Net primary productivity estimation method based on neural network and casa model

Abstract

The application provides a net primary productivity estimation method based on a neural network and a CASA model, a UAV with a near-infrared light wave band NIR and a red light wave band R is used to take aerial photographs of a research area to obtain R, NIR and panchromatic images; the images are processed to obtain a DOM and a DSM; biomass is obtained through field sampling of the research area, so that an influence factor of the estimation area is calculated; meteorological data are processed to obtain temperature, rainfall and radiation parameters; data are trained to obtain an influence factor model of the area; and the influence factor model and the CASA model are used to process the elevation, normalized vegetation index NDVI, temperature, rainfall and radiation of each pixel to obtain net primary productivity NPP. The application can realize high-precision estimation of the net primary productivity of UAV images.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of carbon sequestration, and particularly relates to a net primary productivity estimation method based on a neural network and a CASA model. BACKGROUND

[0002] The NPP value is an important index for carbon sequestration and vegetation restoration, plays an important role in carbon sequestration research and vegetation restoration, and has important research value. NPP estimation models are mainly divided into four categories, namely, climate productivity models, physiological and ecological process models, light energy utilization rate models, and ecological remote sensing coupling models. The CASA model is a classic representative with good performance in the light energy utilization rate model, but the CASA model lacks reliable physiological and ecological basis, and there are many uncertainties in the process of light energy transmission and conversion, thereby affecting the prediction accuracy. The method calculates the actual light energy utilization rate through a neural network, and combines the CASA model, thereby improving the estimation accuracy of NPP. SUMMARY

[0003] In view of the technical problems existing in the background art, the net primary productivity estimation method based on the neural network and the CASA model provided by the application inverses the value after light energy transmission and conversion through a vegetation index, to a certain extent, makes the process more accurate, and the optimized model uses field sampling data to make the model more reliable.

[0004] In order to solve the above technical problems, the application adopts the following technical solutions to realize:

[0005] A net primary productivity estimation method based on a neural network and a CASA model, steps are as follows:

[0006] Step 1: using Pix4Dmapper software to splice unmanned aerial vehicle multi-spectral data to generate a digital surface model (DSM) and a digital orthographic image (DOM), and then using the measured elevation value, correcting the DSM through ArcGIS;

[0007] Step 2: field sampling is used to obtain the plant biomass AGB and the corresponding elevation data of the research time period through actual sampling of the estimation site;

[0008] Step 3: obtaining radiation, rainfall and temperature data through the NOAA website, the China Meteorological Data Network and the Geographic Remote Sensing Ecological Network;

[0009] Step 4: using unmanned aerial vehicle images with near-infrared light wave band NIR and red light wave band R to calculate the normalized vegetation index NDVI;

[0010] Step 5: obtaining the measured NPP through formula (1) t Data;

[0011] NPP t = AGB x 0.45 (1)

[0012] Again, the impact factor coefficient λ is obtained by formula (2)

[0013] λ = NPP t / NPP a (2)

[0014] In the formula, NPP a is the simulated value of the CASA model;

[0015] Step 6: As shown in the figure, the neural network is trained using the elevation data, NDVI and λ, and the neural network is a radial basis function neural network RBF, and a prediction model is obtained; Figure 2

[0016] Step 7: The elevation and NDVI data of the study area are brought into the prediction model to obtain the impact factor coefficient λ of the study area;

[0017] Step 8: As shown in the figure, the simulated value and the predicted value of the CASA model are used to calculate the NPP value; Figure 3 Figure 4

[0018] Step 9: As shown in the figure and Table 1, the prediction performance of the method is evaluated according to the commonly used evaluation indicators of the prediction model. Figures 5-7

[0019] Preferably, the decomposition step of step 1 is:

[0020] Step 1.1: Use a drone with a camera carrying NIR and R bands to take aerial photos of the study area;

[0021] Step 1.2: Import the aerial photos of the drone into Pix4Dmapper software;

[0022] Step 1.3: Use the automatic image correction function of Pix4Dmapper software to quickly check the drone photos;

[0023] Step 1.4: Add control points to adjust the image;

[0024] Step 1.5: Finally, perform automatic aerial triangulation, point cloud encryption to generate DOM and DSM;

[0025] Step 1.6: Use the measured elevation to correct by ArcGIS;

[0026] Preferably, the decomposition step of step 2 is:

[0027] ​​​​Step 2.1: 8 sample areas are selected in the estimated area, and 8 1m*1m quadrats are selected in each sample area. The selection rule is to randomly select and select part of the area as the sample point in the place where the vegetation is dense and sparse, so as to make the training model accurate;

[0028] Step 2.2: 25 0.2m*0.2m small quadrats are set in each 1m*1m quadrat, and 3 small quadrats are selected along the diagonal line to collect all the plant bodies;

[0029] Step 2.3: The plants are placed in an 80℃ constant temperature box to dry to constant weight, and the dry weight is weighed.

