Soil organic matter mapping feature screening method based on DS evidence theory

Abstract

The invention discloses a soil organic matter mapping feature screening method (DSFS) based on a DS evidence theory. The method is used for improving the precision and stability of soil organic matter (SOM) space prediction. According to the method, results of six feature selectors (Pearson, VIF, RFE, Boruta, RF and XGBoost) are integrated, a recognition framework is constructed, a quality function of each feature is calculated, multi-source evidences are fused by using an improved Dempster combination rule, and finally an optimized feature subset is output. Experiments show that the DSFS is obviously superior to a traditional method in six machine learning models (RF, SVR and the like), the prediction precision of the DSFS occupies the first position with the highest R2 (0.655) and the lowest RMSE (4.187 g / kg), and the DSFS shows high robustness in high-dimensional data. The method has the core advantages that multi-model decisions are integrated through the evidence theory, high-contribution remote sensing variables such as Sentinel-2 are accurately recognized, and an efficient and reliable feature screening scheme is provided for digital soil mapping.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing and geographic information, and particularly relates to a soil organic matter mapping feature screening method based on DS evidence theory. BACKGROUND

[0002] It is of great significance to obtain the spatial distribution of soil organic matter (SOM) quickly and accurately and draw a high-precision spatial distribution map for soil fertility evaluation, soil carbon pool research, precision agriculture and efficient and sustainable use of soil resources.

[0003] Traditional SOM mapping methods rely on soil researchers to conduct field soil investigation and sampling. This method requires a large number of intensive sampling points, and then experts with rich soil knowledge manually draw maps according to the information of these sampling points. This method is time-consuming, labor-intensive, inefficient, and subjective, resulting in low mapping accuracy and long mapping cycle. With the significant progress of geographic information systems, remote sensing, geostatistics, machine learning and high-performance computing capabilities, as well as the widespread use of remote sensing data, digital elevation models and environmental covariates, digital soil mapping has gradually replaced traditional mapping methods with absolute advantages of efficiency, economy and environmental protection.

[0004] Digital soil mapping is a process of generating high-resolution soil property distribution maps using remote sensing, geographic information systems, machine learning and field survey data through spatial modeling techniques. Incorporating more environmental covariates into machine learning models can increase the likelihood of finding the best variables affecting SOM, but adding too many irrelevant variables can increase model running time and reduce model prediction performance. Therefore, feature selection methods can effectively reduce redundant features in the model, identify information-rich covariates and exclude covariates that add noise or complexity to the model.

[0005] However, existing SOM mapping processes generally ignore the feature screening step or rely only on a single feature selection method, which does not have high prediction accuracy. For the above reasons, the present application proposes a soil organic matter mapping feature screening method based on DS evidence theory. SUMMARY

[0006] The present application aims to provide a soil organic matter mapping feature screening method based on DS evidence theory to provide a feature screening method with high prediction accuracy and high stability.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0008] A soil organic matter mapping feature screening method based on DS evidence theory, comprising the following steps of: using a plurality of feature selectors to screen a feature subset to establish a recognition framework, calculating a quality function of each method, and finally outputting a final feature fusion result according to a combination rule of Dempster, and the specific process is as follows:

[0009] S1, creating a recognition framework: creating a recognition framework Θ, the framework includes a complete set of mutually exclusive events, and the features screened by the feature selector are defined as the event set, and the representation is as follows:

[0010] Θ={θ1,θ2,θ3,…,θ N}

[0011] The set composed of all subsets of Θ is called the power set 2 Θ of Θ:

[0012]

[0013] Each component element is called a base element of Θ, and the power set 2 Θ is set to all single features and all feature sets, and the important features screened are X1, X2, X3, and the definition is as follows:

[0014]

[0015] S2, calculating the quality function: assigning a quality function to each element in the power set 2 Θ , if then:

[0016] m:2 Θ →[0,1]

[0017] and satisfies the following conditions:

[0018]

[0019] Where m is the quality function in 2 Θ , indicating the degree of trust of the evidence to the element X, if m(X)>0, X is called a focal element;

