CN114357885B - Photosynthetic effective radiation scattering proportion prediction method fusing multisource data - Google Patents

Abstract

The invention relates to a method for predicting the scattering ratio of photosynthetically active radiation by integrating multi-source data, and belongs to the technical field of scattering radiation ratio estimation. First, obtain the site data set, satellite data set, reanalysis data set and angle data; based on the obtained data, you can directly construct a random forest decision tree of the measured site PAR and PAR _dif and the satellite data and reanalysis data PAR and PAR _dif . According to The PAR and PAR _dif predicted by the decision tree are calculated to generate the scattered radiation ratio; or the scattered radiation ratio of the measured site, satellite data and reanalysis data is first generated based on the acquired data, and then the scattered radiation ratio of the measured site, satellite data and reanalysis data is constructed. Compared with the random forest decision tree, the scattered radiation ratio is directly predicted based on the decision tree. Both methods can solve the problem of low accuracy of scattered radiation ratio provided by existing satellite remote sensing and reanalysis products. They can also solve the problem of sparse ground observation data sites, and are more conducive to regional or global scattered radiation ratio research. .

Description

A method for predicting photosynthetically active radiation scattering ratio that integrates multi-source data

Technical field

The invention relates to a method for predicting the scattering ratio of photosynthetically active radiation by integrating multi-source data, and belongs to the technical field of scattering radiation ratio estimation.

Background technique

The scattered radiation ratio is an important indicator for studying the change of vegetation photosynthesis ability with the sun. The scattered radiation ratio can be accurately obtained based on the observation data of existing ground flux stations. However, due to the sparse distribution of the stations, the scattered radiation of the ground flux stations is used. Ratio data are not conducive to research on regional or global distribution; although satellite observation and reanalysis products can provide global scattered radiation ratio data, the scattered radiation data in satellite observation data are calculated through inversion algorithms, and reanalysis The scattered radiation ratio data in the data is also calculated through a certain model. Therefore, there will be a certain error in the scattered radiation ratio data in the satellite observation data and reanalysis data, causing the scattered radiation ratio to be low in accuracy and unable to guarantee regional or global scattered radiation. The specific accuracy is not conducive to subsequent research.

Contents of the invention

The purpose of the present invention is to provide a method for predicting photosynthetically active radiation scattering ratio that integrates multi-source data to solve the problem of low accuracy of scattered radiation ratio provided by existing satellite data and reanalysis data.

The present invention proposes a method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data. The method includes the following steps:

1) Obtain the site data set, satellite data set, reanalysis data set and angle data corresponding to the same time and space as the satellite data set and reanalysis data set; the site data set includes PAR and PAR _dif data and the corresponding site coordinates and sites Data collection time, the satellite data set includes PAR and PAR _dif data, and the reanalysis data set includes PAR and PAR _dif data;

2) Perform spatio-temporal matching on the satellite data set and reanalysis data set based on the obtained site coordinates and site data collection time; and convert the corresponding angle data in space and time after successful matching into the cosine of the solar zenith angle;

3) Calculate the corresponding scattered radiation ratio based on the PAR and PAR _dif data in the site data set, and calculate the corresponding scattered radiation ratio based on the satellite data set and reanalysis data set with successful spatiotemporal matching; combine the obtained site scattered radiation ratio, satellite The data scattering radiation ratio, the reanalysis data scattering radiation ratio and the solar zenith angle cosine corresponding to the three are used as the multi-source scattered radiation ratio training data set;

4) Input the obtained multi-source scattered radiation ratio training data set into the random forest decision tree for training, and obtain the random forest of the site scattered radiation ratio, satellite data scattered radiation ratio, reanalysis data scattered radiation ratio, and solar zenith angle cosine. Decision tree, predict the scattered radiation ratio in the time and space to be measured based on the random forest decision tree.

This invention proposes a prediction method for photosynthetically active radiation scattering ratio that integrates multi-source data and uses a random forest decision tree to establish the regression between the scattered radiation ratio data of the actual measured site and the scattered radiation ratio data of the satellite data set and the reanalysis data set. Relationship, the scattered radiation ratio is obtained. The random forest decision tree directly predicts the scattered radiation ratio under the time and space to be measured, and can obtain more accurate scattered radiation ratios under different time and space. The present invention uses a variety of scattered radiation ratio data for prediction, which avoids the existing problem of low scattered radiation ratio prediction accuracy and is more conducive to regional or global scattered radiation ratio research.

Further, the site data set is the International Flux Observation Research Network data set, the satellite data set is the CERES data set or the BESS data set, and the reanalysis data set is the MERRA2 data set.

Further, in order to ensure the accuracy of data selection and the accuracy of subsequent random forest regression model establishment, in step 3), the multi-source scattering radiation ratio training data set needs to be concentrated into the site data set, satellite data set and reanalysis data set. The scattered radiation ratio data with a scattered radiation ratio less than 0 and a scattered radiation ratio greater than 1 are eliminated.

