Quantitative method and system for constructing influence mechanism of water diversion on NDVI

Abstract

The invention provides a quantitative method and system for constructing an influence mechanism of water diversion on NDVI, and belongs to the technical field of quantitative identification of NDVI change attribution, and the method mainly comprises the steps of remote sensing image acquisition and preprocessing, NDVI data inversion, multi-source NDVI data fusion, NDVI main influence factor screening, NDVI influence mechanism construction, and future water diversion interval adjustment. According to the method, the advantages of various data sources are fused, and a more comprehensive and reliable data source is provided for construction of an NDVI influence mechanism. According to the method, influence factors are reasonably screened, and interference of secondary factors is removed, so that the reliability of a quantitative analysis result is ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of quantitative identification of NDVI change attribution, and specifically relates to a quantitative method and system for constructing a mechanism for the impact of water diversion on NDVI. Background Art

[0002] In recent years, the response mechanism of vegetation growth to a changing environment has gradually become a research hotspot. Affected by climate change and human activities, changes in desert vegetation have become complex. Quantitatively attributing the impact of climate change and human activities on desert vegetation is of great significance for the rational planning of vegetation restoration measures and timely adjustment of the scale of ecological water diversion.

[0003] The impact of desert water diversion on vegetation includes natural factors and human factors. Natural factors mainly include rainfall, temperature, wind speed, etc. These factors control the water, heat and energy conditions for vegetation growth, and then affect the respiration and photosynthesis of vegetation. Due to the harsh conditions in the Kubuqi Desert, the main human factor is the amount of ecological water diversion. NDVI (Normalized Difference Vegetation Index) is an important tool for studying changes in vegetation cover and is currently the most widely used vegetation index. It can better reflect the dynamic changes of regional underlying surface vegetation. However, studies usually use qualitative methods to analyze the impact mechanism of desert NDVI, and the factors are not comprehensive and the methods are relatively single. Other quantitative analysis methods for the impact mechanism of NDVI have certain limitations and uncertainties. The dual impact mechanism of natural factors and human factors is not clear and needs to be further improved. How to obtain a more applicable NDVI impact mechanism to provide a more accurate reference for engineering practice is particularly important. Summary of the invention

[0004] In view of this, the present invention provides a quantitative method and system for constructing the impact mechanism of water diversion on NDVI, so as to solve the problem that in the prior art, qualitative methods are usually used to analyze the impact mechanism of desert NDVI, and the factors are not comprehensive and the methods are relatively single.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A quantitative method to construct the mechanism of water diversion's impact on NDVI includes:

[0007] Step 1: Obtain remote sensing image data from multiple data sources in the region and pre-process each remote sensing image;

[0008] In step 1, the preprocessing of each remote sensing image specifically includes: performing geometric correction, radiation calibration, atmospheric correction, and stitching and cropping on each remote sensing image.

[0009] Step 2: Calculate the NDVI value of each remote sensing image after preprocessing, and calculate the annual NDVI mean of each data source in the region based on the NDVI value, and obtain the time series of the NDVI mean of each data source in the region;

[0010] In step 2, the NDVI value of each remote sensing image is calculated by the following formula:

[0011]

[0012] In the formula, ρ nir is the spectral reflectance in the near-infrared band, ρ r is the spectral reflectance in the red light band.

[0013] Step 3: (1) Use the interpolation method to fill in the missing NDVI mean values ​​in the time series of each data source, so as to process the time series of the NDVI mean values ​​of multiple data sources into equal-length sequences; According to step 2, multi-year NDVI sequences of different data sources can be obtained. Since the lengths of various NDVI sequences are different, the NDVI variation trends of different satellite remote sensing inversions are consistent, that is, the NDVI interannual variation curves are similar. Therefore, first use the most complete data (such as MODIS) as the benchmark data sequence, with the starting year recorded as t0, the data sources to be interpolated are Landsat8 (starting year recorded as t1) and Sentinel2 (starting year recorded as t2), and the common end year of the three types of data is recorded as t n , MODIS has data for both t1 and t2. Interpolation can be performed using the following formula.

[0014]

[0015] In the formula, t0→t1 and t0→t2 both represent sequences, t1-t n and t2-t n Each represents two points. ΔNDVI represents the change in NDVI during the corresponding period.

[0016] (2) Calculating the optimal weight of each data source based on the variance method and spatial characteristics; In the above (1), different satellite remote sensing data can be processed into sequences of equal length. In order to carry out multi-source satellite data fusion and achieve complementary advantages, the present invention focuses on the data quality and spatial resolution of various satellite data sources, and adopts the variance method and assigns weights based on spatial characteristics.

[0017] (2.1) The variance method is used to assign weights. The weight is the percentage of the variance of each NDVI series to the total variance. The product with the smallest variance is given the highest weight, and the product with the largest variance is given the lowest weight. The weight of Sentinel-2 is a1, the weight of Landsat8 is b1, and the weight of MODIS is c1.

[0018] (2.2) Weights are assigned based on spatial characteristics. Data sources with higher spatial resolution can usually provide more detailed information and can be assigned higher weights. The weight is the proportion of the spatial resolution of each data source. The weight of Sentinel-2 is a2, the weight of Landsat8 is b2, and the weight of MODIS is c2.

[0019] Combining the above two methods, we can consider the main characteristics of various data sources and obtain the final weight through arithmetic averaging. The weight of Sentinel-2 is denoted as a (where ), the weight of Landsat8 is b (where, ), the weight of MODIS is c (where, ).

[0020] (3) According to the optimal weights, the corresponding NDVI means in the time series of multiple data sources are fused to obtain NDVI fused data. Based on the optimal weights obtained in the above steps (1)-(2), the fused data is calculated according to the following formula.

[0021] NDVI=a·NDVI Sentinel-2 +b NDVI landsat8 +c·NDVI MODIS (4).