[0020] Sort the features screened by each feature selector according to importance from high to low, the position of a certain feature is called rank, and there are N features in total, and the absolute value of 0.999-rank / N is used to represent the basic probability, and then normalized processing is performed;

[0021] S3, evidence synthesis rule: DS evidence theory is based on the assumption that the evidences are independent of each other, a kind of evidence synthesis rule is set, the information of multiple evidence sources is fused, there are n groups of evidences, and the synthesis rule between the evidences is as follows:

[0022]

[0023] wherein, K is a conflict factor, which is a measure of the degree of conflict between two evidences;

[0024] Using the improved Dempster rule MDS, and using the squeeze method of MDS to limit the growth of the combination of propositions, the prior distribution of the characteristics in this method is set to 0.0001, and the sum of the remaining independent characteristics is 0.9999, that is , the sum of the mass functions of {X1}, {X2}, {X3} is 0.9999, and {X1, X2, X3} is 0.0001, and finally the mComb function in EvCombR is used for feature fusion and output result.

[0025] Further, the plurality of feature selectors are set to six, and the six feature selectors use Pearson filtering method, VIF filtering method, RFE wrapping method, boruta wrapping method, RF embedding method and XGBoost embedding method respectively.

[0026] Principle and beneficial effects of the technical solution: the DSFS algorithm of the present application is superior to the traditional method in terms of prediction accuracy and stability, and can provide a transferable method paradigm for complex geographical region cultivated land SOM mapping, and compared with the traditional method, the DSFS algorithm shows stronger robustness in high-dimensional data processing, and can make more accurate and comprehensive judgment and analysis on the observed object, effectively screen and retain variable information with important value. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is the best variable combination of six machine learning models;

[0028] Figure 2 is the prediction performance of six machine learning models with different data;

[0029] Figure 3 is the SOM prediction accuracy of four types of feature selection methods based on the best variable combination;

[0030] Figure 4 is the SOM prediction accuracy (R 2 ) of the machine learning model based on the best variable combination for feature selection;

[0031] Figure 5 is the SOM prediction accuracy (RMSE) of the machine learning model based on the best variable combination for feature selection;

[0032] Figure 6 is the SOM prediction accuracy of four types of feature selection methods based on all variables;

[0033] Figure 7 The prediction accuracy (R) of the SOM machine learning model based on feature selection of all variables. 2 );

[0034] Figure 8 The prediction accuracy (RMSE) of the SOM machine learning model that selects features based on all variables;

[0035] Figure 9 The average prediction accuracy of the SOM (Signal-Oriented Model) prediction method, which is based on four types of feature selection methods, including all variables and the optimal combination of variables.

[0036] Figure 10 The results of feature selection using the DSFS algorithm based on six prediction models are shown. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:

[0038] A feature selection method for soil organic matter mapping based on Dempster's evidence theory is presented. This method belongs to the ensemble method, which integrates multiple feature selectors to form a new model. To reduce the possibility of important features being incorrectly selected or missed by a single feature selection method, this invention integrates multiple feature selectors based on evidence theory to form a new DSFS algorithm. It mainly includes three steps: First, a feature subset selected by six feature selectors is used to establish the recognition framework, namely, feature subsets selected by Pearson filtering, VIF filtering, RFE wrapping, Boruta wrapping, RF embedding, and XGBoost embedding. Then, the quality function of each method is calculated. Finally, the final feature fusion result is output according to Dempster's combination rules. The detailed steps are as follows:

[0039] S1. Create a recognition framework: Create a recognition framework Θ, which includes all mutually exclusive and complete event sets. Define the features filtered by the feature selector as the event set, as shown below:

[0040] Θ = {θ1, θ2, θ3, ..., θ N}

[0041] The set consisting of all subsets of Θ is called the power set of Θ. Θ :

[0042]

[0043] Each of the constituent elements is called a primitive of Θ, and the power set 2 Θ Let the set be all individual features and the set of all their features. Let the important features after filtering be X1, X2, and X3. Then define:

[0044]

[0045] S2, Calculate the quality function: assign a quality function to each element in 2 Θ If Then:

[0046] m:2 Θ →[0,1]

[0047] And satisfy the following conditions:

[0048]

[0049] Where m is the quality function in 2 Θ The degree of trust that the evidence has in the element X, if m(X) > 0, X is called the focus element;

[0050] Sort the features selected by each feature selector in order of importance from high to low, the position of a certain feature is called rank, set the total number of features N, use the absolute value of 0.999-rank / N to represent the basic probability, and then perform normalization processing, as shown in Table 1:

[0051] Table 1 Basic probability distribution of six feature selectors

[0052]

[0053] S3, Evidence synthesis rule: DS evidence theory is based on the assumption that evidence is independent of each other, set a kind of evidence synthesis rule to fuse multiple evidence sources information, set n groups of evidence, then the synthesis rule between the evidence is as follows:

[0054]

[0055] Where, K is the conflict factor, which is a measure of the degree of conflict between two evidences; when there is conflict between two evidences, assume that one evidence shows that a factor is extremely important, and the quality function is high, while the other evidence shows that the factor is not important, even the quality function is 0, which will cause the fusion result to appear a large deviation.

[0056] The combination rule of Dempster is based on the common elements among multiple evidence sources, but the conflicting elements are ignored by the normalization factor K, so it is easy to cause paradox of the evidence combination result. In order to obtain more accurate fusion results as much as possible, the improved Dempster rule MDS is selected, which combines evidence theory and Bayesian method, allows to consider non-Bayesian prior knowledge in evidence combination, and uses the'squeeze' method of MDS to limit the combination growth of propositions, in the method, the prior distribution of the characteristics is set to 0.0001, and the sum of the remaining independent characteristics is 0.9999, that is In the sum of the mass functions of {X1}, {X2}, {X3} is 0.9999, and {X1, X2, X3} is 0.0001, and finally the mComb function in EvCombR is used for feature fusion and output of the result.

[0057] Specific implementation process:

[0058] In order to verify the stability and accuracy of the method of the application, six machine learning models are used for SOM prediction, and the six machine learning models are RF, SVR, Cubist, INLA-SPDE, XGBoost and LightGBM. On the basis of obtaining the optimal variable combination of the six machine learning models, the DSFS method of the application, Pearson filtering method, VIF filtering method, RFE wrapping method, boruta wrapping method, RF embedding method, XGBoost embedding method and RRA integration method are used for feature screening and comparison verification, so as to show the stability and accuracy of the method of the application.

[0059] I. Data and methods

[0060] Data

[0061] Soil sample data: In a certain research area, 214 surface soil samples (0-20 cm) were collected by random distribution. In each sampling point (10 m x 10 m), one soil sample was taken from each of the four corners and the center point, and then mixed thoroughly. One of the soil samples was taken as a representative sample and placed in a sealed bag with a label. The elevation and geographic coordinates of the sampling location center were determined using a portable global positioning system (GPS), and recorded on the label of the sealed bag. The samples were brought back to the laboratory for further chemical analysis. After natural air drying in the laboratory, the 214 sealed bags were cleaned of impurities, and then divided into three parts after passing through a 2 mm aperture sieve. The SOM content was determined by potassium dichromate volumetric-dilution thermal method.

[0062] Environmental covariate sources: The prediction variables used in the experiment include Sentinel-1, Sentinel-2, Landsat-8, DEM, and climate factor derived remote sensing prediction variables and topographic climatic characteristics. Among them, the Sentinel-1, Sentinel-2 and Landsat-8 environmental variables are obtained from Google Earth Engine, the DEM elevation data is downloaded from PIE-Engine, and the climate covariates are derived from the National Earth System Science Data Center. Finally, these data are processed and analyzed in ArcGIS10.8. The data source information is shown in Table 2:

[0063] Table 2 Data source information

[0064]

[0065]

[0066] The environmental variables are extracted from the obtained data, and the summary of the environmental variables is shown in Table 3:

[0067] Table 3 Summary of soil organic matter prediction environmental variables

[0068]

[0069] Method

[0070] Machine learning algorithm: The key to building a SOM prediction model is to extend the sampling point data to the surface. Six machine learning models are used for SOM prediction, namely RF, SVR, Cubist, INLA-SPDE, XGBoost and LightGBM models. By comparing the accuracy of the six models under different variable grouping, the best variable combination of each model is obtained.

