Cotton field soil moisture inversion method and system combining microwave and optical data

Abstract

The invention discloses a cotton field soil moisture inversion method and system combining microwave and optical data, and relates to the technical field of remote sensing measurement of soil moisture content, and the method comprises the steps: obtaining satellite data which comprises radar satellite data and optical satellite data; a water cloud model is adopted for radar satellite data to remove vegetation influences, and radar characteristic parameters are obtained; performing polarization decomposition on the optical satellite data to obtain polarization decomposition characteristic parameters, and obtaining a vegetation index of the optical satellite data; the characteristic parameters comprise radar characteristic parameters, polarization decomposition characteristic parameters and vegetation indexes; screening the feature parameters by adopting a feature selection algorithm, presetting an importance threshold, and selecting the feature parameters greater than the importance threshold as a sample set; according to the method, the soil moisture inversion is carried out through combination of the optical satellite data and the radar satellite data, and the soil moisture inversion precision is improved.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing measurement of soil moisture content, and more particularly to a cotton field soil moisture inversion method and system combining microwave and optical data. Background Art

[0002] As an important surface parameter, soil moisture plays a vital role in the global surface water cycle and carbon cycle, and affects climate processes such as precipitation, temperature and evaporation, affecting the entire natural and human life at all times. Measuring global soil moisture helps improve our understanding of hydrological processes, ecosystem functions, and the connection between the earth's water, energy and carbon cycles.

[0003] Soil moisture determines the growth of crops. Monitoring farmland moisture is of great significance for agricultural science research such as crop yield estimation and drought and flood prediction. Affected by factors such as climate and topography, the temporal and spatial distribution of surface moisture is extremely uneven, and droughts and floods often occur. The time series of surface moisture can serve as a good indicator for pre-disaster warning, and the ability to estimate losses and study disasters after disasters cannot be ignored. Real-time, accurate, and large-scale monitoring of soil moisture is of great value for production and life, disaster prevention and control, and natural science research.

[0004] Therefore, how to accurately invert soil moisture is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the present invention provides a cotton field soil moisture inversion method and system combining microwave and optical data, which improves the inversion accuracy by combining optical satellite data with radar satellite data to perform soil moisture inversion.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A cotton field soil moisture inversion method combining microwave and optical data, comprising:

[0008] Acquiring satellite data, wherein the satellite data includes radar satellite data and optical satellite data;

[0009] The water cloud model is used to remove the influence of vegetation on the radar satellite data to obtain radar characteristic parameters; polarization decomposition is performed on the optical satellite data to obtain polarization decomposition characteristic parameters, and vegetation index of the optical satellite data is obtained at the same time; the characteristic parameters include the radar characteristic parameters, polarization decomposition characteristic parameters, and vegetation index;

[0010] Using a feature selection algorithm to screen the feature parameters, presetting an importance threshold, and selecting feature parameters greater than the importance threshold as a sample set;

[0011] The neural network is trained by the sample set, the trained neural network is used as an inversion model, and the satellite data to be inverted is input to perform soil moisture inversion.

[0012] Preferably, the method further includes preprocessing the radar satellite data and the optical satellite data, specifically including: the optical satellite data preprocessing is performed in ENVI, and the main preprocessing process includes radiation calibration, atmospheric correction, and orthorectification; the radar satellite data preprocessing is performed in SNAP 7.0 software, and the preprocessing process mainly includes orbit correction, thermal noise removal, multi-view processing, alignment, radiation calibration, filtering processing, and terrain correction.

[0013] Preferably, the method further includes converting the pre-processed radar satellite data into decibels, specifically including:

[0014] dB=log 10 (P / P0);

[0015] Among them, P is the target value and P0 is the reference value.

[0016] Preferably, the obtaining of radar characteristic parameters specifically includes: extracting VH polarization backscattering coefficient, VV polarization backscattering coefficient, and incident angle by calculating surface roughness and soil moisture content;

[0017]

[0018]

[0019]

[0020] Among them, m veg is the vegetation moisture content, NDVI is the normalized difference vegetation index, is the total radar backscatter coefficient under the vegetation-covered surface, is the direct backscattering coefficient of the vegetation layer, is the direct backscattering coefficient of the soil, γ 2 is the double-layer attenuation factor of the crop layer, A and B are parameters that depend on the vegetation type; m veg is the vegetation water content, Z S is the combined roughness, A V and B V is a coefficient that depends only on the angle of incidence, is the VH polarization backscattering coefficient, is the VV polarization backscattering coefficient.