[0022] Step 4: Obtain the alternative influencing factors that affect the NDVI fusion data, take the NDVI fusion data as the dependent variable, and each alternative influencing factor as the independent variable, and screen out the main influencing factors of the NDVI fusion data from the alternative influencing factors based on correlation test, autocorrelation test, F test and multicollinearity test;

[0023] The climatic factors that affect desert vegetation are mainly water and heat conditions, and the human factors are mainly ecological water diversion. Therefore, first determine enough factors that can affect vegetation growth. The present invention uses precipitation (Pre), average temperature (Tem), maximum temperature (Tem-max), minimum temperature (Tem-min), wind speed (Win), relative humidity (Rhu), sunshine hours (Ssd), annual water diversion (Div), cumulative water diversion (Divs), etc. as candidate factors. Secondly, the NDVI fusion data obtained by formula (4) is used as the dependent variable, and each candidate factor is used as an independent variable to carry out a multi-method test based on the multivariate linear regression equation, so as to gradually screen the main influencing factors.

[0024] 4.1 The principle of correlation test is as follows:

[0025] The determination coefficient of the regression model is used to measure the influence of each influencing factor on the NDVI fusion data. 2 The calculation formula is as follows:

[0026]

[0027] In the formula, represents the model prediction value at time i, The mean of the long series of NDVI, y i represents the measured value of NDVI fusion data at the i-th moment, where the determination coefficient R 2 The value range is between 0 and 1, and the determination coefficient R 2 The closer the value is to 1, the stronger the explanatory power of the alternative influencing factors on the NDVI fusion data. 2 The closer it is to 0, the weaker the explanatory power of the alternative influencing factors on the NDVI fusion data.

[0028] 4.2 The principle of autocorrelation test is as follows:

[0029] In multiple linear regression, the Durbin-Watson (usually abbreviated as DW) statistic is an indicator for testing whether there is first-order autocorrelation in the residual sequence. The present invention evaluates the autocorrelation of each candidate influencing factor by calculating DW. The calculation formula is as follows:

[0030]

[0031] In the formula, e t is the residual of the tth observation (i.e., the difference between the measured value and the model predicted value), and n is the total number of observations;

[0032] Among them, a DW value close to 0 indicates that the residual sequence has a strong positive autocorrelation, a DW value close to 2 indicates that the residual sequence is close to independence, and a DW value close to 4 indicates that the residual sequence has a strong negative autocorrelation. Therefore, when DW is close to 2, it indicates that the autocorrelation test has passed.

[0033] The principle of 4.3F test sieve is as follows:

[0034] The present invention adopts F test to perform variance analysis (Analysis of Variance, ANOVA), and the principle of the method is as follows:

[0035]

[0036] In the formula, SSR is the sum of squares of deviations between the alternative influencing factors, SSE is the sum of squares of deviations within the alternative influencing factors, m represents the total number of NDVI fusion data, and k represents the number of NDVI fusion data in the training group. Then, the test statistic F is calculated, and the critical value F' corresponding to the significance level of 0.05 is determined. If F>F', the F test is passed.

[0037] 4.4 The principle of multicollinearity test is as follows:

[0038] The variance inflation coefficient is used to measure the collinearity of each candidate influencing factor. The calculation formula is as follows:

[0039]

[0040] In the formula, R 2 is the coefficient of determination. This formula is used to calculate the linear correlation between each candidate factor and other factors, thereby measuring the degree of collinearity. When the VIF value is large, it indicates that there is strong collinearity between the corresponding independent variable and other independent variables. When the VIF value is less than 10, the multicollinearity test is passed.

[0041] 4.5 Finally, by comprehensively considering the above indicators, the main influencing factors with greater impact on NDVI were screened out.

[0042] Step 5: Based on the main influencing factors of the screened NDVI fusion data, a variety of models of the relationship between NDVI and the main influencing factors are constructed in combination with linear and nonlinear theories to form a set of NDVI influencing mechanisms;

[0043] The present invention combines linear and nonlinear theories to construct a set of NDVI impact mechanisms. Linear methods include multivariate linear regression equations. Nonlinear methods include deep neural networks, random forests, and support vector machines. The principles of each method are as follows.

[0044] (5.1)Multivariate linear regression:

[0045] y=β0+β1X1+β2X2+…+β n X n (9)

[0046] Where X1, X2, …, X n represents the main influencing factors selected in step 4, β0 represents the constant of the regression equation, β1, β2, …, β n It represents the optimal weight of each main influencing factor, and y represents the NDVI result predicted by the multivariate linear regression model.

[0047] (5.2) Random Forest:

[0048] The present invention uses random forest to construct a nonlinear mechanism. Given a training set X = X1, X2, ..., X n and the response variable Y (X is the main influencing factor; Y is NDVI), Bootstrap will repeat B times (b = 1, 2, ..., B) with replacement random sampling to construct the sample set [X b , Y], and then perform tree fitting f b, before fitting, the training data is projected into a random subspace to increase the variation between different CARTs. To avoid overfitting, the final prediction of RF is defined as the average output of each CART, and the general function is:

[0049]

[0050] In the formula, is the CART calculation for the i input sample X during the pth iteration i The prediction results generated.

[0051] (5.3) Support vector machine:

[0052] The present invention adopts support vector machine to construct the impact mechanism. The basic principle is to separate the main influencing factors screened in step 4 by finding a hyperplane (the optimal hyperplane decision function is as follows) and maximize the boundary (i.e., interval) between the two categories, so as to better explain the changes in NDVI.

[0053]

[0054] Where Sgn() is the sign function; b * To determine the parameters of the optimal partitioning hyperplane; x, x i ∈R N is an N-dimensional vector, x is a point on the hyperplane, x i is a sample data set, (x·x i ) is the dot product of two vectors; y i ∈{1, 2, ..., k} is the k-class partition.