Furthermore, in order to ensure the reliability of data training in the random forest regression model, the training method of the random forest decision tree in step 4) is the K-fold cross-validation method.

Further, in order to achieve accurate matching of the satellite data set, the reanalysis data set and the site data set, the spatiotemporal matching method of the satellite data set, the reanalysis data set and the site data set is: first obtain the spatiotemporal information of the site data set, The bilinear interpolation method or nearest neighbor method is used to extract point data in corresponding time and space from satellite data sets and reanalysis data sets.

The present invention proposes a method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data. The method includes the following steps:

1) Obtain the site data set, satellite data set, reanalysis data set and angle data in time and space corresponding to the satellite data set and reanalysis data set; the site data set includes PAR and PAR _dif data and the corresponding site coordinates and site data At acquisition time, the satellite data set includes PAR and PAR _dif data, and the reanalysis data set includes PAR and PAR _dif data;

2) Perform spatio-temporal matching on the satellite data set and reanalysis data set based on the acquired site coordinates and site data collection time; convert the corresponding angle data in space and time after successful matching into the cosine of the solar zenith angle; convert the obtained site PAR and PAR _dif , satellite data PAR and PAR _dif , reanalysis data PAR and PAR _dif and the solar zenith angle cosine corresponding to the three are used as multi-source PAR and PAR _dif training data sets;

3) Input the multi-source PAR and PAR _dif data into the random forest decision tree for training, obtain the PAR and PAR _dif random forest decision tree, and predict the PAR and PAR _dif under the time and space to be measured;

4) Calculate the scattered radiation ratio in the space and time to be measured through the predicted PAR and PAR _dif .

The present invention provides a method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data. It uses a random forest decision tree to establish the regression relationship between the PAR and PAR _dif data of the actual measurement site and the PAR and PAR _dif data in the satellite data set, and obtains PAR and PAR _dif random forest decision trees predict PAR and PAR _dif under the time and space to be measured, and then calculate the scattered radiation ratio under the time and space to be measured. This method avoids the scattered radiation ratio calculated through PAR and PAR _dif data in satellite data sets. For the problem of low accuracy, accurate scattered radiation ratio can be obtained by predicting accurate PAR and PAR _dif , which is more conducive to regional or global scattered radiation ratio research.

Further, the site data set is the International Flux Observation Research Network data set, the satellite data set is the CERES data set or the BESS data set, and the reanalysis data set is the MERRA2 data set.

Furthermore, in order to ensure the accuracy of data selection and the accuracy of subsequent random forest regression model establishment, in step 2), when performing data matching, the scattered radiation in the site data set, satellite data set and reanalysis data set needs to be The PAR and PAR _dif data corresponding to the ratio less than 0 and the scattered radiation ratio greater than 1 are eliminated.

Further, in order to ensure the reliability of data training in the random forest regression model, the training method of the random forest decision tree in step 3) is the K-fold cross-validation method.

Further, in order to achieve accurate matching of the satellite data set, the reanalysis data set and the site data set, the spatiotemporal matching method of the satellite data set, the reanalysis data set and the site data set is: first obtain the spatiotemporal information of the site data set, The bilinear interpolation method or nearest neighbor method is used to extract point data in corresponding time and space from satellite data sets and reanalysis data sets.

Description of drawings

Figure 1 is a flow chart of the method for predicting photosynthetically active radiation scattering proportion by merging multi-source data according to the present invention;

Figure 2 is a comparison chart of the mean K _d values directly predicted by the multi-product RFM model and measured at the site in Experiment 1 of the present invention;

Figure 3(a) is a box plot comparing the indicator R of the K _d results directly predicted by the multi-product and single-product models in Experiment 1 of the present invention;

Figure 3(b) is a box plot comparing the indicator MSE of the K _d results directly predicted by the multi-product and single-product models in Experiment 1 of the present invention;

Figure 3(c) is a box plot comparing the indicator RMSE of the K _d results directly predicted by the multi-product and single-product models in Experiment 1 of the present invention;

Figure 3(d) is a box plot comparing the indicator MAE of the K _d results directly predicted by the multi-product and single-product models in Experiment 1 of the present invention;

Figure 3(e) is a box plot comparing the indicator R ² of the multi-product and single-product models directly predicting K _d results in Experiment 1 of the present invention;

Figure 4 shows the K _d distribution under different vegetation types predicted in Experiment 2 of the present invention;

Figure 5 is a scatter plot of K _d and PAR predicted in Experiment 1 of the present invention;

Figure 6 is a scatter plot of K _d and PAR predicted in Experiment 2 of the present invention.

Detailed ways

The method for predicting photosynthetically active radiation scattering ratio by integrating multi-source data provided by the present invention includes constructing a K _d model to directly predict K _d and constructing a PAR and PAR _dif model to indirectly predict K _d . The present invention will be further described in detail below with reference to the accompanying drawings.