[0030] Preferably, the decomposition step of step 4 is:

[0031] Step 4.1: The estimated area is cropped using the spliced NIR and R band DOM;

[0032] Step 4.2: The cropped band image is normalized for subsequent calculation;

[0033] Step 4.3: Perform band operation calculation by formula (3);

[0034]

[0035] The normalized vegetation index NDVI is obtained.

[0036] Preferably, the decomposition step of step 6 is:

[0037] Step 6.1: Form a matrix X∈R m×n , where m represents the number of eigenvalues, and n represents the number of sample points;

[0038] Step 6.2: Divide the matrix X=[x1,x2,...,x n ] into training set and test set according to 5:5;

[0039] Step 6.3: Input layer: input the training set data into RBFNN for model training;

[0040] Step 6.4: Hidden layer: use Gaussian kernel function as activation function to activate input sequence. The connection relationship between the hidden layer and the output layer is the ordinary neural network, and the weight between them can be trained and changed,

[0041]

[0042] Where μ is the center point and σ is the radial basis width.

[0043] Step 6.5: the center mu of the RBF neural network can be changed, and its position is determined through self-organizing learning; the linear weight of the output layer is determined through supervised learning;

[0044] Step 6.6: the sequence output obtained through the hidden layer is used to obtain the simulation value of the influence factor.

[0045] Preferably, the decomposition step of step 8 is:

[0046] Step 8.1: the NDVI, temperature, rainfall and radiation data are brought into the CASA model to obtain NPP1;

[0047] Step 8.2: the simulation obtained influence factor coefficient lambda is combined with NPP1 to obtain NPP,

[0048] NPP = lambda * NPP1 (5)

[0049] Preferably, in step 9, the optimized CASA model outputs the predicted NPP value, and the mean absolute percentage error MAPE and the determination coefficient R 2 As an evaluation index, the error between the predicted value and the true value is evaluated, and the calculation formula is as follows:

[0050]

[0051]

[0052] Wherein, yi and pi are the true value and the predicted value of the i-th element respectively, the greater the value of MAPE, the greater the error of the model prediction value, and R 2 The value of R 2 is between 0 and 1, and the closer to 1, the smaller the error of the model prediction value.

[0053] The present application can achieve the following beneficial effects:

[0054] The present application combines field sampling data with the CASA model, which makes up for the problems that the CASA model lacks reliable physiological and ecological basis, and there are many uncertainties in the process of light energy transmission and conversion. Through NDVI and elevation data, the influence factor corresponding to each pixel point of the remote sensing image is calculated, the calculation method of the actual light energy utilization rate in the CASA model is changed, and the simulation accuracy of the model is improved.

[0055] Most of the existing NPP estimation methods use remote sensing data as the input source, and the resolution of satellite images is low, while the present method uses unmanned aerial image as the input source, and the resolution is higher than that of satellite image, which can better estimate the NPP of small area (such as the water-level-fluctuation zone and crop area), so as to provide basis for ecological environment monitoring and evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0056] The application will be further described below in conjunction with the accompanying drawings and examples:

[0057] Figure 1 A net primary productivity NPP estimation model framework of the application;

[0058] Figure 2 An influence factor inversion flowchart of the application;

[0059] Figure 3 A CASA optimization model prediction chart of the application;

[0060] Figure 4 A different model prediction chart of the application;

[0061] Figure 5 A comparison of predicted values and true values of different models of the application Figure 1 ;

[0062] Figure 6 A comparison of predicted values and true values of different models of the application Figure 2 ;

[0063] Figure 7 A comparison of predicted values and true values of different models of the application Figure 3 ;

[0064] Figure 8 An example NPP inversion chart of the application. DETAILED DESCRIPTION

[0065] Example 1:

[0066] A preferred scheme is shown as follows, a net primary productivity estimation method based on a neural network and a CASA model, the steps are as follows: Figures 1 to 8