[0071] Feature selection method: Feature selection can reduce redundant data in the model, improve model training speed, and improve SOM prediction accuracy. Based on the optimal variable combination of the six models, eight different variable selection methods (Pearson filtering method, VIF filtering method, RFE wrapping method, Boruta wrapping method, RF embedding method, XGBoost embedding method, RRA integration method and the DSFS integration method of the present application) are compared to verify the accuracy and stability of the DSFS selection method of the present application.

[0072] II. Best variable combination and prediction accuracy of machine learning model

[0073] The progressive variable combination strategy was adopted, and 63 environmental variables were divided into four basic datasets of terrain climate, Sentinel-1, Landsat-8, and Sentinel-2 according to data sources. Fifteen feature spaces (SC1-SC15) were constructed by systematic cross-combination. Among them, SC1 (terrain climate), SC2 (Sentinel-1), SC4 (Landsat-8), and SC8 (Sentinel-2) were single data source benchmark combinations, and the rest were multi-source data composite combinations. SC4 marked the introduction of Landsat-8 spectral features, and SC8 represented the participation of Sentinel-2 hyperspectral data in modeling. SC6, SC7, SC10, SC12, SC13, and SC15 integrated the synergistic information of optical and radar remote sensing data, among which SC15 had the largest feature dimension as a full variable set. The specific situation is shown in Table 4:

[0074] Table 4 Variable grouping

[0075]

[0076] The six machine learning algorithms of RF, SVR, Cubist, INLA-SPDE, XGBoost, and LightGBM in this experiment were based on R 2 The model performance was evaluated by two indicators of R Figure 1 Figure 1 It can be seen that the R 2 values of all models showed an increasing trend in the evolution of SC1-SC15 variable combinations, while the RMSE showed a decreasing trend, verifying the rule that the prediction accuracy is enhanced with the complexity of variable combination.

[0077] The best variable combination of different machine learning algorithm models has different characteristics: RF and LightGBM reached the peak value (R 2 = 0.632 / 0.625, RMSE = 4.325 / 4.374 g / kg) at SC13 (Sentinel-1+ Sentinel-2+ Landsat-8), and their R 2 curves showed a monotonous upward trend before SC13; Cubist reached the optimal performance (R 2 = 0.625, RMSE = 4.361 g / kg) at SC15 (full variable combination), showing an increasing pattern throughout; XGBoost, SVR, and INLA-SPDE achieved the highest accuracy (R 2 ​= 0.623 / 0.587 / 0.647, RMSE = 4.322 / 4.591 / 4.227 g / kg). In summary, the optimal combination of each model is distributed in the third gradient (SC11, SC13, SC15), and all contain high-resolution optical remote sensing data.

[0078] From the above results, it can be seen that the optimal combination of all models integrates Sentinel-2 and Landsat-8 optical remote sensing data. Further analysis found that the fusion of Sentinel-1 radar data can produce significant synergistic effect: compared with single-source optical combination SC4, SC6 combination (Sentinel-1 + Landsat-8) makes the INLA-SPDE model R 2 improved by 2.64%; while SC13 (Sentinel-1 + Sentinel-2 + Landsat-8) combination makes the RF model R 2 improved by 1.44%.

[0079] As shown in Figure 2 The prediction performance of six machine learning models with different data is shown in Figure 2 It can be seen that the SC1-SC3 combination containing only terrain climate and Sentinel-1 variables has the lowest prediction accuracy, the SC4-SC7 combination introducing Landsat-8 makes the accuracy of each model significantly improved, and the SC8-SC15 combination integrating Sentinel-2 further improves the accuracy (R 2 maximization and RMSE minimization). This confirms that optical remote sensing data (especially Sentinel-2 with 10m spatial resolution) has better representation ability than SAR data, and its high spatial resolution characteristics can effectively capture the complex surface heterogeneity of hilly and mountainous farmland.