[0055] (5.4) Deep Neural Networks

[0056] The set of influencing mechanisms proposed in the present invention includes a deep neural network. This method imitates the structure and working principle of the human brain neural network. Through hierarchical feature learning and weight adjustment, high-performance solutions for complex tasks can be achieved. The influencing factors selected in step 4 are input into the model to stimulate the operation of the neural network. Each neuron layer receives the output of the previous layer as input and calculates the output through a series of nonlinear transformations and weight adjustments. Finally, training is performed through the back propagation algorithm, that is, by calculating the error between the predicted output and the actual output, and using the gradient descent method to update the weights and bias values ​​in the network until the network reaches a predetermined performance level, thereby better revealing the changes in NDVI and achieving efficient simulation.

[0057] Step 6: Based on the NDVI fusion data, comprehensively evaluate the performance of various models in the NDVI impact mechanism set using Pearson correlation coefficient, root mean square error, and mean absolute error metrics to obtain the final NDVI impact mechanism set;

[0058] In the present invention, various models (multiple linear regression, random forest, support vector machine, deep neural network) included in the impact mechanism set are divided into a training set (the first 80% of the data sequence) and a validation set (the last 20% of the data sequence). The performance of the models is evaluated using Pearson correlation coefficient (PCC), root mean square error (RMSE), and mean absolute error (MAE) for the validation set. The principles of various methods are as follows.

[0059] (6.1) Pearson correlation coefficient (PCC)

[0060] The Pearson correlation coefficient reflects the strength of the linear relationship between the measured NDVI values and the model-simulated values. The absolute value of its range is from 0 to 1. The closer it is to 1, the better the performance of the model.

[0061] Generally, 0.8 < PCC ≤ 1.0 means very strong correlation; 0.6 < PCC ≤ 0.8 means strong correlation; 0.4 < PCC ≤ 0.6 means medium correlation; 0.2 < PCC ≤ 0.4 means weak correlation; 0.0 ≤ PCC ≤ 0.2 indicates very weak or no correlation; PCC ≤ 0.0 indicates negative correlation.

[0062]

[0063] (6.2) Root mean square error (RMSE)

[0064] The root mean square error is used to evaluate the deviation degree between the measured NDVI values and the model-simulated values. Its value is always non-negative. The smaller the value, the smaller the error, and vice versa.

[0065]

[0066] (6.3) Mean absolute error (MAE)

[0067] The mean absolute error (MAE) is used to evaluate the difference between the measured NDVI values and the model-simulated values and measure the magnitude of the average error. The mean absolute error can avoid the problem of error cancellation and thus can accurately reflect the actual error magnitude.

[0068]

[0069] In formulas (12)-(14), N represents the data volume of NDVI, S i represents the model-simulated value at the i-th moment, and O i represents the measured NDVI value at the i-th moment. represents the mean of the NDVI measured series, Represents the mean of the NDVI simulation series.

[0070] Step 7: Input the future meteorological data and the NDVI data of the corresponding period into the NDVI influencing mechanism set determined in step 6, and calculate the future annual water diversion corresponding to different models.

[0071] A quantitative system for constructing the mechanism of water diversion's impact on NDVI, including:

[0072] Preprocessing module: obtain remote sensing image data from multiple data sources in the region and preprocess each remote sensing image;

[0073] Calculation module: Calculate the NDVI value of each remote sensing image after preprocessing, and calculate the annual NDVI mean of each data source in the region based on the NDVI value, and obtain the time series of the NDVI mean of each data source in the region;

[0074] Fusion module: Use interpolation method to fill in the missing NDVI mean values ​​in the time series of each data source, so as to process the time series of NDVI mean values ​​of multiple data sources into equal-length sequences;

[0075] Calculate the optimal weight of each data source based on variance method and spatial characteristics;

[0076] According to the optimal weight, the corresponding NDVI means in the time series of multiple data sources are fused to obtain NDVI fused data;

[0077] Screening module: Obtain the alternative influencing factors that affect the NDVI fusion data, take the NDVI fusion data as the dependent variable, and each alternative influencing factor as the independent variable, and screen out the main influencing factors of the NDVI fusion data from the alternative influencing factors based on correlation test, autocorrelation test, F test and multicollinearity test;

[0078] Model building module: Based on the main influencing factors of the screened NDVI fusion data, a variety of models of the relationship between NDVI and the main influencing factors are constructed in combination with linear and nonlinear theories to form a set of NDVI influencing mechanisms;

[0079] Evaluation module: Based on NDVI fusion data, the Pearson correlation coefficient, root mean square error, and mean absolute error indicators are used to comprehensively evaluate the performance of various models of the NDVI impact mechanism set to obtain the final NDVI impact mechanism set;

[0080] Water diversion acquisition module: input future meteorological data and NDVI data of the corresponding period into the NDVI impact mechanism set determined by the evaluation module, and calculate the future annual water diversion corresponding to different models.

[0081] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0082] (1) In the present invention, the advantages of multiple data sources are integrated to provide a more comprehensive and reliable data source for the construction of the NDVI impact mechanism.

[0083] (2) In the present invention, the influencing factors are rationally screened and the interference of minor factors is removed, thereby ensuring the reliability of the quantitative analysis results.

[0084] (3) The “influence mechanism set” constructed by the present invention can combine linear and nonlinear methods to fully explore the relationship between NDVI and its main influencing factors with high accuracy.

[0085] (4) In the present invention, the future water diversion range can be obtained based on the mechanism set, thereby providing a reference for the development of engineering practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0087] Figure 1 is a flow chart of the present invention;

[0088] Figure 2 is a flow chart of step 3 of the present invention;

[0089] Figure 3 This is a schematic diagram of the research area in Example 1 of the present invention;

[0090] Figure 4 This is a schematic diagram of the interannual variation of the NDVI fusion results in Example 1 of the present invention;

[0091] Figure 5 Schematic diagram of principal component regression test results in Example 1 of the present invention. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0093] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0094] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0095] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0096] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0097] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0098] Example 1

[0099] like Figure 1-4 As shown, an embodiment of the present invention discloses a quantitative method for constructing a mechanism of the impact of water diversion on NDVI, including:

[0100] Step 1: Obtain remote sensing image data from multiple data sources in the region and pre-process each remote sensing image;

[0101] In step 1, the preprocessing of each remote sensing image specifically includes: performing geometric correction, radiation calibration, atmospheric correction, and stitching and cropping on each remote sensing image.