Example of directly predicting K _d :

The invention provides a method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data. The specific process for directly predicting K _d by constructing a K _d random forest regression model is shown in Figure 1. First, obtain the PAR and PAR _dif data of the three data sets: site data set, satellite data set, and reanalysis data set, as well as the angle data in space and time corresponding to the satellite data set and reanalysis data set; then based on the obtained site coordinates and During the site data collection time, space-time matching is performed on the satellite data set and the reanalysis data set; and after successful matching, the angle data in the corresponding time and space are converted into the cosine of the solar zenith angle, and then the corresponding PAR and PAR _dif data in the site data set are calculated. Scattered radiation ratio, and calculate the corresponding scattered radiation ratio based on the successfully matched satellite data set and reanalysis data set; unify the obtained site scattered radiation ratio, satellite data scattered radiation ratio, reanalysis data scattered radiation ratio and the three The corresponding cosine of the solar zenith angle is used as the multi-source scattered radiation ratio training data set; finally, the obtained multi-source scattered radiation ratio training data set is input into the random forest decision tree for training, and the site scattered radiation ratio and the satellite data scattered radiation ratio are obtained. , and then analyze the random forest decision tree of the scattering radiation ratio of the data to predict the scattering radiation ratio in the time and space to be measured, so as to solve the problem of low accuracy of the scattering radiation ratio provided by existing satellite data.

Step 1. Get data

1) Site measured data set

The site measured data obtained by this invention is the International Flux Observation Research Network Data Set (FLUXNET). FLUXNET can provide continuous measured data of a large number of surface sites since 1991, including total photosynthetically active radiation PAR (μmol m ^-2 s ^-1 ) and scattered photosynthetically active radiation PAR _dif (μmol m ^-2 s ^-1 ) data. This data set is based on half-hour time intervals. Today, the total number of FLUXNET sites has exceeded 200, which are widely distributed around the world. In this example, the PAR (μmol m ^-2 s ^-1 ) and PAR _dif (μmol m ^-2 s ^-1 ) daily data of 42 sites in the FLUXNET2015 product from 2000 to 2010 are selected, and the corresponding sites are also included. Location coordinates and data acquisition time. Among them, PAR is photosynthetically active radiation, and PAR _dif is the scattered radiation in photosynthetically active radiation.

2)Satellite data set

The satellite data set selected in this invention is at least one of the CERES data set and the BESS data set.

CERES dataset:

The Clouds and the Earth's Radiant Energy System (CERES) is the world's only project whose primary purpose is to observe the Earth's Radiation Budget (ERB) instrument and record ERB global climate data. The Earth Radiation Budget (ERB) is an indicator used to describe the amount of solar radiation absorbed by the Earth and the magnitude and difference of infrared energy emitted into space due to the Earth's climate. The CERES project provides satellite-based ERB observation data. It uses measurements from the CERES instrument, which directly measures reflected solar radiation, on several satellites, as well as data recorded by other instruments, to generate a complete set of ERB product data sets that are used in climate, weather and other disciplines. Research. In this example, the photosynthetically active radiation scattered radiation (PAR _dif (W m ^-2 ) and photosynthetically active radiation direct radiation (PAR _dir (W m ^-2 ) from 2000 to 2010 in the CERES product data set SYN1deg-Level3 are selected) The global daily data has a spatial resolution of 1° × 1°. Solve the sum of PAR _dif and PAR _dir to obtain the corresponding PAR, and finally use the PAR _dif (W m ^-2 ) in the CERES product data set and the calculated PAR ( W m ^-2 ).

BESS data set:

BESS (Breathing Earth System Simulator, BESS) is a simplified process model that combines atmospheric and canopy radiation transfer, canopy photosynthesis, transpiration and energy balance. It combines atmospheric radiative transfer models and artificial neural networks with atmospheric products from MODIS to generate 5 km daily products. In this embodiment, the PAR (mol m ^-2 d ^-1 ) and PAR _dif (mol m ^-2 d ^-1 ) data with a resolution of 5km from 2000 to 2010 in BESS_Rad are selected.

3) Reanalyze the data set

MERRA2 reanalysis data set:

MERRA2 (Modern-Era Retrospective analysis for Research and Applications, MERRA2) is a long-term series of reanalysis data sets of the United States National Aeronautics and Astronautics, produced and released by the NASA office. This data set contains a variety of meteorological variables, such as net radiation, temperature, relative humidity, wind speed, etc. At the same time, MERRA2 data covers the whole world, with a spatial resolution of 0.5° × 0.625° and a temporal resolution of 1 hour. In this example, the global hourly average data of photosynthetically active radiation scattered radiation (PAR _dif (W m ^-2 )) and photosynthetically active radiation direct radiation (PAR _dir (W m ^-2 )) in MERRA2 from 2000 to 2010 are selected. , solve the sum of PAR _dif and PAR _dir to obtain the corresponding PAR, and finally use the PAR _dif (W m ^-2 ) and calculated PAR (W m ^-2 ) in the MERRA2 product data set.