[0080] According to the above, we can conclude that the optimal variable combination of all models contains Sentinel-2 and Landsat-8 optical remote sensing data, which confirms the core role of optical variables in SOM spatial inversion, and Sentinel-2 has a more prominent contribution due to its spatial resolution advantage.

[0081] III. Prediction accuracy and screening result analysis of feature selection model

[0082] Based on the above comparative analysis of six machine learning models under 15 variable combinations, the optimal feature combination is determined and the contribution gradient of multi-source remote sensing data to SOM spatial prediction is revealed. Through the construction of four types of eight feature selection frameworks, the stability of the accuracy of the four types of eight selection methods is evaluated.

[0083] Comparison of feature selection accuracy based on the best variable combination

[0084] Based on the optimal variable combination framework of six machine learning models (available in Section 2), this section constructs four types of feature optimization methods (including eight specific algorithms, namely Pearson filtering method, VIF filtering method, RFE wrapping method, boruta wrapping method, RF embedding method, XGBoost embedding method, RRA integration method and DSFS integration method of the present application), generates 48 feature subsets (6 models x 8 methods) by eliminating low contribution variables, and retains the full variable set (63 environmental factors) as a benchmark control. The 6x9 feature matrix (including the control group) is used for multi-model coupling verification to systematically evaluate the optimization effect of feature optimization on SOM prediction accuracy. Model performance is quantified by R 2 and RMSE dual indicators, and the accuracy distribution characteristics of different feature sets are visualized by box plots.

[0085] As Figure 3 shown in Figure 1, the SOM prediction accuracy of the four types of feature selection methods based on the optimal variable combination is shown in Figure 1, where the dashed line represents the mean baseline of the accuracy of all optimized models, providing a quantitative reference for method comparison. As can be seen from the figure, the average accuracy of the baseline model (full variable input of 63 environmental factors) in the six machine learning algorithms is R 2 = 0.603, RMSE = 4.496 g / kg. Feature selection analysis shows that the optimization performance of the filtering method (Pearson, VIF) has significant limitations, with the average R 2 only increasing by 1.66% to 0.613, and the RMSE decreasing by 1.36% to 4.435 g / kg. This phenomenon is due to the fact that the filtering method only evaluates the linear correlation between the independent variables and the SOM, and fails to capture the complex nonlinear interaction mechanisms of the soil-environment system, resulting in limited explanatory power of the feature subsets. This result confirms the applicability boundary of feature selection methods based on simple statistics in mapping modeling.

[0086] In contrast, the wrapping and embedding methods exhibit better feature selection performance. Based on the RFE and Boruta model-driven strategies of the wrapping method, the average prediction accuracy is improved to R 2 = 0.637 (5.64% higher than the baseline model), and the RMSE is reduced to 4.295 g / kg (error reduction of 4.47%). The embedding method achieves R 2 = 0.647 (7.30% improvement) and RMSE = 4.231 g / kg (5.89% error reduction) through the built-in feature importance evaluation mechanism of XGBoost and RF. Both types of methods achieve the co-evolution of feature subsets and prediction performance through the iterative optimization of the objective function during model training, and their selection mechanisms can effectively capture the nonlinear interaction effects between soil properties and environmental factors, confirming the methodological advantages of model-dependent feature optimization strategies in complex ecological modeling.

[0087] The integrated feature selection methods (DSFS, RRA) show the best prediction performance, with the average R 2 up to 0.652 (8.13% higher than the baseline) and the RMSE reduced to 4.201 g / kg (6.56% lower error). The advantage is due to the integrated strategy that effectively corrects the false negative errors in individual feature selection through the multi-model collaborative decision mechanism. The traditional packaging method and the embedding method may mistakenly delete the interactive features with high actual contribution due to the limitation of a single scoring function. Specifically, the DSFS method integrates the feature importance ranking of six basic models through the normalization factor and the Dempster combination rule, so as to improve the stable selection probability of the feature subset and significantly enhance the robustness of the feature subset. These results confirm the technical advantage of effectively mining implicit environmental driving factors and provide a method basis for high-precision SOM spatial modeling.