[0021] Preferably, the polarization decomposition characteristic parameters specifically include the target object polarization entropy, the average scattering angle and the inverse entropy of the polarization entropy.

[0022] Preferably, the method further includes dividing the sample set into a training set and a validation set, wherein the data of the training set is used to train the neural network, and the data of the validation set is used to test and evaluate the accuracy of the model.

[0023] Preferably, the inversion model includes a convolution layer, a pooling layer, and a fully connected layer, which are connected in sequence. The number of neurons in the input layer is set to 21, and the number of neurons in the output layer is set to 1. It includes 5 convolution layers, 5 pooling layers and one fully connected layer. The convolution kernel size is 1×3, MAnormalization is added for regularization, LeakyRelu is selected as the activation function, and the maximum pooling parameter step size is 2.

[0024] A cotton field soil moisture inversion system combining microwave and optical data, comprising:

[0025] A data acquisition module, which acquires satellite data, wherein the satellite data includes radar satellite data and optical satellite data;

[0026] A characteristic parameter acquisition module, which uses a water cloud model to remove vegetation effects on the radar satellite data to obtain radar characteristic parameters; performs polarization decomposition on the optical satellite data to obtain polarization decomposition characteristic parameters, and simultaneously obtains a vegetation index of the optical satellite data; the characteristic parameters include the radar characteristic parameters, polarization decomposition characteristic parameters, and vegetation index;

[0027] A parameter selection module is used to sort the characteristic parameters by importance, preset an importance threshold, and select characteristic parameters greater than the importance threshold as a sample set;

[0028] The inversion module trains a neural network through the sample set, uses the trained neural network as an inversion model, and inputs the satellite data to be inverted to perform soil moisture inversion.

[0029] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a cotton field soil moisture inversion method and system that combines microwave and optical data. By combining optical satellite data with radar satellite data and combining with the water cloud model, the influence of vegetation moisture content on soil moisture is reduced, thereby improving the accuracy of soil moisture inversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0031] Figure 1 The accompanying drawing is a schematic diagram of the step flow chart provided by the present invention.

[0032] Figure 2 The accompanying drawing is a schematic diagram of the importance of radar data features provided by the present invention.

[0033] Figure 3 The accompanying drawing is a schematic diagram of the importance of radar+optical data features provided by the present invention.

[0034] Figure 4 The accompanying drawing is a schematic diagram of the structure of the neural network inversion model provided by the present invention.

[0035] Figure 5 The accompanying drawing is a scatter plot of radar data neural network data provided by the present invention.

[0036] Figure 6 The accompanying drawing is a scatter plot of random forest data of radar data provided by the present invention.

[0037] Figure 7 The accompanying drawing is a scatter plot of radar data support vector machine data provided by the present invention.

[0038] Figure 8 The accompanying drawing is a scatter plot of the extreme gradient boosting data of radar data provided by the present invention.

[0039] Fig. 9 The accompanying drawing is a scatter plot of multi-source data neural network data provided by the present invention.

[0040] Fig.10 The accompanying drawing is a scatter plot of multi-source data random forest data provided by the present invention.

[0041] Fig.11 The accompanying drawing is a scatter plot of multi-source data support vector machine data provided by the present invention.

[0042] Fig.12 The accompanying drawing is a scatter plot of the extreme gradient boosting data of multi-source data provided by the present invention.

[0043] Fig.13 The accompanying drawing is a diagram of soil moisture inversion results provided by the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0045] The embodiment of the present invention discloses a cotton field soil moisture inversion method combining microwave and optical data, such as Figure 1 As shown, including:

[0046] S1: Acquire satellite data, including radar satellite data and optical satellite data;

[0047] S2: Use the water cloud model to remove the vegetation effect on radar satellite data and obtain radar characteristic parameters; perform polarization decomposition on optical satellite data to obtain polarization decomposition characteristic parameters, and simultaneously obtain the vegetation index of optical satellite data; the characteristic parameters include radar characteristic parameters, polarization decomposition characteristic parameters, and vegetation index;

[0048] S3: Use feature selection algorithm to screen feature parameters, preset importance threshold, and select feature parameters greater than the importance threshold as sample set;

[0049] S4: Train the neural network through the sample set, use the trained neural network as the inversion model, and input the satellite data to be inverted to perform soil moisture inversion.