[0102] Step 2: Calculate the NDVI value of each remote sensing image after preprocessing, and calculate the annual NDVI mean of each data source in the region based on the NDVI value, and obtain the time series of the NDVI mean of each data source in the region;

[0103] In step 2, the NDVI value of each remote sensing image is calculated by the following formula:

[0104]

[0105] In the formula, ρ nir is the spectral reflectance in the near-infrared band, ρ r is the spectral reflectance in the red light band.

[0106] Step 3: (1) Use the interpolation method to fill in the missing NDVI mean values ​​in the time series of each data source, so as to process the time series of the NDVI mean values ​​of multiple data sources into equal-length sequences; According to step 2, multi-year NDVI sequences of different data sources can be obtained. Since the lengths of various NDVI sequences are different, the NDVI variation trends of different satellite remote sensing inversions are consistent, that is, the NDVI interannual variation curves are similar. Therefore, first use the most complete data (such as MODIS) as the benchmark data sequence, with the starting year recorded as t0, the data sources to be interpolated are Landsat8 (starting year recorded as t1) and Sentinel2 (starting year recorded as t2), and the common end year of the three types of data is recorded as t n , MODIS has data for both t1 and t2. Interpolation can be performed using the following formula.

[0107]

[0108] In the formula, t0→t1 and t0→t2 both represent sequences, t1-t n and t2-t n Each represents two points. ΔNDVI represents the change in NDVI during the corresponding period.

[0109] (2) Calculating the optimal weight of each data source based on the variance method and spatial characteristics; In the above (1), different satellite remote sensing data can be processed into sequences of equal length. In order to carry out multi-source satellite data fusion and achieve complementary advantages, the present invention focuses on the data quality and spatial resolution of various satellite data sources, and adopts the variance method and assigns weights based on spatial characteristics.

[0110] (2.1) The variance method is used to assign weights. The weight is the percentage of the variance of each NDVI series to the total variance. The product with the smallest variance is given the highest weight, and the product with the largest variance is given the lowest weight. The weight of Sentinel-2 is a1, the weight of Landsat8 is b1, and the weight of MODIS is c1.

[0111] (2.2) Weights are assigned based on spatial characteristics. Data sources with higher spatial resolution can usually provide more detailed information and can be assigned higher weights. The weight is the proportion of the spatial resolution of each data source. The weight of Sentinel-2 is a2, the weight of Landsat8 is b2, and the weight of MODIS is c2.

[0112] Combining the above two methods, we can consider the main characteristics of various data sources and obtain the final weight through arithmetic averaging. The weight of Sentinel-2 is denoted as a (where ), the weight of Landsat8 is b (where, ), the weight of MODIS is c (where, ).

[0113] (3) According to the optimal weights, the corresponding NDVI means in the time series of multiple data sources are fused to obtain NDVI fused data. Based on the optimal weights obtained in the above steps (1)-(2), the fused data is calculated according to the following formula.

[0114] NDVI=a·NDVI Sentinel-2 +b NDVI landsat8 +c·NDVI MODIS (4).

[0115] Step 4: Obtain the alternative influencing factors that affect the NDVI fusion data, take the NDVI fusion data as the dependent variable, and each alternative influencing factor as the independent variable, and screen out the main influencing factors of the NDVI fusion data from the alternative influencing factors based on correlation test, autocorrelation test, F test and multicollinearity test;

[0116] The climatic factors that affect desert vegetation are mainly water and heat conditions, and the human factors are mainly ecological water diversion. Therefore, first determine enough factors that can affect vegetation growth. The present invention uses precipitation (Pre), average temperature (Tem), maximum temperature (Tem-max), minimum temperature (Tem-min), wind speed (Win), relative humidity (Rhu), sunshine hours (Ssd), annual water diversion (Div), cumulative water diversion (Divs), etc. as candidate factors. Secondly, the NDVI fusion data obtained by formula (4) is used as the dependent variable, and each candidate factor is used as an independent variable to carry out a multi-method test based on the multivariate linear regression equation, so as to gradually screen the main influencing factors.

[0117] 4.1 The principle of correlation test is as follows:

[0118] The determination coefficient of the regression model is used to measure the influence of each influencing factor on the NDVI fusion data. 2 The calculation formula is as follows:

[0119]

[0120] In the formula, represents the model prediction value at time i, The mean of the long series of NDVI, yi represents the measured value of NDVI fusion data at the i-th moment, where the determination coefficient R 2 The value range is between 0 and 1, and the determination coefficient R 2 The closer the value is to 1, the stronger the explanatory power of the alternative influencing factors on the NDVI fusion data. 2 The closer it is to 0, the weaker the explanatory power of the alternative influencing factors on the NDVI fusion data.

[0121] 4.2 The principle of autocorrelation test is as follows:

[0122] In multiple linear regression, the Durbin-Watson (usually abbreviated as DW) statistic is an indicator for testing whether there is first-order autocorrelation in the residual sequence. The present invention evaluates the autocorrelation of each candidate influencing factor by calculating DW. The calculation formula is as follows:

[0123]

[0124] In the formula, e t is the residual of the tth observation (i.e., the difference between the measured value and the model predicted value), and n is the total number of observations;

[0125] Among them, a DW value close to 0 indicates that the residual sequence has a strong positive autocorrelation, a DW value close to 2 indicates that the residual sequence is close to independence, and a DW value close to 4 indicates that the residual sequence has a strong negative autocorrelation. Therefore, when DW is close to 2, it indicates that the autocorrelation test has passed.

