A digital elevation model elevation error correction method based on particle swarm optimization random forest

Abstract

The application discloses a digital elevation model elevation error correction method based on a particle swarm optimization random forest, and belongs to the technical field of remote sensing. The particle swarm optimization random forest method is used for correction, so that the precision of the corrected digital elevation model is further improved. SRTM is selected as the digital elevation model used in the test, ICESat-2 strong light beam ground surface photon data is used as the reference elevation control point used in the test, Globeland30 is used as the global ground surface cover data used in the test, airborne LIDAR DTM published by NEON is used as the verification data used in the test. The method proposed in the application and the correction method based on polynomial regression are tested, the root mean square error is used as the verification index, the elevation error of SRTM can be effectively reduced, the elevation error of the corrected SRTM is reduced by 42% to 46% compared with the elevation error before correction, and the correction precision is better than that of the correction method based on polynomial regression.

Description

TECHNICAL FIELD

[0001] The application relates to a digital elevation model elevation error correction method based on particle swarm optimization random forest, and belongs to the technical field of remote sensing. BACKGROUND

[0002] As a digital expression of the ground elevation, the digital elevation model is widely used in the fields of geomorphology, hydrology, surveying and mapping, disaster monitoring and control and the like. However, the elevation of the digital elevation model is influenced by observation means, terrain conditions, vegetation coverage and the like, and the elevation error cannot be ignored, and the elevation accuracy is quite different in different regions.

[0003] For the elevation error of the digital elevation model, some scholars try to correct the elevation error of the digital elevation model by using various high-precision elevation data such as high-precision GPS measurement points, airborne laser radar elevation data and high-precision DEM data. However, the above elevation data are influenced by factors such as limited distribution area, great difficulty in acquisition and production, and are difficult to correct the elevation error of the SRTM in a large range and any region. The ICESat satellite height measurement data are gradually applied to the correction of the digital elevation model due to the advantages of covering the globe and high height measurement accuracy, but the ICESat is disabled in 2009, and the data are also stopped updating accordingly, and it is impossible to realize more current SRTM correction. At present, the ICESat-2 height measurement satellite is still in operation, and can provide high-precision height measurement data in a global range. Magruder et al. establish a digital elevation model elevation error correction model by using a polynomial regression method based on the ICESat-2 data and in combination with Landsat8 images. However, since the relationship between the elevation error of the digital elevation model and its influencing factors is often a complex nonlinear relationship, the polynomial regression equation expressed by simple mathematics is difficult to fully express the relationship, which causes that the elevation accuracy of the digital elevation model corrected by using the method still has great limitations.

[0004] As a machine learning algorithm that can be used to solve the nonlinear regression problem, the random forest has the advantages of high precision, strong anti-noise ability and low risk of overfitting, but the precision of the method is influenced by the set super parameter, and therefore the particle swarm optimization random forest method is used to correct the elevation error of the digital elevation model by fusing the particle swarm algorithm to find the optimal super parameter combination value of the random forest. SUMMARY

[0005] In order to realize a higher-precision digital elevation model correction result, the application provides a digital elevation model elevation error correction method based on particle swarm optimization random forest.

[0006] The technical scheme adopted by the application to achieve the above object is:

[0007] A digital elevation model elevation error correction method based on particle swarm optimization random forest, comprising the following steps:

[0008] a. Extracting reference elevation control points from satellite data, calculating the digital elevation model elevation error relative to the reference elevation control points, and extracting the corresponding latitude, longitude, terrain parameters and land cover type parameters of the reference elevation control points;

[0009] b. Constructing a digital elevation model elevation error correction model based on particle swarm optimization random forest;

[0010] c. Using the latitude, longitude, terrain parameters and land cover parameters at the reference elevation control points as the input data of the correction model, and the elevation error as the target data of the correction model, to establish a training set for the training model;

[0011] d. Training the correction model with the training set, applying the digital elevation model to the trained correction model, and correcting the elevation error.

[0012] In step a, the extracted terrain parameters are the slope Sl, aspect As and terrain relief Re of the digital elevation model, and the extracted land cover type parameters come from global land cover data Gl; for the calculation of the terrain parameters corresponding to the reference elevation control points, the bilinear interpolation method is used to extract at the reference elevation control points; for the land cover type parameters, they are directly extracted at the reference elevation control points.

[0013] In step b, the elevation error correction model is:

[0014]

[0015] In the formula, H corrected is the elevation of the corrected digital elevation model, H original is the original digital elevation model elevation, H error is the predicted digital elevation model elevation error, PSO is the particle swarm algorithm, RF is the random forest algorithm; Lat and Lon are the latitude and longitude respectively, Sl, As, Re, Gl are the slope, aspect, terrain relief and land cover type parameters respectively. Through parameter optimization of the PSO algorithm, the RF elevation error model with the optimal parameter combination is trained by using the training set, and according to the response variables [Lat, Lon, Sl, As, Re, Gl] of each SRTM pixel, it is used for prediction of the SRTM elevation error result H error , and then the obtained elevation error is combined with the original SRTM elevation H original to obtain the corrected SRTM elevation H corrected .