[0088] The performance evaluation of the feature selection method shows that the DSFS and RRA integrated methods perform best in the prediction of cultivated land SOM. As shown in Figure 3 , the DSFS model ranks first in terms of the highest R 2 (0.655) and the lowest RMSE (4.187 g / kg), with an R 2 up to 8.66% and a 6.87% lower error than the baseline model; the RRA model ranks second (R 2 = 0.650, RMSE = 4.214 g / kg), with an accuracy improvement of 7.83%. Therefore, the method of the present application has higher accuracy compared with the existing feature selection methods.

[0089] As shown in Figure 4 , the SOM prediction accuracy (R 2 ) of the machine learning model based on the best variable combination for feature selection, as shown in Figure 5 , the SOM prediction accuracy (RMSE) of the machine learning model based on the best variable combination for feature selection. From Figure 4 and Figure 5 , it can be seen that the DSFS algorithm shows excellent stability advantage in the six machine learning models. The algorithm achieves the best prediction performance (the highest R 2, the lowest RMSE), and also ranked in the top in the remaining models. Its outstanding performance comes from the multi-source information fusion mechanism under the D-S evidence theory framework, which quantifies the feature importance confidence through the probability distribution function, effectively integrates the multi-model decision boundary, and realizes the accurate identification of the key driving factors of SOM in hilly and mountainous areas. Compared with traditional methods, DSFS shows stronger robustness in high-dimensional data processing, and can make more accurate and comprehensive judgments and analysis on the observed objects, effectively filtering and retaining variable information with important value.

[0090] Comparison of feature optimization accuracy based on all features

[0091] This section builds four types of feature optimization strategies under the framework of full variable input, systematically evaluates the influence of feature subsets on model generalization ability. Through multi-dimensional screening of 63 environmental variables, 8 feature subsets are generated, based on six machine learning models to build 48 predictors (8 subsets x 6 models), and R 2 and RMSE double index quantify model performance.

[0092] As shown in Figure 6 , the prediction accuracy of SOM based on four types of feature selection methods based on all variables is shown. To reveal the optimization effect of feature selection, the full variable input model group is set as the benchmark, and the precision distribution difference is visualized by box plot. The dashed line in the figure represents the benchmark line of the average precision of the optimized model. This verification framework verifies and compares with the benchmark through multi-model coupling, and quantitatively analyzes the robustness of different feature optimization methods in SOM prediction in complex terrain.

[0093] As can be seen from Figure 6 , DSFS shows the optimal and robust prediction performance at the single model scale. Its model performance indicators R 2 , the average RMSE is 4.214 g / kg (range 0.247 g / kg). DSFS model constructs a multi-model confidence fusion framework through D-S evidence theory, and always maintains the optimal accuracy in cross-dataset testing. The feature subset selected by it has high stability, which confirms that this method can effectively overcome the interference of data heterogeneity and provide a robust feature selection scheme for SOM inversion based on multi-source remote sensing collaboration.

[0094] As shown in Figure 7 , the prediction accuracy of SOM based on six machine learning models for feature selection based on all variables (R 2 ) is shown. As shown in Figure 8 , the prediction accuracy of SOM based on six machine learning models for feature selection based on all variables (RMSE) is shown. From Figure 7 and Figure 8It can be shown that the DSFS algorithm exhibits excellent robustness. The algorithm achieves the optimal R 2 with the lowest RMSE, and consistently ranks in the top three in the remaining models, demonstrating its multi-model adaptation advantage. In contrast, the RRA has a 2.3-4.9 percentage point lower R 2 value than DSFS, revealing that its group decision mechanism is prone to ranking bias when the heterogeneity of the base model increases. This further demonstrates that the DSFS algorithm exhibits higher stability in ensemble methods, effectively overcoming the interference caused by model structure differences, and maintaining the accuracy and consistency of environmental variable identification under different data distribution conditions.