[0126] The principle of 4.3F test sieve is as follows:

[0127] The present invention adopts F test to carry out analysis of variance (Analysis of Variance, ANOVA), and the principle of the method is as follows:

[0128]

[0129] In the formula, SSR is the sum of squares of deviations between the alternative influencing factors, SSE is the sum of squares of deviations within the alternative influencing factors, m represents the total number of NDVI fusion data, and k represents the number of NDVI fusion data in the training group. Then, the test statistic F is calculated, and the critical value F' corresponding to the significance level of 0.05 is determined. If F>F', the F test is passed.

[0130] 4.4 The principle of multicollinearity test is as follows:

[0131] The variance inflation coefficient is used to measure the collinearity of each candidate influencing factor. The calculation formula is as follows:

[0132]

[0133] In the formula, R 2 is the coefficient of determination. This formula is used to calculate the linear correlation between each candidate factor and other factors, thereby measuring the degree of collinearity. When the VIF value is large, it indicates that there is strong collinearity between the corresponding independent variable and other independent variables. When the VIF value is less than 10, the multicollinearity test is passed.

[0134] 4.5 Finally, by comprehensively considering the above indicators, the main influencing factors with greater impact on NDVI were screened out.

[0135] Step 5: Based on the main influencing factors of the screened NDVI fusion data, a variety of models of the relationship between NDVI and the main influencing factors are constructed in combination with linear and nonlinear theories to form a set of NDVI influencing mechanisms;

[0136] The present invention combines linear and nonlinear theories to construct a set of NDVI impact mechanisms. Linear methods include multivariate linear regression equations. Nonlinear methods include deep neural networks, random forests, and support vector machines. The principles of each method are as follows.

[0137] (5.1)Multivariate linear regression:

[0138] y=β0+β1X1+β2X2+…+β n X n (9)

[0139] Where X1, X2, …, X n represents the main influencing factors selected in step 4, β0 represents the constant of the regression equation, β1, β2, …, β n It represents the optimal weight of each main influencing factor, and y represents the NDVI result predicted by the multivariate linear regression model.

[0140] (5.2) Random Forest:

[0141] The present invention uses random forest to construct a nonlinear mechanism. Given a training set X = X1, X2, ..., X n and the response variable Y (X is the main influencing factor; Y is NDVI), Bootstrap will repeat B times (b = 1, 2, ..., B) with replacement random sampling to construct the sample set [X b , Y], and then perform tree fitting f b , before fitting, the training data is projected into a random subspace to increase the variation between different CARTs. To avoid overfitting, the final prediction of RF is defined as the average output of each CART, and the general function is:

[0142]

[0143] In the formula, is the CART calculation for the i input sample X during the pth iteration i The prediction results generated.

[0144] (5.3) Support vector machine:

[0145] The present invention adopts support vector machine to construct the impact mechanism. The basic principle is to separate the main influencing factors screened in step 4 by finding a hyperplane (the optimal hyperplane decision function is as follows) and maximize the boundary (i.e., interval) between the two categories, so as to better explain the changes in NDVI.

[0146]

[0147] Where Sgn() is the sign function; b * To determine the parameters of the optimal partitioning hyperplane; x, x i ∈R N is an N-dimensional vector, x is a point on the hyperplane, x i is a sample data set, (x·x i ) is the dot product of two vectors; y i ∈{1, 2, ..., k} is the k-class partition.

[0148] (5.4) Deep Neural Networks

[0149] The set of influencing mechanisms proposed in the present invention includes a deep neural network. This method imitates the structure and working principle of the human brain neural network. Through hierarchical feature learning and weight adjustment, high-performance solutions for complex tasks can be achieved. The influencing factors selected in step 4 are input into the model to stimulate the operation of the neural network. Each neuron layer receives the output of the previous layer as input and calculates the output through a series of nonlinear transformations and weight adjustments. Finally, training is performed through the back propagation algorithm, that is, by calculating the error between the predicted output and the actual output, and using the gradient descent method to update the weights and bias values ​​in the network until the network reaches a predetermined performance level, thereby better revealing the changes in NDVI and achieving efficient simulation.

[0150] Step 6: Based on the NDVI fusion data, the Pearson correlation coefficient, root mean square error, and mean absolute error indicators are used to comprehensively evaluate the performance of various models of the NDVI impact mechanism set to obtain the final NDVI impact mechanism set;

[0151] The present invention divides various models (multiple linear regression, random forest, support vector machine, deep neural network) incorporated into the impact mechanism set into a training set (the first 80% of the data sequence) and a validation set (the last 20% of the data sequence), and the performance of the models is evaluated using the Pearson correlation coefficient (PCC), root mean square error (RMSE), and mean absolute error (MAE) for the validation set. The principles of various methods are as follows.

[0152] (6.1) Pearson correlation coefficient (PCC)

[0153] The Pearson correlation coefficient reflects the strength of the linear relationship between the measured NDVI values and the model simulation values. The absolute value of it ranges from 0 to 1. The closer it is to 1, the better the performance of the model.

[0154] Generally, 0.8 < PCC ≤ 1.0 means very strong correlation; 0.6 < PCC ≤ 0.8 means strong correlation; 0.4 < PCC ≤ 0.6 means medium correlation; 0.2 < PCC ≤ 0.4 means weak correlation; 0.0 ≤ PCC ≤ 0.2 indicates very weak or no correlation; PCC ≤ 0.0 indicates negative correlation.

[0155]

[0156] (6.2) Root mean square error (RMSE)

[0157] The root mean square error is used to evaluate the deviation degree between the measured NDVI values and the model simulation values. Its value is always non - negative. The smaller the value, the smaller the error, and vice versa.