[0016] In step d, the specific process of model training and correction is as followsFigure 1 As shown: first, the original training set is randomly divided into 5 groups, 4 of which are used for training of the random forest model, and 1 of which is used for model precision verification, wherein the random forest model is:

[0017] H error-RF = f RF (Lat, Lon, Sl, As, Re, Gl)

[0018] The verification evaluation index adopts the mean square error regression loss (MSE):

[0019]

[0020] In the formula, N is the number of data used for verification, H error-RF is the elevation error predicted by the random forest model, is the elevation error of the digital elevation model relative to the reference elevation control point. In turn, each group of data is used for precision verification, and finally the mean of the 5 precision verification results is used as the fitness function value of the model, and the smaller the fitness function value, the higher the model precision.

[0021] Then determine the solution space range of the hyperparameters (the number of decision trees and the maximum number of selectable feature variables for node division) that affect the precision of the random forest model, and through the iterative update of the speed and position of the particles set by the particle swarm algorithm in the solution space, compare the fitness function values under different hyperparameter combinations, search for the hyperparameter combination with the smallest fitness function value, and use the entire training set to train the random forest model under the optimal hyperparameter combination.

[0022] Finally, according to the [Lat, Lon, Sl, As, Re, Gl] corresponding to each pixel of the digital elevation model, the model trained is used to predict the elevation error result H error of the pixel, and the original elevation H original of the pixel is added, so that the corrected elevation H corrected is obtained, and the elevation error correction of the digital elevation model is completed.

[0023] The beneficial effects of the present application: the present application uses the particle swarm optimization random forest method for correction to further improve the precision of the corrected digital elevation model, selects SRTM as the digital elevation model used in the test, ICESat-2 strong beam ground photon data as the reference elevation control point used in the test, Globeland30 as the global land cover data used in the test, and NEON published airborne LIDAR DTM as the verification data used in the test, and through the test of the method (PSO-RF) proposed in the present application and the correction method based on polynomial regression (PR), the root mean square error (RMSE) is used as the verification index, and the results are as follows Figure 2It is shown that the method can effectively reduce the elevation error of SRTM, and the elevation error of the corrected SRTM is reduced by 42%-46% compared with that before correction, and the correction accuracy is better than that of the correction method based on polynomial regression. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a flow chart of digital elevation model elevation error correction.

[0025] Figure 2 is the original SRTM, PR corrected SRTM and PSO-RF corrected SRTM elevation accuracy evaluation results of three survey lines. DETAILED DESCRIPTION

[0026] The specific embodiments of the present application are further explained in conjunction with the embodiments and drawings, but are not used to limit the present application.

[0027] The implementation process of the present application is to realize the correction of digital elevation model elevation error based on particle swarm optimization random forest by using computer. Taking SRTM digital elevation model, ICESat-2 satellite altimetry data and Globeland30 land cover data in a certain area as an example, the method is used to correct the elevation error of SRTM, including the following steps:

[0028] Step a, read in the ICESat-2 satellite altimetry data, select the photons classified as ground (classed_pc_flag=1) as reference elevation control points, and obtain the elevation error of SRTM relative to the reference elevation control points at the reference elevation control points

[0029] Step b, perform geographic processing on SRTM to obtain slope, slope direction and terrain relief data, use bilinear interpolation to extract the terrain parameters [Sl, As, Re] at the reference elevation control points, and directly take the value of the Globeland30 pixel where the reference elevation control point is located as the land cover type parameter Gl at the reference elevation control point.

[0030] Step c, construct a digital elevation model elevation error correction model based on particle swarm optimization random forest.

[0031] Step d, use the latitude and longitude, terrain parameters and land cover parameters at the reference elevation control points as the input data of the correction model, and the elevation error as the target data of the correction model, to establish the training set of the training model.

[0032] Step e, randomly divide the original training set into 5 groups, of which 4 groups are used for training of the random forest model, and 1 group is used for model accuracy verification, wherein the random forest model is: H error-RF = f RF(Lat, Lon, Sl, As, Re, Gl). Each set of data is used sequentially for accuracy validation. The mean of the five accuracy validation results is then used as the model's fitness function value; a smaller fitness function value indicates higher model accuracy. Next, the solution space range of the hyperparameters affecting the accuracy of the random forest model (number of decision trees and maximum number of selectable feature variables for node partitioning) is determined. The velocity and position of particles in the solution space are iteratively updated using the particle swarm optimization algorithm. The fitness function values ​​under different hyperparameter combinations are compared, and the hyperparameter combination with the smallest fitness function value is searched. This hyperparameter combination is taken as the optimal hyperparameter combination for the random forest model, and the random forest model with the optimal parameter combination is trained using the entire training set.

[0033] Step f: Using the trained model, predict the elevation error result H corresponding to each pixel in the digital elevation model based on [Lat, Lon, Sl, As, Re, Gl]. error Adding the original elevation H of the pixel original The corrected elevation H can then be obtained. corrected This completes the elevation error correction of the digital elevation model.

