covering the climate type and soil characteristic spatial variation of the research area in the research area, and randomly generating a management measure and climate variation scene combination for each grid point; simulating the state variable change of the ecological system at each grid point and under the scene by using the ecological system process model verified in the step S1 based on the soil data and the historical climate data at the scene corresponding to each grid point;

s3: integrating input data and output data of the ecosystem process model scene simulation in the step S2, and training different types of machine learning models to enable the machine learning models to fully learn and simulate state variable changes of the ecosystem process model simulation; evaluating the performance of machine learning, adaptively eliminating the poor-performance machine learning model, and obtaining an optimal integrated model according to a weighting evaluation method of different model weights;

s4: and taking the optimal integrated model as a proxy model of the process model of the ecosystem, and performing proxy simulation on the basis of spatial data to generate a high-precision digital drawing product required by a research purpose under a high-resolution spatiotemporal scale.

2. The method for high-precision spatiotemporal simulation coupled with an ecological process model and a machine learning algorithm according to claim 1, wherein the step S1 is specifically as follows:

aiming at a specific research purpose, determining a research area to be simulated and a suitable ecosystem process model; acquiring observation data of the ecological system in-situ site in the research area according to the requirements of the ecological system process model, and correcting and verifying key parameters of the ecological system process model by using a differential evolution algorithm.

3. The method of high-precision spatiotemporal simulation coupling an ecological process model and a machine learning algorithm of claim 2, wherein the ecosystem in-situ site observation data comprises one or more of meteorological data, soil property data, farmland ecosystem data, ecosystem state variable data.

4. The method for high-precision spatiotemporal simulation of a coupled ecological process model and machine learning algorithm of claim 3, wherein the meteorological data includes daily temperature, precipitation and solar radiation, the soil property data includes soil organic carbon, pH and soil volume weight, the field ecosystem data includes management data and yield data for stages of seeding, fertilizing and irrigating, and the ecosystem state variable data includes soil organic carbon and greenhouse gas emission observations.

5. The method of high-precision spatiotemporal simulation of a coupled ecological process model and machine learning algorithm of claim 2, in which the key parameters include crop variety parameters, soil organic carbon and N in a field ecosystem ₂ O-exhaust process parameters.

6. The method for high-precision spatio-temporal simulation of a coupled ecological process model and a machine learning algorithm according to claim 1, wherein the step S2 is specifically as follows:

in the research area, generating spatial grids according to a set spatial resolution, randomly selecting a plurality of grid points capable of covering the climate type and soil characteristic spatial variation of the research area from all the grids, and randomly generating a management measure and climate change scene combination at each grid point; and (3) collecting and preparing soil data and historical climate data corresponding to the selected grid points, and driving the ecosystem process model verified in the step 1 to simulate interested state variables under historical and future situations.

7. The method for high-precision spatiotemporal simulation of a coupled ecological process model and machine learning algorithm of claim 6, wherein the management measure scenarios refer to various possible human activity management, in a field ecosystem, the management measure scenarios include a crop system scenario, a crop seeding stage scenario, a fertilization scenario, an irrigation scenario, and a straw returning scenario; the climate change situation refers to a future climate change situation and is selected and determined from a global climate model and a shared economic path emission situation.

8. A high-precision space-time simulation method coupling an ecological process model and a machine learning algorithm according to claim 1, wherein in the step S3, the machine learning model comprises a LASSO model, a support vector machine, a random forest, an XGBoost, a convolutional neural network, and a long-short term memory network.

9. The method for high-precision spatio-temporal simulation of coupled ecological process model and machine learning algorithm as claimed in claim 1, wherein in step S3, the root mean square error RMSE and the decision coefficient R are calculated respectively ² To evaluate the performance of each machine learning model to predict the process model output by setting a screening criterion, R ² The models with the ranking lower than the average level of each model are abandoned, and a plurality of models with the models represented on the average level are screened out; and according to the model expression, calculating the weight of each screened model, wherein the weight is 1/RMSE, and constructing a weighted average integration model to obtain an optimal integration model.

10. The method for high-precision spatio-temporal simulation of a coupled ecological process model and a machine learning algorithm according to claim 1, wherein the step S4 is specifically as follows:

and collecting and preparing high-precision spatial data, driving the constructed agent model, and realizing rapid simulation through parallel operation to generate a high-precision digital mapping product required by the research purpose.

Images