[0067] Step 1: use a drone carrying a camera with NIR and R bands to take aerial photographs of the study area;

[0068] Step 2: select 8 sample areas in the estimation area, and select 8 1m*1m quadrats in each sample area, and the selection rule is to randomly select and select some areas as sample points in places with dense and sparse vegetation, so as to make the training model accurate;

[0069] Step 3: set 25 0.2m*0.2m small quadrats in each 1m*1m quadrat, select 3 small quadrats along the diagonal, and collect all the plant bodies;

[0070] Step 4: record the latitude, longitude and elevation of the sample point;

[0071] Step 5: place the plants in an 80℃ constant temperature oven to dry to constant weight, and weigh the dry weight;​

[0072] Step 6: Obtain the measured NPP by formula (1) t Data;

[0073] Step 7: Obtain the impact factor coefficient λ by formula (2)

[0074] Step 8: Import the aerial unmanned aerial vehicle pictures into Pix4Dmapper software;

[0075] Step 9: Use the automatic image correction function of Pix4Dmapper software to quickly check the unmanned aerial vehicle pictures;

[0076] Step 10: Add control points and adjust the images;

[0077] Step 11: Finally, perform automatic aerial triangulation, point cloud encryption, and generate DOM and DSM;

[0078] Step 12: Use the measured elevation to correct by ArcGIS;

[0079] Step 13: Obtain radiation, rainfall, and temperature data through the NOAA website, China Meteorological Data Network, and Geographical Remote Sensing Ecological Network;

[0080] Step 14: Perform Kriging interpolation on the radiation, rainfall, and temperature data to obtain radiation, rainfall, and temperature raster data for the study area;

[0081] Step 15: Use the spliced NIR and R band DOM to crop the estimation area;

[0082] Step 16: Perform normalization processing on the cropped band images for subsequent calculations;

[0083] Step 17: Perform band operation calculations by formula (3) to obtain the normalized vegetation index NDVI.

[0084] Step 18: Form a matrix X ∈ R m×n , where m represents the number of eigenvalues, and n represents the number of sample points.

[0085] Step 19: Randomly divide the matrix X = [x1, x2,..., xn] into training and testing sets according to 5:5; n

[0086] Step 20: Input layer: input the training set data into RBFNN for model training;

[0087] ​Step 21: Hidden layer: Use formula (4) as the activation function to activate the input sequence. The connection between the hidden layer and the output layer is the same as that of a normal neural network. The weights between them can be changed through training.

[0088] Step 22: The center μ of the RBF neural network can vary, and its position is determined through self-organization learning; the linear weights of the output layer are determined through supervised learning.

[0089] Step 23: Output the sequence obtained through the hidden layer to obtain the simulated value of the influence factor.

[0090] Step 24: Input the NDVI, temperature, rainfall, and radiation data into the CASA model to obtain NPP1;

[0091] Step 25: Combine the influence factor coefficient λ obtained from the simulation with NPP1 as shown in formula (5) to obtain NPP;

[0092] Step 26: As shown in Table 1 and Figures 5-7 As shown, the predictive performance of this method is evaluated based on commonly used evaluation indicators for prediction models, such as formulas (6) and (7).

[0093]

[0094] Table 1 Comparison of Prediction Performance of Different Models

[0095] Step 27: Read the corrected DSM and NDVI image data for the study area;

[0096] Step 28: Input the data into the impact factor prediction model to obtain the impact factor coefficient λ of the study area;

[0097] Step 29: Input the raster data of radiation, precipitation, temperature, and NDVI into the CASA model to obtain NPP1;

[0098] Step 30: Combine the simulated values ​​from the CASA model with the simulated values ​​from the impact factors to calculate the NPP value for the study area, as shown below. Figure 8 ;