[0034] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A method for correcting elevation errors in a digital elevation model based on particle swarm optimization and random forest, characterized in that, The method includes the following steps: a. Extract reference elevation control points from altimetry satellite data, calculate the elevation error of the digital elevation model relative to the reference elevation control points, and extract the latitude and longitude, terrain parameters, and land cover type parameters corresponding to the reference elevation control points; b. Construct a digital elevation model based on particle swarm optimization random forest for elevation error correction. The elevation error correction model is as follows: In the formula, H corrected To correct the elevation of the digital elevation model, H original The elevation of the original digital elevation model, H error To predict the elevation error of the digital elevation model (SRTM), PSO (Particle Swarm Optimization) and RF (Random Forest) algorithms were used. Lat and Lon represent latitude and longitude, respectively, and Sl, As, Re, and Gl represent slope, aspect, topographic relief, and land cover type parameters, respectively. Through parameter optimization using the PSO algorithm, the optimal parameter combination of the RF elevation error model was trained using the training set. The response variables [Lat, Lon, Sl, As, Re, Gl] of each SRTM pixel were used to calculate the SRTM elevation error result H. error In the prediction, the obtained elevation error is then combined with the original SRTM elevation H. original The corrected SRTM elevation H was obtained. corrected ; c. Use the latitude and longitude, topographic parameters, and land cover parameters at the reference elevation control point as input data for the correction model, and the elevation error as the target data for the correction model to establish a training set for the training model. d. Train the correction model using the training set, and apply the digital elevation model to the trained correction model to correct the elevation error.

2. The elevation error correction method for a digital elevation model based on particle swarm optimization random forest according to claim 1, characterized in that, In step a, the extracted terrain parameters are the slope Sl, aspect As, and topographic relief Re of the digital elevation model, and the extracted land cover type parameters are from the global land cover data Gl. For the calculation of terrain parameters corresponding to the reference elevation control points, bilinear interpolation is used to extract them at the reference elevation control points. For the land cover type parameters, they are extracted directly at the reference elevation control points.

3. A method for correcting elevation errors in a digital elevation model based on particle swarm optimization and random forest, as described in claim 1 or 2, characterized in that... In step d, the specific process for model training and correction is as follows: First, the original training set is randomly divided into 5 groups, of which 4 groups are used for training the random forest model and 1 group is used for model accuracy verification. The random forest model is: H error-RF =f RF (Lat,Lon,Sl,As,Re,Gl) The validation and evaluation index uses the mean squared error regression loss (MSE). In the formula, N is the number of data points used for verification, and H... error-RF The elevation error predicted by the random forest model. The elevation error of the digital elevation model relative to the reference elevation control point is used. Each set of data is used for accuracy verification in turn. Finally, the mean of the five accuracy verification results is used as the fitness function value of the model. The smaller the fitness function value, the higher the model accuracy. Then, the solution space range of the hyperparameters that affect the accuracy of the random forest model is determined. The velocity and position of the particles in the solution space are iteratively updated by the particle swarm algorithm. The fitness function values ​​under different combinations of hyperparameters are compared. The hyperparameter combination with the smallest fitness function value is searched. This hyperparameter combination is taken as the optimal hyperparameter combination of the random forest model. The random forest model with the optimal parameter combination is trained using the entire training set. Finally, the trained model predicts the elevation error H of each pixel based on the [Lat, Lon, Sl, As, Re, Gl] values ​​corresponding to that pixel in the digital elevation model. error Adding the original elevation H of the pixel original The corrected elevation H is obtained. corrected This completes the elevation error correction of the digital elevation model.

4. The elevation error correction method for a digital elevation model based on particle swarm optimization random forest according to claim 1, characterized in that, In step d, the specific process for model training and correction is as follows: First, the original training set is randomly divided into 5 groups, of which 4 groups are used for training the random forest model and 1 group is used for model accuracy verification. The random forest model is: H error-RF =f RF (Lat,Lon,Sl,As,Re,Gl) The validation and evaluation index uses the mean squared error regression loss (MSE). In the formula, N is the number of data points used for verification, and H... error-RF The elevation error predicted by the random forest model. The elevation error of the digital elevation model relative to the reference elevation control point is used. Each set of data is used for accuracy verification in turn. Finally, the mean of the five accuracy verification results is used as the fitness function value of the model. The smaller the fitness function value, the higher the model accuracy. Then, the solution space range of the hyperparameters that affect the accuracy of the random forest model is determined. The velocity and position of the particles in the solution space are iteratively updated by the particle swarm algorithm. The fitness function values ​​under different combinations of hyperparameters are compared. The hyperparameter combination with the smallest fitness function value is searched. This hyperparameter combination is taken as the optimal hyperparameter combination of the random forest model. The random forest model with the optimal parameter combination is trained using the entire training set. Finally, the trained model predicts the elevation error H of each pixel based on the [Lat, Lon, Sl, As, Re, Gl] values ​​corresponding to that pixel in the digital elevation model. error Adding the original elevation H of the pixel original The corrected elevation H is obtained. corrected This completes the elevation error correction of the digital elevation model.

Images