4)Angle data

The angle data obtained by the present invention is the angle data in the same time and space corresponding to the obtained site data set, satellite data set and reanalysis data set, including data acquisition time, longitude and latitude, solar altitude, and solar declination angle. Through the above obtained The data calculates the cosine of the sun's zenith angle, which is used to represent time and latitude and longitude.

Step 2. Data processing

In order to ensure the accuracy of the random forest decision tree between the subsequent establishment of the site data set, the satellite data set, and the reanalysis data set, the present invention needs to first extract the satellite data set and reanalysis data that match the site location and site data acquisition time. Set to obtain multi-source PAR and PAR _dif data after spatio-temporal matching. Furthermore, in order to ensure the accuracy of data selection, the quality of the matched data needs to be screened. In addition, the present invention also uses angle data to use the three characteristics of time and longitude and latitude with the sun. Expressed as the cosine of the zenith angle.

First, perform spatio-temporal matching on all the acquired data, and first extract the satellite data set and re-analysis data set at the same time and location of the corresponding site data set. For example, the longitude and latitude of the corresponding site BE-Bra is known, and the location of the site BE-Bra is known. The time of the acquired data is also determined. Assume that the position coordinates of BE-Bra are (A°S, B°E), and the acquisition time of one of the data is January 1, 2000. Then the matching extracted CERES data is concentrated at the position ( A°S, B°E), the time is the data on January 1, 2000. Similarly, other data matching methods are the same. In this embodiment, only 42 FLUXNET sites have observation data of PAR and PAR _dif , while CERES, BESS, and MERRA2 are all planar data, so point data needs to be extracted from the planar data. In this embodiment, double lines are first used The data set was unified to the same spatial resolution, that is, 1° × 1°, using sexual interpolation, and the nearest neighbor method was used to extract data from 42 sites in these three types of planar data. As another implementation manner, the nearest neighbor method can also be used to unify the data sets to the same resolution and extract point data in the planar data. Finally, the spatiotemporally matched FLUXNET site PAR and PAR _dif data, BESS PAR and PAR _dif data, CERES PAR and PAR _dif data, and MERRA2 reanalyzed PAR and PAR _dif data were obtained.

Since the units of PAR and PAR _dif data obtained in the station, satellite, and reanalysis data sets are not completely consistent, the units of PAR and PAR _dif of BESS are (mol m -2 d ^-1 ), and the units of PAR and PAR _dif of FLUXNET are (mol m ^-2 d -1 ). The unit is (μmol m ^-2 s ^-1 ), the unit of PAR and PAR _dif of CERES is (W m ^-2 ), and the unit of PAR and PAR _dif of MERRA2 is (W m ^-2 ). In order to ensure that the subsequent regression model is established For accuracy, all data units need to be unified to μmol m ^-2 s ^-1 . The unit conversion formula is as follows:

The PAR/PAR _dif data unit conversion formula for CERES and MERRA2 is:

1W m ^-2 = 4.56μmol m ^-2 s ^-1 (1)

The BESS PAR/PAR _dif data unit conversion formula is:

1mol m ^-2 d ^-1 =11.5741μmol m ^-2 s ^-1 (2)

Furthermore, in order to facilitate the establishment of subsequent models, the present invention expresses the three characteristics of time and longitude and latitude as the cosine of the solar zenith angle. The cosine of the solar zenith angle (cos Z) is based on the solar zenith angle and the solar azimuth angle. Calculated based on the principle of supplementary angles, the solar zenith angle is mainly corrected by annual correction, longitude correction and time correction. The revised formula is:

sin(A+Z)＝1 (5)

Where Z is the solar zenith angle, A is the solar altitude angle, h _⊙ is the solar altitude, δ is the solar declination angle, The local geographical latitude, τ is the solar hour angle at that time.

Furthermore, because the observation data of corresponding parts of the satellite data set and the reanalysis data set are missing (the result is 0 or -9999), in order to avoid errors and ensure the accuracy of data selection and subsequent model establishment, it is necessary to All data are quality screened. This invention eliminates data with a value of 0 or -9999 in PAR and PAR _dif data. At the same time, the data at the corresponding site will no longer be used for model construction.