[0099] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for estimating net primary productivity based on neural network and CASA model, characterized in that The method comprises the following steps: Step 1: using Pix4Dmapper software to splice unmanned aerial vehicle multispectral data, generate digital surface model and orthographic image, and then using measured elevation value, correcting the DSM through ArcGIS; Step 2: field surveying, obtaining the plant biomass AGB of the research period and the elevation data of the corresponding sample points through actual sampling of the estimated site; Step 3: obtaining radiation, rainfall and temperature data through NOAA website, China Meteorological Data Network and Geographical Remote Sensing Ecological Network; Step 4: using the unmanned aerial vehicle image with near-infrared wave band NIR and red wave band R to calculate the normalized vegetation index NDVI; Step 5: Obtain measured NPP by equation (1) t Data; NPP t = AGB x 0.45 (1) And then obtaining the influence factor coefficient λ through formula (2) λ = NPP t / NPP a (2) In the formula, NPP a is the simulated value of the CASA model; Step 6: using the elevation data, NDVI and λ to train the neural network, using radial basis neural network RBF to obtain the prediction model; Step 7: inputting the elevation and NDVI data of the research area into the prediction model to obtain the influence factor coefficient λ of the research area; Step 8: calculating the NPP value by using the simulation value and the prediction value of the CASA model; Step 9: evaluating the prediction performance of the method according to the commonly used evaluation index of the prediction model. 2.The method of estimating NPP based on neural network and CASA model according to claim 1, wherein: The decomposition steps of step 1 are as follows: Step 1.1: using the unmanned aerial vehicle with NIR and R wave band camera to take aerial photographs of the research area; Step 1.2: importing the aerial photographs of the unmanned aerial vehicle into Pix4Dmapper software; Step 1.3: using the automatic image correction function of Pix4Dmapper software to quickly check the unmanned aerial vehicle pictures; Step 1.4: adding control points to adjust the images; Step 1.5: finally, automatic aerial triangulation, point cloud encryption, generating DOM and DSM; Step 1.6: correcting through ArcGIS using the measured elevation. 3.The method of estimating NPP based on neural network and CASA model according to claim 1, wherein: The decomposition steps of step 2 are as follows: Step 2.1: selecting 8 sample areas in the estimated area, and selecting 8 1m*1m quadrats in each sample area, and selecting the rules are randomly selecting and selecting some areas as sample points in places with dense and sparse vegetation, so as to make the training model accurate; Step 2.2: setting 25 0.2m*0.2m small quadrats in each 1m*1m quadrat, selecting 3 small quadrats along the diagonal line, and collecting all the plants; Step 2.3: drying the plants in the 80℃ constant temperature box to constant weight, and weighing the dry weight. 4.The method of estimating NPP based on neural network and CASA model according to claim 1, wherein: The decomposition steps of step 4 are as follows: Step 4.1: using the spliced NIR and R wave band DOM to clip the estimated area; Step 4.2: normalizing the clipped wave band image for subsequent calculation; Step 4.3: performing wave band operation calculation through formula (3); To obtain the normalized vegetation index NDVI. 5.The method of estimating the NPP based on the neural network and the CASA model according to claim 1, wherein: The decomposition steps of step 6 are as follows: Step 6.1: Form a matrix X ∈ R m×n where m represents the number of eigenvalues, and n represents the number of sample points. Step 6.2: Randomly split the matrix X = [x1, x2,..., x n ] into training and testing sets in the ratio 5:5; Step 6.3: input layer: inputting the training set data into RBFNN for model training; Step 6.4: hidden layer: using Gaussian kernel function as activation function to activate the input sequence, the connection relationship between the hidden layer and the output layer is the ordinary neural network connection relationship, and the weight between them can be trained to change, Wherein, μ is the center point, and σ is the radial basis width. Step 6.5: The center μ of the RBF neural network can be changed, and its position is determined by self-organizing learning; the linear weight of the output layer is determined by supervised learning; Step 6.6: The sequence output obtained through the hidden layer is used to obtain the simulation value of the influence factor. 6.The method of estimating the NPP based on the neural network and the CASA model according to claim 1, wherein: The decomposition steps of step 8 are as follows: Step 8.1: The NDVI, temperature, rainfall, and radiation data are input into the CASA model to obtain NPP1; Step 8.2: The simulation obtained influence factor coefficient λ is combined with NPP1 to obtain NPP, NPP = λ × NPP1 (5). 7.The net primary productivity estimation method based on neural network and CASA model according to claim 1, wherein: In step 9, the optimized CASA model outputs the predicted NPP value, using the mean absolute percentage error (MAPE) and the coefficient of determination (R 2 As an evaluation index, the error between the predicted value and the true value is evaluated, and the calculation formula is as follows: wherein y i and p i are the real value and the predicted value of the i-th element, respectively, and the greater the value of MAPE, the greater the error of the model prediction value, and the value of R 2 is between 0 and 1, and the closer R 2 is to 1, the smaller the error of the model prediction value.

Images