[0050] In a specific embodiment, it also includes preprocessing of radar satellite data and optical satellite data, specifically including: optical satellite data preprocessing is performed in ENVI, and the main preprocessing process includes radiation calibration, atmospheric correction, and orthorectification; radar satellite data preprocessing is performed in SNAP7.0 software, and the preprocessing process mainly includes orbit correction, thermal noise removal, multi-view processing, alignment, radiation calibration, filtering processing, and terrain correction.

[0051] In a specific embodiment, the pre-processed radar satellite data is further decibelized, specifically including:

[0052] dB=log 10 (P / P0);

[0053] Among them, P is the target value and P0 is the reference value.

[0054] In a specific embodiment, obtaining radar characteristic parameters specifically includes: extracting VH polarization backscattering coefficient, VV polarization backscattering coefficient, and incident angle by calculating surface roughness and soil moisture content;

[0055]

[0056]

[0057]

[0058] Among them, m veg is the vegetation moisture content, NDVI is the normalized difference vegetation index, is the total radar backscatter coefficient under the vegetation-covered surface, is the direct backscattering coefficient of the vegetation layer, is the direct backscattering coefficient of the soil, γ 2 is the double-layer attenuation factor of the crop layer, A and B are parameters that depend on the vegetation type; m veg is the vegetation water content, Z S is the combined roughness, A V and B V is a coefficient that depends only on the angle of incidence, is the VH polarization backscattering coefficient, is the VV polarization backscattering coefficient.

[0059] In a specific embodiment, the polarization decomposition characteristic parameters specifically include the target object polarization entropy, the average scattering angle, and the inverse entropy of the polarization entropy.

[0060] In a specific embodiment, the sample set is further divided into a training set and a validation set, the data of the training set is used to train the neural network, and the data of the validation set is used to test and evaluate the accuracy of the model.

[0061] In a specific embodiment, the inversion model includes a convolution layer, a pooling layer, and a fully connected layer. The convolution layer, the pooling layer, and the fully connected layer are connected in sequence. The number of neurons in the input layer is set to 21, and the number of neurons in the output layer is 1. It includes 5 convolution layers, 5 pooling layers and one fully connected layer. The convolution kernel size is 1×3, MA normalization is added for regularization, LeakyRelu is selected as the activation function, and the maximum pooling parameter step size is 2.

[0062] A cotton field soil moisture inversion system combining microwave and optical data, comprising:

[0063] A data acquisition module is used to acquire satellite data, including radar satellite data and optical satellite data;

[0064] The characteristic parameter acquisition module uses a water cloud model to remove the influence of vegetation on radar satellite data and obtain radar characteristic parameters; polarization decomposition is performed on optical satellite data to obtain polarization decomposition characteristic parameters, and the vegetation index of optical satellite data is obtained at the same time; the characteristic parameters include radar characteristic parameters, polarization decomposition characteristic parameters, and vegetation index;

[0065] The parameter selection module sorts the feature parameters by importance, presets the importance threshold, and selects the feature parameters that are greater than the importance threshold as the sample set;

[0066] The inversion module trains the neural network through the sample set, uses the trained neural network as the inversion model, and inputs the satellite data to be inverted to perform soil moisture inversion.

[0067] In a specific embodiment, the optical satellite data used in the present invention is the GF-6WFV data product, which includes a 16m medium-resolution wide-band camera with 8 multispectral bands. Compared with the previous Gaofen series satellites, the purple band (0.40-0.45μm), yellow band (0.59-0.63μm), red edge band 1 (0.69-0.73μm) and red edge band 2 (0.73-0.77μm) are newly added. The preprocessing is performed in ENVI, and the main preprocessing processes include: radiation calibration, atmospheric correction, orthorectification, etc.

[0068] The radar satellite data used in the present invention is the dual-polarization (VH and VV) slant range single look complex product (Single Look Complex, SLC) in Sentinel-1A wide imaging mode, the data level is Level-1, C band, the resolution is 5×20m, and the incident angle range is 30.05°~46.17°. The preprocessing process is carried out in the SNAP 7.0 software developed by the European Space Agency. The preprocessing process mainly includes: orbit correction, thermal noise removal, multi-view processing, registration, radiation calibration, filtering processing, terrain correction, etc. The image resolution after data preprocessing is 10m, and it needs to be resampled to the same 16m resolution as the optical satellite data.