Step 3. Generate multi-source scattered radiation ratio training data set

Calculate the corresponding scattered radiation ratio based on the PAR and PAR _dif data in the site data set, and calculate the corresponding scattered radiation ratio based on the PAR and PAR _dif in the satellite data set with successful spatiotemporal matching and the PAR and PAR _dif in the reanalysis data set. Calculate The formula is:

The obtained site scattered radiation ratio, satellite data scattered radiation ratio, reanalysis data scattered radiation ratio and the solar zenith angle cosine corresponding to the three are used as the multi-source scattered radiation ratio training data set. In this process, the site scattered radiation ratio, the satellite data scattered radiation ratio, the reanalysis data scattered radiation ratio, and the cosine of the solar zenith angle corresponding to the three are summarized. If K _d is less than or equal to 0 in any of the data sets, or greater than 1, K _d less than or equal to 0 or greater than 1 will be eliminated from this data set, and data from other data sets that match its time and space will be eliminated; the remaining site scattered radiation ratios, satellite data scattered radiation after elimination will The ratio, reanalysis data scattering radiation ratio and the solar zenith angle cosine corresponding to the three are used as the multi-source scattered radiation ratio training data set.

In this embodiment, the PAR data set obtained through step 2 has 94,948 entries, and the PAR _dif data set has 52,232 entries. Then the K _d data set calculated by comparing PAR _dif with PAR is also 52232. The K _d data of the satellite, the K _d data of the reanalysis product, the K _d data of the FLUXNET site and the solar sky corresponding to the three are unified The cosine data of the vertex angle are summarized, and K _d less than or equal to 0 or greater than 1 is eliminated. Finally, the K _d data of 39 stations that can be used and the corresponding satellite K _d data, reanalysis data K _d data and the three are obtained. The uniformly corresponding cosine of the solar zenith angle is used as a multi-source scattered radiation ratio training data set for subsequent use.

Step 4. Build a random forest decision tree and train it

Random forest is a model of the integrated learning part of machine learning. It is a statistical learning method that uses the Bootsrap resampling method to randomly repeatedly sample from the original sample, and then based on the weight of the sample features, with the maximum weight Features serve as root nodes, and branches are established sequentially to build a decision tree. Finally, the final prediction result is obtained through voting. The present invention adopts the random forest method to obtain more accurate prediction results, and avoids the problem of data set mismatch caused by overfitting.

The random forest (Random forest, RFM model) used in this invention is a model in the Sci-kitLearn library based on Python. The multi-source scattered radiation ratio training data set obtained in step 3 is input into the RFM model to realize random forest decision-making. After training the tree, the trained random forest decision tree can be used to predict K _d in any time and space, providing powerful data for subsequent regional or global scattering radiation ratio research.

During the training process, the random forest regression model randomly divides the input data set and builds a separate decision tree to obtain the prediction result. The prediction result of each decision tree is the mean value from the root node to the output leaf node. Then, the one with the most votes The result is the final prediction. Before model training, the present invention disrupts the data set to avoid overfitting of the simulation results, and uses the K-fold cross validation method for data training. In this embodiment, the model training method is the ten-fold cross-validation method, that is, the acquired training data set is divided into 10 parts, of which 9 parts are used for training and 1 part is used for verification. The training and verification data are continuously replaced until each One set of data is used for verification to avoid large differences in model accuracy due to different selections of training data. At the same time, random forest selects data sets completely randomly when training the model. The present invention monitors the learning process, and there is no serious over-fitting phenomenon, so there is no need to regularize the data set. This invention uses the grid search method to adjust hyperparameters and avoids overfitting caused by the parameters of the model itself.

In order to verify whether the K _d directly predicted by the K _d random forest regression model constructed by the algorithm of the present invention is accurate, Experiment 1 is further illustrated.

Experiment 1:

The experiment builds a single RFM model and a multi-product RFM model. The single RFM model refers to using only one kind of data and site data from CERES, BESS, and MERRA2 to build the RFM model; the multi-product model refers to using CERES, BESS, and MERRA2. The three types of data are used together with site data to construct an RFM model; the experiment compares the prediction accuracy of the multi-product RFM model with that of a single RFM model to prove the accuracy and reliability of predicting K _d using the RFM model trained with satellite data, reanalysis data and site data. . The ten-fold cross-validation method is used to verify the prediction accuracy of the model. Ten-fold cross-validation refers to dividing the data set into 10 parts (9 parts for training and 1 part for verification) when building the model, and continuously replacing the verification data until Each piece of data is used for verification; and the four statistical indicators of correlation coefficient (R), root mean square error (RMSE), absolute mean error (MAE), mean square error (MSE) and coefficient of determination (R ² ) are used To evaluate the RFM model, the range of R ² is between 0 and 1, which can be used to judge the goodness of fit of the regression model based on the sample; comparing the R ² and RMSE of the single RFM model and the multi-product RFM model, you can get the difference between them Difference, RMSE and MAE can be used to express the degree of deviation between the predicted results and the observed results.