[0158]

[0159] (6.3) Mean absolute error (MAE)

[0160] The mean absolute error (MAE) is used to evaluate the difference between the measured NDVI values and the model simulation values and measure the magnitude of the average error. The mean absolute error can avoid the problem of error cancellation and thus can accurately reflect the actual error magnitude.

[0161]

[0162] In equations (12) - (14), N represents the amount of NDVI data, S i represents the model simulation value at time i, and O i represents the measured NDVI value at time i. represents the mean of the measured NDVI sequence, represents the mean of the NDVI simulation sequence.

[0163] Step 7: Input the future meteorological data and the NDVI data of the corresponding period into the NDVI influencing mechanism set determined in step 6, and calculate the future annual water diversion corresponding to different models.

[0164] Example 2

[0165] This embodiment proposes a quantitative system for constructing the mechanism of water diversion's impact on NDVI, including:

[0166] Preprocessing module: obtain remote sensing image data from multiple data sources in the region and preprocess each remote sensing image;

[0167] Calculation module: Calculate the NDVI value of each remote sensing image after preprocessing, and calculate the annual NDVI mean of each data source in the region based on the NDVI value, and obtain the time series of the NDVI mean of each data source in the region;

[0168] Fusion module: Use interpolation method to fill in the missing NDVI mean values ​​in the time series of each data source, so as to process the time series of NDVI mean values ​​of multiple data sources into equal-length sequences;

[0169] Calculate the optimal weight of each data source based on variance method and spatial characteristics;

[0170] According to the optimal weight, the corresponding NDVI means in the time series of multiple data sources are fused to obtain NDVI fused data;

[0171] Screening module: Obtain the alternative influencing factors that affect the NDVI fusion data, take the NDVI fusion data as the dependent variable, and each alternative influencing factor as the independent variable, and screen out the main influencing factors of the NDVI fusion data from the alternative influencing factors based on correlation test, autocorrelation test, F test and multicollinearity test;

[0172] Model building module: Based on the main influencing factors of the screened NDVI fusion data, a variety of models of the relationship between NDVI and the main influencing factors are constructed in combination with linear and nonlinear theories to form a set of NDVI influencing mechanisms;

[0173] Evaluation module: Based on NDVI fusion data, the Pearson correlation coefficient, root mean square error, and mean absolute error indicators are used to comprehensively evaluate the performance of various models of the NDVI impact mechanism set to obtain the final NDVI impact mechanism set;

[0174] Water diversion acquisition module: input future meteorological data and NDVI data of the corresponding period into the NDVI impact mechanism set determined by the evaluation module, and calculate the future annual water diversion corresponding to different models.

[0175] Example 3

[0176] like Figure 1-4 As shown, this embodiment is a specific implementation of embodiment 1.

[0177] 1. Remote sensing image acquisition and preprocessing. In order to cope with the threat of ice and flood control of the Yellow River and the ecological threat of the Kubuqi Desert, Ordos has explored effective ecological restoration measures. Since 2014, part of the ice and flood water of the Yellow River has been introduced into the low-lying areas on the northern edge of the Kubuqi Desert to alleviate the pressure of ice and flood control during the ice and flood season, and at the same time improve the ecological environment of the northern edge of the Kubuqi Desert (ecological management area). Based on this background, this embodiment takes the Kubuqi Desert Ecological Management Area as the research area (geographical location such as Figure 3 Acquire Sentinel-2, Landsat8, and MODIS satellite images before and after water diversion (2004-2023), and carry out preprocessing work such as geometric correction, radiometric calibration, atmospheric correction, and stitching and cropping.

[0178] 2. NDVI data inversion. According to the bands of different remote sensing images, the NDVI is calculated using formula (1). At the same time, the mean NDVI value of the vegetation growing season (June-September) in the study area is calculated, and the time series of the mean NDVI value before and after water diversion (2004-2023) is obtained. The results are shown in the following table. Among them, the data of Sentinel-2 (spatial resolution of 10 meters) starts in 2016, the data of Landsat8 (spatial resolution of 30 meters) starts in 2013, and the data of MODIS (spatial resolution of 1000 meters) starts in 2004.

[0179] Table 1 Inversion results of multi-source satellite remote sensing NDVI data

[0180]

[0181]

[0182] 3. Multi-source NDVI data fusion. The implementation process of this step is as follows Figure 2 shown.

[0183] (1) Interpolation of NDVI sequences of equal length. According to step 2, multi-year NDVI sequences from different data sources can be obtained. Since the time lengths of the data sources are different, the NDVI trends and change characteristics inverted from different satellite data sources are basically the same. Therefore, the following method is used for interpolation: First, the most complete data (such as MODIS) is used as the benchmark data sequence, with the starting year marked as t0. The data sources to be interpolated are Landsat8 (the starting year is marked as t1) and Sentinel2 (the starting year is marked as t2). The common end year of the three types of data is marked as t n, MODIS has data for both t1 and t2. Interpolation can be performed using formulas (2)-(3), and the interpolation results are shown in the following table.

[0184] Table 2 Results of equal-length interpolation of NDVI series

[0185]

[0186] (2) Multiple methods to calculate the optimal weights. The above steps can process different satellite remote sensing data into sequences of equal length. In order to carry out multi-source data fusion and achieve complementary advantages, multiple methods are combined to calculate the optimal weights.

[0187] 1) The variance method is used to assign weights. The weight is the percentage of the variance of each NDVI series in the total variance. The product with the smallest variance is given the highest weight, and the product with the largest variance is given the lowest weight. The weight of Sentinel-2 is a1, the weight of Landsat8 is b1, and the weight of MODIS is c1.

[0188] 2) Weights are assigned based on spatial characteristics. Data sources with high spatial resolution can usually provide more detailed information and can be assigned higher weights. The weight is the proportion of the spatial resolution of each data source. The weight of Sentinel-2 is a2, the weight of Landsat8 is b2, and the weight of MODIS is c2.