[0095] Feature optimization based on two methods

[0096] To systematically evaluate the optimization effect of feature optimization strategies on the precision of cultivated land SOM inversion, this section compares the performance differences of six prediction models under the selection of full variable set and optimal feature subset, and constructs a quantitative precision change analysis framework.

[0097] As Figure 9 shown, the average precision of SOM prediction based on all variables and the best variable combination of four types of feature selection methods. From the figure, it can be seen that the Pearson correlation coefficient method obtains an average R 2 = 0.615, RMSE = 4.424 g / kg, and after optimal variable combination screening, its R 2 increases by 1.33% to 0.623, and RMSE decreases by 1.04% to 4.378 g / kg. The performance improvement of the VIF algorithm is more significant, with R 2 increasing by 4.76% from 0.575 to 0.602, and RMSE decreasing by 3.45% from 4.653 g / kg to 4.492 g / kg. The RFE algorithm has a slight increase of 0.76% (0.640 to 0.645) in R 2 , and a decrease of 0.64% (4.280 to 4.253 g / kg) in RMSE. The RRA algorithm has an R 2 increase of 2.07% (0.637 to 0.650) and a decrease of 1.83% (4.293 to 4.214 g / kg) in RMSE. It is worth noting that the precision difference of Boruta, RF, XGBoost, and DSFS algorithms under the two types of data sets does not exceed 0.6%, because Boruta, RF, and XGBoost models highly depend on the accuracy of the prediction model itself when selecting variables, especially the DSFS algorithm, which is based on the ensemble of the above three models. 2The accuracy of DSFS is the highest and the RMSE index is stable, which proves the robustness of DSFS in high-dimensional feature space.

[0098] Feature selection results of DSFS algorithm

[0099] As Figure 10 The following is the explanation and analysis of the DSFS algorithm selection results:

[0100] Based on the best variable combination of the RF prediction model, the DSFS feature selection algorithm is used to sort the importance of all 49 variables. When the top 21 variables are selected as the input of the RF prediction model, the highest prediction accuracy is achieved. Among the 21 retained variables, 13 are from Sentinel-2, accounting for 62%; 8 are from Landsat-8, accounting for 38%. In addition, there are no Sentinel-1 related variables in the feature set.

[0101] Based on the best variable combination of the SVR prediction model, the DSFS feature selection algorithm is used to sort the importance of all 44 variables. When the top 7 variables are selected as the input of the SVR prediction model, the highest prediction accuracy is achieved. The 7 variables are: S_RVI, S_NDVI, L_RVI, S_B4, S_CI, S_BSI and S_B11. Among them, 6 are from Sentinel-2, accounting for 86%; 1 is from Landsat-8, accounting for 14%.

[0102] Based on the best variable combination of the Cubist model, the DSFS feature selection algorithm is used to sort the importance of all 63 variables. When the top 19 variables are selected as the input of the Cubist prediction model, the highest prediction accuracy is achieved. Among the 19 retained variables, 13 are from Sentinel-2, accounting for 68%; 6 are from Landsat-8, accounting for 32%. In addition, there are no terrain climate and Sentinel-1 related variables in the feature set.

[0103] Based on the best variable combination of the INLA-SPDE model, the DSFS feature selection algorithm is used to sort the importance of all 44 variables. When the top 9 variables are selected as the input of the INLA-SPDE prediction model, the highest prediction accuracy is achieved. The 9 variables are: S_RVI, S_NDVI, L_RVI, S_B4, S_CI, S_BSI, S_B11, S_B2 and S_NDRE1. Among the 9 retained variables, 8 are from Sentinel-2, accounting for 89%; 1 is from Landsat-8, accounting for 11%.

[0104] Based on the optimal variable combination of the XGBoost model, the DSFS feature screening algorithm is used to sort the importance of all 44 variables. When the top 6 variables are selected as the input of the XGBoost prediction model, the highest prediction accuracy is achieved, and the 6 variables are S_RVI, S_NDVI, L_RVI, S_B4, S_CI and S_B11. Among the 6 reserved variables, 5 are from Sentinel-2, accounting for 83%; 1 is from Landsat-8, accounting for 17%.