The evaluation index results of single RFM model and multi-product RFM model obtained from the experiment are shown in Table 1:

Table 1

As can be seen from Table 1, the MSE, RMSE, and MAE of the single RFM model and the multi-product RFM model are all small. However, the comparison shows that the prediction results of the multi-product RFM model have the highest correlation with the site observation data. The average R is smaller than the average R of the multi-product RFM model. The average R ² of the three single RFM models is smaller than the average R ² of the multi-product RFM model. At the same time, comparing the MSE, RMSE, and MAE indicators of different models, we can see that the multi-product The three index values obtained by the RFM model are all smaller than the MSE, RMSE, and MAE index values obtained by the other three single RFM models. Moreover, from the comparison box plots of various indicators of the K _d results directly predicted by the multi-product and single-product models shown in Figures 3(a)-3(e), we can more intuitively see the MSE, RMSE, and The MAE indicators are all smaller than the MSE, RMSE, and MAE indicators obtained by the other three single-product RFM models. The R and R ² obtained by the multi-product RFM model are both greater than the R and R ² obtained by the other three single-product RFM models. From Table 1 and Figure The experimental results shown in 3(a)-3(e) all illustrate that the prediction results of the multi-product RFM model are better than the prediction results of the RFM model built using only one product, further proving that compared with a single RFM model, through multi-product The RFM model predicts scattered radiation with higher accuracy than At the same time, it is proved that the method of the present invention is used to directly predict the reliability and practicability of the scattered radiation ratio, and can provide strong data support for subsequent research on the scattered radiation ratio.

Since it can be concluded from Table 1 and Figures 3(a)-3(e) that the multi-product RFM model has better prediction results than the single RFM model, a multi-product RFM model was established to predict K _d and the site measured K _d for 39 sites _. The comparison chart of the mean values is shown in Figure 2. It can be found that, in addition to the unsatisfactory prediction results of some stations with too high or too low observation values, the prediction results of other stations such as CA-Qfo, DK-Sor, GF-Guy, etc. have shown good fitting properties, which may be It is the random forest regression model that has differences in predicting different regions. Generally speaking, except for some sites where the observed values are too high or too low, the prediction results of most sites fit well with the measured results, proving that the use of multi-product RFM models to directly predict K _d has practical value.

Example of indirect prediction of K _d :

The invention provides a method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data. The specific process of indirectly predicting K _d by constructing a PAR and PAR _dif random forest regression model is shown in Figure 1. First, obtain the PAR and PAR _dif data of the three data sets: site data set, satellite data set, and reanalysis data set, as well as the angle data in time and space corresponding to the satellite data set and reanalysis data set; then according to the obtained site coordinates and site During the data collection time, space-time matching is performed on the satellite data set and the reanalysis data set; and after successful matching, the corresponding space-time angle data is converted into the cosine of the solar zenith angle; the obtained site PAR and PAR _dif , satellite data PAR and PAR _dif , The reanalysis data PAR and PAR _dif and the cosine of the solar zenith angle corresponding to the three are used as the multi-source PAR and PAR _dif training data set; the multi-source PAR and PAR _dif data are input into the random forest decision tree for training, and PAR and PAR are obtained _{The dif} random forest decision tree predicts PAR and PAR _dif in the time and space to be measured; the scattered radiation ratio at the time to be measured and the coordinates to be measured is calculated through the predicted PAR and PAR _dif to solve the problem of PAR and PAR provided by existing satellite data. The low accuracy of PAR _dif causes the problem of low accuracy of K _d .

Step 1. Get data

1) Site measured data set

The site measured data obtained by this invention is the International Flux Observation Research Network Data Set (FLUXNET), as described in the embodiment of directly predicting K _d . In this example, the daily data of PAR (μmol m ^-2 s ^-1 ) and PAR _dif (μmol m ^-2 s ^-1 ) of 42 sites in the FLUXNET2015 product from 2000 to 2010 are used. Among them, PAR is photosynthetically active radiation, and PAR _dif is the scattered radiation in photosynthetically active radiation.

2)Satellite data set

The satellite data acquired by the present invention is at least one satellite data set in the CERES data set and the BESS data set, specifically as described in the embodiment of directly predicting K _d . In this example, the PAR _dif (W m ^-2 ) and the calculated PAR (W m ^-2 ) in the CERES product data set are used. The PAR (mol m ^-2 d) with a resolution of 5 km in BESS_Rad from 2000 to 2010 is used. ^-1 ) and PAR _dif (mol m ^-2 d ^-1 ).

3) Reanalyze the data set

The reanalysis data obtained by the present invention is the MERRA2 reanalysis data set, specifically as described in the embodiment of directly predicting K _d . In this embodiment, the PAR _dif (W m ^-2 ) and PAR (W m ^-2 ) data in the MERRA2 reanalysis data set are used.

4)Angle data

The angle data acquired by the present invention includes acquisition time, longitude and latitude, solar altitude, and solar declination angle. Specifically, as described in the embodiment of directly predicting K _d , the cosine of the solar altitude angle is calculated through the above acquired data to represent the time and Latitude and longitude.