[0189] Combining the above two methods, we can consider the main characteristics of various data sources and obtain the final weight through arithmetic averaging. The weight of Sentinel-2 is denoted as a (where ), the weight of Landsat8 is b (where, ), the weight of MODIS is c (where, ). The weight calculation results are shown in the following table.

[0190] Table 3 Calculation results of optimal weights for each remote sensing data source

[0191]

[0192] (3) Calculate the fusion data. Based on the optimal weights obtained in the above steps, the fusion data is calculated according to formula (4). The result is as follows: Figure 4 shown.

[0193] 4. Screening of the main influencing factors of NDVI. This example uses the CN05.1 daily-scale meteorological data (precipitation (Pre), average temperature (Tem), maximum temperature (Tem-max), minimum temperature (Tem-min), wind speed (Win), relative humidity (Rhu), and sunshine hours (Ssd)) of the Kubuqi Desert Ecological Management Area from 2004 to 2023, the annual water diversion (Div), the cumulative water diversion (Div s ) (desert water diversion factor) as alternative factors, the NDVI fusion data obtained by formula (3) was used as the dependent variable, and formulas (4)-(7) were used to carry out multiple tests (correlation test, autocorrelation test, F test, multicollinearity test) based on the multivariate linear regression equation, so as to gradually screen the main influencing factors.

[0194] The test results show that when NDVI fusion data is used as the dependent variable and precipitation, average temperature, wind speed, relative humidity, sunshine hours, and annual water diversion are used as independent variables, the correlation coefficient R 2 is 0.87 (high correlation, passed the correlation test); DW is 1.92 (close to 2, passed the autocorrelation test); the significance level of the F test is lower than 0.01 (passed the F test); the VIF values ​​of each variable are lower than 10 (passed the multicollinearity test); in addition, the random error term obeys the normal distribution (passed the random error term normal distribution test and heteroscedasticity test), the results are as follows Figure 5 shown.

[0195] 5. Construct NDVI impact mechanism. This embodiment constructs a set of NDVI impact mechanisms based on linear and nonlinear theories (Formula (8-10)). Linear methods include multivariate linear regression equations. Nonlinear methods include deep neural networks, random forests, and support vector machines. Based on historical data sets, training sets and test sets are divided, and the first 80% of the data is used as the training set and the last 20% as the test set to continuously optimize the impact mechanism of NDVI. The multivariate linear regression equation is as follows.

[0196] Y=0.0126+0.3173X1-0.0868X2-0.0698X3-1.0463X4+0.0818X5-0.0406X6

[0197] Where Y represents NDVI, X1 represents annual precipitation, X2 represents average temperature, X3 represents wind speed, X4 represents relative humidity, X5 represents sunshine hours, and X6 represents annual water diversion.

[0198] Finally, the effects of various models in the impact mechanism set are evaluated using formulas (12)-(14), and the results are shown in Table 4. Among them, if PCC≥0.6, RMSE≤0.2, and MAE≤0.2 are satisfied simultaneously, it indicates that the simulation effect of the model is good and can be further used for the prediction of water diversion volume.

[0199] Table 4 Evaluation Results of the NDVI Impact Mechanism Set

[0200]

[0201] Based on the calculation results in Table 4 and combined with the above evaluation criteria, deep neural network and random forest are selected for subsequent water diversion volume prediction.

[0202] 6. Adjust the future water diversion area. Research shows that when 0 < NDVI ≤ 0.2, it belongs to the low vegetation coverage area; when 0.2 < NDVI ≤ 0.4, it belongs to the medium-low vegetation coverage area; when 0.4 < NDVI ≤ 0.6, it belongs to the medium vegetation coverage area; when 0.6 < NDVI ≤ 0.8, it belongs to the medium-high vegetation coverage area; when 0.8 < NDVI ≤ 1, it belongs to the high vegetation coverage area. In this embodiment, considering the climate conditions of the desert and the growth characteristics of vegetation comprehensively, the local vegetation reaching the medium coverage level (NDVI: 0.4 - 0.6) is taken as the future planning goal. At the same time, the prediction results of PRECIS (Providing Regional Climates for Impacts Studies) for various meteorological elements (precipitation, average temperature, relative humidity, sunshine hours, wind speed) from 2025 to 2034 are input into the mechanism set determined in step 6, and the annual water diversion volume and comprehensive water diversion status corresponding to various models from 2025 to 2050 are calculated. The results are shown in Table 5. Further, the future water diversion volume range is set to be 31.78 - 78.53 million m 3 , and the results can provide a reference for engineering practice.

[0203] Table 5 Prediction Results of Optimal Water Diversion Volume from 2025 to 2050 (Water Diversion Volume (10,000 m 3 ))

[0204]

[0205]

[0206] The circuits, electronic components, and modules involved are all existing technologies, which can be fully realized by those skilled in the art without further elaboration. The content protected by the present invention does not involve the improvement of software and methods either.

[0207] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0208] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A quantitative method for constructing the mechanism of water diversion's impact on NDVI, characterized in that: include: Step 1: Obtain remote sensing image data from multiple data sources in the region and pre-process each remote sensing image; Step 2: Calculate the NDVI value of each remote sensing image after preprocessing, and calculate the annual NDVI mean of each data source in the region based on the NDVI value, and obtain the time series of the NDVI mean of each data source in the region; Step 3: Use the interpolation method to fill in the missing NDVI mean values ​​in the time series of each data source, so as to process the time series of NDVI mean values ​​of multiple data sources into sequences of equal length; Calculate the optimal weight of each data source based on variance method and spatial characteristics; According to the optimal weight, the corresponding NDVI means in the time series of multiple data sources are fused to obtain NDVI fused data; Step 4: Obtain the alternative influencing factors that affect the NDVI fusion data, take the NDVI fusion data as the dependent variable, and each alternative influencing factor as the independent variable, and screen out the main influencing factors of the NDVI fusion data from the alternative influencing factors based on correlation test, autocorrelation test, F test and multicollinearity test; Step 5: Based on the main influencing factors of the screened NDVI fusion data, a variety of models of the relationship between NDVI and the main influencing factors are constructed in combination with linear and nonlinear theories to form a set of NDVI influencing mechanisms; Step 6: Based on the NDVI fusion data, the Pearson correlation coefficient, root mean square error, and mean absolute error indicators are used to comprehensively evaluate the performance of various models of the NDVI impact mechanism set to obtain the final NDVI impact mechanism set; Step 7: Input the future meteorological data and the NDVI data of the corresponding period into the NDVI influencing mechanism set determined in step 6, calculate the future annual water diversion corresponding to different models, and comprehensively consider the results of multiple models to derive the water diversion range.