[0105] Based on the optimal variable combination of the LightGBM model, the DSFS feature screening algorithm is used to sort the importance of all 49 variables. When the top 7 variables are selected as the input of the LightGBM prediction model, the highest prediction accuracy is achieved, and the 7 variables are S_B11, S_BSI, S_NDVI, S_CI, L_RVI, L_NDVI and L_NDWI. In addition, among the 7 reserved variables, 1 is from Sentinel-1, accounting for 6%; 4 are from Sentinel-2, accounting for 57%; 3 are from Landsat-8, accounting for 43%. In addition, the Sentinel-1 related variables are not included in the feature set.

[0106] The above results can show that the DSFS algorithm proves its excellent feature selection ability in various prediction models (including RF, SVR, Cubist, INLA-SPDE, XGBoost and LightGBM). It can accurately identify the factors that have the greatest impact on SOM, i.e. Sentinel-2 variables, and based on these key features, the highest SOM inversion accuracy is achieved. Figure 3 It can be shown.

[0107] In summary, the DS evidence theory based integrated feature screening method of the application is superior to the traditional method in terms of prediction accuracy and stability. Compared with the traditional method, the DSFS algorithm shows stronger robustness in high-dimensional data processing, and can make more accurate and comprehensive judgments and analysis on the observed objects, effectively screening and retaining variable information with important value.

[0108] The above-mentioned is only an embodiment of the application, and the specific technical solutions or characteristics of the scheme known in the art are not described in detail. For those skilled in the art, without departing from the technical solutions of the application, a number of modifications and improvements can be made, which should be regarded as the protection scope of the application. The protection scope of the present application should be subject to the content of its claims, and the specific embodiments described in the specification can be used to explain the content of the claims.

Claims

1. A method for soil organic matter mapping feature selection based on DS evidence theory, characterized in that, The feature subset screened by using the several feature selectors is used to establish an identification framework, a quality function of each method is calculated, and finally a final feature fusion result is output according to a combination rule of Dempster, and the specific process is as follows: S1, creating an identification framework: an identification framework Θ is created, the framework includes all mutually exclusive and complete event sets, the features screened by the feature selectors are defined as the event sets, and the features are represented as follows: Θ = {θ1, θ2, θ3, …, θ N} The set of all subsets of Θ is called the power set of Θ, denoted by 2 Θ : where each constituent element is called a basis element of Θ, the power set 2 Θ Set to all individual features and all their sets of features, set the important features after screening X1, X2, X3, then define: S2, compute quality function: assign a quality function to each element of the power set 2 Θ if then: m:2 Θ →[0,1] And the following conditions are met: where m is 2 Θ The mass function m(X) in (1) represents the degree of belief of the evidence in the element X. If m(X) > 0, X is called a focal element. The features screened by each feature selector are sorted according to importance from high to low, the position of a certain feature is called rank, there are N features in total, the absolute value of 0.999-rank / N is used to represent the basic probability, and then normalization processing is performed; S3, evidence synthesis rule: the DS evidence theory is based on the assumption that the evidences are independent of each other, a kind of evidence synthesis rule is set, multiple evidence source information is fused, there are n groups of evidences, and the synthesis rule between the evidences is as follows: wherein, K is a conflict factor, which is a measure of the degree of conflict between two evidences; Using the improved Dempster rule MDS, and using the squeeze method of MDS to limit the growth of the combination of propositions, the prior distribution of the characteristics in this method is set to 0.0001, and the sum of the remaining independent characteristics is 0.9999, that is In the sum of the mass functions of {X1}, {X2}, {X3} is 0.9999, {X1, X2, X3} is 0.0001, and finally the mComb function in EvCombR is used for feature fusion and output result.

2. The soil organic matter mapping feature screening method based on DS evidence theory according to claim 1, characterized in that: The several feature selectors are set to six, and the six feature selectors use Pearson filtering method, VIF filtering method, RFE wrapping method, boruta wrapping method, RF embedding method and XGBoost embedding method respectively.

Images