Step 2. Data processing

In order to ensure the accuracy of data selection and subsequent model construction, this invention needs to first extract satellite data and re-analysis data that match the site data in space and time, and obtain multi-source PAR and PAR _dif data after spatio-temporal matching. It also needs to be unified Coordinates of PAR and PAR _dif in all site datasets, satellite datasets and reanalysis datasets. Furthermore, in order to ensure the accuracy of data selection, it is necessary to quality screen the data after spatio-temporal matching. In addition, the present invention also uses the obtained angle data to express the three characteristics of time and longitude and latitude as the cosine of the solar zenith angle. The specific methods include spatio-temporal matching of satellite data and site data, unit conversion, angle data processing, and data filtering, as described in the embodiment of directly predicting K _d . Finally, the present invention uses the PAR _dif and PAR data of 39 sites and the corresponding preprocessed satellite PAR _dif and PAR data, and then analyzes the PAR _dif and PAR data.

Step 3. Build a random forest decision tree and train it

Establish corresponding data sets based on the data obtained in step 2, and combine the PAR _dif and PAR data of the FLUXNET site, the PAR _dif and PAR data of at least one data set in the satellite data set, the PAR _dif and PAR data of the MERRA2 reanalysis data set, and the three The cosine of the solar zenith angle corresponding to each other is used as a training set, and is input into the RFM model to implement decision tree training to generate the final PAR _dif and PAR random forest decision tree, which can then predict PAR _dif and PAR in a certain region or globally. During the model training process, the K-fold cross-validation method is used for training.

Step 4. Predict the scattered radiation ratio

According to the PAR _dif and PAR predicted in step 3, the scattered radiation ratio K _d is calculated according to the predicted PAR _dif and PAR according to formula (6).

Experiment 2:

The present invention predicts and generates PAR and PAR _dif in different time and space by constructing the RFM model of PAR _dif and PAR, and then indirectly obtains the scattered radiation radiation ratio in the corresponding time and space through calculation. In order to verify the practicality of this method for indirectly predicting the scattered radiation ratio, the experiment will observe and analyze the indirect prediction results based on the K _d obtained indirectly according to the site IGBP standard classification, and obtain the K _d distribution results under different vegetation types as shown in Figure 3. In Figure 4, the star mark is the mean value of K _d , the left side of the Y-axis is the vegetation type classified by the IGBP standard, and the shaded part is the concentration interval of K _d (0.4-0.6). According to the IGBP classification standard, the vegetation types of the 266 sites in FLUXNET mainly include ((Deciduous broad-leaved forest, DBF=27), (Farmland, CRO=29), (Evergreen coniferous forest, ENF=55), (Evergreen Broadleaf forest, EBF=18), (Grassland, GRA=46), (Savanna, SAV=9), (Sparse woodland, WSA=7), (Mixed forest, MF=9), (Sparse shrubland, OSH=14), (Wetland, WET=44), (Others, OTHERS=8)) eleven types. The absorption distribution of PAR by different vegetation types varies greatly. As the number of trees increases, the mean value of K _d shows an upward trend. Comparing SAV and WSA with ENF, DBF, EBF, and MF, it can be found that the mean value increases by 25.6% ± 3.65%. Globally, the average K _d of FLUXNET sites is 0.4964, with an overall range of 0.4-0.6 (shaded part in Figure 4), which is in good agreement with the actual research results and proves that this method is used to indirectly predict the scattered radiation ratio. Reliable, it can be applied to subsequent research on scattered radiation ratio in a certain region or globally.

The above two methods (direct prediction and indirect prediction) are used to predict and generate K _d respectively. Comparing the experimental results obtained by direct prediction (Experiment 1) and indirect prediction (Experiment 2), the directly predicted K d and K _d are obtained as shown in Figure 5. Scatter plot of PAR, Figure 6 shows a scatter plot of indirectly predicted K _d versus PAR. In Figures 5 and 6, the sloped line is the linear regression of K _d , and the shaded part is the distribution area of K _d . It can be seen that the K _d predicted by Experiment 1 and Experiment 2 are concentrated between 0.4-0.6, and with the As RAR increases, K _d will slowly decrease. The trends obtained in Experiment 1 and Experiment 2 are the same. The K _d Mean predicted by Experiment 1 is 0.5125, and the K _d Mean predicted by Experiment 2 is 0.4907. The difference between the two is only 0.02, proving that regardless of Whether it is direct prediction or indirect prediction, the results obtained by these two methods are not much different, and both can be applied to the subsequent study of scattered radiation ratio.