2. A quantitative method for constructing the mechanism of water diversion's impact on NDVI according to claim 1, characterized in that: In step 1, the preprocessing of each remote sensing image specifically includes: Each remote sensing image is subjected to geometric correction, radiation calibration, atmospheric correction, and stitching and cropping.

3. A quantitative method for constructing the mechanism of water diversion's impact on NDVI according to claim 1, characterized in that: In step 2, the NDVI value of each remote sensing image is calculated by the following formula: In the formula, ρ nir is the spectral reflectance in the near-infrared band, ρ r is the spectral reflectance in the red light band.

4. A quantitative method for constructing the mechanism of water diversion's impact on NDVI according to claim 1, characterized in that: In step 4, the main influencing factors of NDVI fusion data were screened out based on correlation test, including: The determination coefficient of the regression model is used to measure the influence of each influencing factor on the NDVI fusion data. 2 The calculation formula is as follows: In the formula, represents the model prediction value at time i, The mean of the long series of NDVI, y i represents the measured value of NDVI fusion data at the i-th moment, where the determination coefficient R 2 The value range is between 0 and 1, and the determination coefficient R 2 The closer the value is to 1, the stronger the explanatory power of the alternative influencing factors on the NDVI fusion data. 2 The closer it is to 0, the weaker the explanatory power of the alternative influencing factors on the NDVI fusion data.

5. A quantitative method for constructing the mechanism of water diversion's impact on NDVI according to claim 1, characterized in that: In step 4, the main influencing factors of NDVI fusion data were screened out based on the autocorrelation test, including: By calculating DW, the autocorrelation of each candidate influencing factor is evaluated. The calculation formula is as follows: In the formula, e t is the residual of the t-th observation, and n is the total number of observations; Among them, a DW value close to 0 indicates that the residual sequence has a strong positive autocorrelation, and a DW value close to 2 indicates that the residual sequence is close to independence.

6. A quantitative method for constructing the mechanism of water diversion's impact on NDVI according to claim 1, characterized in that: In step 4, the main influencing factors of NDVI fusion data were screened out based on the F test, including: The F test was used for variance analysis as follows: In the formula, SSR is the sum of squares of deviations between the alternative influencing factors, SSE is the sum of squares of deviations within the alternative influencing factors, m represents the total number of NDVI fusion data, and k represents the number of NDVI fusion data in the training group. Then, the test statistic F is calculated, and the critical value F' corresponding to the significance level of 0.05 is determined. If F>F', the F test is passed.

7. A quantitative method for constructing the mechanism of the impact of water diversion on NDVI according to claim 1, characterized in that: In step 4, the main influencing factors of NDVI fusion data were screened out based on multicollinearity test, including: The variance inflation coefficient is used to measure the collinearity of each candidate influencing factor. The calculation formula is as follows: In the formula, R 2 It is the determination coefficient. When the VIF value is large, it indicates that there is strong collinearity between the corresponding independent variable and other independent variables. When the VIF value is less than 10, the multicollinearity test is passed.

8. A quantitative method for constructing the mechanism of the impact of desert water diversion on NDVI according to claim 1, characterized in that: In step 5, the models of the relationship between the various NDVIs and the main influencing factors specifically include the following models: multiple linear regression equation, deep neural network, random forest and support vector machine.

9. A quantitative system for constructing the mechanism of water diversion's influence on NDVI, used to implement the quantitative method for constructing the mechanism of water diversion's influence on NDVI as claimed in claim 1, characterized in that: include: Preprocessing module: obtain remote sensing image data from multiple data sources in the region and preprocess each remote sensing image; Calculation module: Calculate the NDVI value of each remote sensing image after preprocessing, and calculate the annual NDVI mean of each data source in the region based on the NDVI value, and obtain the time series of the NDVI mean of each data source in the region; Fusion module: Use interpolation method to fill in the missing NDVI mean values ​​in the time series of each data source, so as to process the time series of NDVI mean values ​​of multiple data sources into equal-length sequences; Calculate the optimal weight of each data source based on variance method and spatial characteristics; According to the optimal weight, the corresponding NDVI means in the time series of multiple data sources are fused to obtain NDVI fused data; Screening module: Obtain the alternative influencing factors that affect the NDVI fusion data, take the NDVI fusion data as the dependent variable, and each alternative influencing factor as the independent variable, and screen out the main influencing factors of the NDVI fusion data from the alternative influencing factors based on correlation test, autocorrelation test, F test and multicollinearity test; Model building module: Based on the main influencing factors of the screened NDVI fusion data, a variety of models of the relationship between NDVI and the main influencing factors are constructed in combination with linear and nonlinear theories to form a set of NDVI influencing mechanisms; Evaluation module: Based on NDVI fusion data, the Pearson correlation coefficient, root mean square error, and mean absolute error indicators are used to comprehensively evaluate the performance of various models of the NDVI impact mechanism set to obtain the final NDVI impact mechanism set; Water diversion acquisition module: input future meteorological data and NDVI data of the corresponding period into the NDVI impact mechanism set determined by the evaluation module, and calculate the future annual water diversion corresponding to different models.

Images