Claims (10)

1. A method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data, characterized in that the method includes the following steps: 1) Obtain the site data set, satellite data set, reanalysis data set and angle data corresponding to the same time and space as the satellite data set and reanalysis data set; the site data set includes PAR and PAR _dif data and the corresponding site coordinates and sites Data collection time, the satellite data set includes PAR and PAR _dif data, and the reanalysis data set includes PAR and PAR _dif data; 2) Perform spatio-temporal matching on the satellite data set and reanalysis data set based on the obtained site coordinates and site data collection time; and convert the corresponding angle data in space and time after successful matching into the cosine of the solar zenith angle; 3) Calculate the corresponding scattered radiation ratio based on the PAR and PAR _dif data in the site data set, and calculate the corresponding scattered radiation ratio based on the satellite data set and reanalysis data set with successful spatiotemporal matching; combine the obtained site scattered radiation ratio, satellite The data scattering radiation ratio, the reanalysis data scattering radiation ratio and the solar zenith angle cosine corresponding to the three are used as the multi-source scattered radiation ratio training data set; 4) Input the obtained multi-source scattered radiation ratio training data set into the random forest decision tree for training, and obtain the random forest of the site scattered radiation ratio, satellite data scattered radiation ratio, reanalysis data scattered radiation ratio, and solar zenith angle cosine. Decision tree, predict the scattered radiation ratio in the time and space to be measured based on the random forest decision tree. 2. The photosynthetically active radiation scattering ratio prediction method integrating multi-source data according to claim 1, characterized in that the site data set in step 1) is an international flux observation research network data set, and the satellite The data set is the CERES data set and/or the BESS data set, and the reanalysis data set is the MERRA2 data set. 3. The photosynthetically active radiation scattering ratio prediction method integrating multi-source data according to claim 1, characterized in that in step 3), the multi-source scattering radiation ratio training data set needs to combine the site data set and the satellite data set. , the scattered radiation ratio data with a scattering radiation ratio less than 0 and a scattering radiation ratio greater than 1 are eliminated from the reanalysis data set. 4. The photosynthetically active radiation scattering ratio prediction method for integrating multi-source data according to claim 1, characterized in that the training method of the random forest decision tree in step 4) is the K-fold cross-validation method. 5. The photosynthetically active radiation scattering ratio prediction method for integrating multi-source data according to claim 1 or 3, characterized in that the spatiotemporal matching method of the satellite data set, the reanalysis data set and the site data set is: first obtain the site data The spatiotemporal information of the set is collected, and then the bilinear interpolation method or the nearest neighbor method is used to extract point data corresponding to spatiotemporal data from the satellite data set and the reanalysis data set. 6. A method for predicting the proportion of photosynthetically active radiation scattering that integrates multi-source data, characterized in that the method includes the following steps: 1) Obtain the site data set, satellite data set, reanalysis data set and angle data corresponding to the same time and space as the satellite data set and reanalysis data set; the site data set includes PAR and PAR _dif data and the corresponding site coordinates and sites Data collection time, the satellite data set includes PAR and PAR _dif data, and the reanalysis data set includes PAR and PAR _dif data; 2) Perform spatio-temporal matching on the satellite data set and reanalysis data set based on the acquired site coordinates and site data collection time; convert the corresponding angle data in space and time after successful matching into the cosine of the solar zenith angle; convert the obtained site PAR and PAR _dif , satellite data PAR and PAR _dif , reanalysis data PAR and PAR _dif and the solar zenith angle cosine corresponding to the three are used as multi-source PAR and PAR _dif training data sets; 3) Input the multi-source PAR and PAR _dif data into the random forest decision tree for training, obtain the PAR and PAR _dif random forest decision tree, and predict the PAR and PAR _dif at the time to be measured and the coordinates to be measured; 4) Calculate the scattered radiation ratio at the time to be measured and at the coordinates to be measured through the predicted PAR and PAR _dif . 7. The photosynthetically active radiation scattering ratio prediction method integrating multi-source data according to claim 6, characterized in that the site data set in step 1) is an international flux observation research network data set, and the satellite The data set is the CERES data set and/or the BESS data set, and the reanalysis data set is the MERRA2 data set. 8. The photosynthetically active radiation scattering ratio prediction method for integrating multi-source data according to claim 6, characterized in that in step 2), when performing data spatio-temporal matching, it is necessary to combine the site data set, the satellite data set and In the reanalysis data set, the PAR and PAR _dif data corresponding to the scattered radiation ratio less than 0 and the scattered radiation ratio greater than 1 are eliminated. 9. The photosynthetically active radiation scattering ratio prediction method for integrating multi-source data according to claim 6, characterized in that the training method of the random forest decision tree in step 3) is the K-fold cross-validation method. 10. The photosynthetically active radiation scattering ratio prediction method for integrating multi-source data according to claim 6 or 8, characterized in that the spatiotemporal matching method of the satellite data set, the reanalysis data set and the site data set is: first obtain the site data The spatiotemporal information of the set is collected, and then the bilinear interpolation method or the nearest neighbor method is used to extract point data corresponding to spatiotemporal data from the satellite data set and the reanalysis data set.