[0069] The result obtained after the above preprocessing of radar data is the backscatter coefficient in linear proportional units. The value is usually a very small positive value. The reason is that the transmission distance is far and the power received by the radar is small, which is not conducive to the development of practical applications. The result needs to be converted to decibels (dB). The range of the backscatter coefficient after decibelization can reach a range close to the common Gaussian distribution, which is convenient for visualization and data analysis. The formula for decibelization is:

[0070] dB=log 10 (P / P0) (1)

[0071] In the formula, P and P0 represent the target value and the reference value respectively. For the backscatter coefficient σ0, the meaning of decibel conversion is actually to perform the following logarithmic transformation:

[0072] σ0(dB)=10*log 10 σ0 (2)

[0073] Radar images are penetrating, and can obtain ground information at any weather and time due to the influence of natural environments such as sunlight, clouds, and atmospheric moisture. Radar parameters such as different polarization modes and incident angles, as well as parameters such as the dielectric constant and physical characteristics of target objects, have the most direct impact on radar image information. Therefore, extracting feature information that can characterize object information from radar images is the basis for radar remote sensing inversion. Optical images can enhance a certain feature or detail of vegetation through combined operations between different bands (2 or more bands). Vegetation has a very direct impact on surface scattering information. Vegetation index can be used to participate in the construction of feature parameters in the construction of soil moisture inversion models. At present, there are more than 100 vegetation indices proposed in the field of remote sensing. Due to the different combinations of different sensors and bands, different vegetation indices are suitable for different types of remote sensing data and different application fields.

[0074] Radar images can detect the characteristics of targets by emitting microwave beams to target objects and receiving the returned echo signals. According to the latitude and longitude information of the actual sampling point, the corresponding VH polarization backscatter coefficient can be extracted from the preprocessed radar data. VV polarization backscatter coefficient The incident angle (θ) is prepared for the subsequent construction of feature sample sets.

[0075] The surface parameters that determine the backscattering intensity are mainly surface roughness and soil moisture content. Considering that cotton planting was in the boll-opening period at the time of the present invention, the surface vegetation coverage of farmland in the study area was considered. The influence of vegetation was removed according to the water cloud model to obtain the backscattering coefficient of bare soil. The vegetation water content (Vegetation watercontent) in the water cloud model input parameter was calculated using the Jackson vegetation water content calculation formula based on NDVI. The backscattering coefficient of bare soil after removing vegetation coverage using Sentinel-1A was used to calculate the surface composite roughness. The calculation formula is as follows (Formula 3 is the Jackson formula, Formula 4 is the water cloud model calculation formula, and Formula 5 is the surface roughness calculation formula).

[0076]

[0077] In the formula, m veg is the vegetation moisture content, and NDVI is the normalized difference vegetation index.

[0078]

[0079] Where: is the total radar backscatter coefficient under the vegetation cover surface; is the direct backscattering coefficient of the vegetation layer; is the direct backscattering coefficient of the soil; γ 2 is the double-layer attenuation factor of the crop layer. These parameters are all functions related to the radar wave incident angle θ, among which A and B are parameters that depend on the vegetation type. The general vegetation parameter A value is 0.0012 and B value is 0.091; m veg is the vegetation water content (kg / m 2 ).

[0080]

[0081] In the formula, Z S is the combined roughness, A V and B V is a coefficient that depends only on the angle of incidence.

[0082] There is also a certain relationship between cos(θ) and sin(θ) and soil moisture, and the ratio of the backscattering coefficient When the incident angle is fixed, it is only related to the surface roughness. Therefore, cos(θ), sin(θ) and As characteristic parameters. And In summary, 10 radar characteristic parameters related to backscatter coefficient and incident angle are extracted from radar images.

[0083] Polarization decomposition can decompose the complex scattering process of ground objects into several simpler scattering mechanisms, and there is a certain correlation between polarization characteristic parameters and soil moisture. Sentinel-1A dual-polarization data only supports H / A / Alpha decomposition. Polarization entropy H, average scattering angle α, and the complementary parameter of polarization entropy - anti-entropy AE of target ground objects are extracted through polarization decomposition. Three polarization decomposition characteristic parameters are extracted from radar data.

[0084] Vegetation index can reflect the growth status of surface vegetation in a certain sense, and the growth status of surface vegetation also has a certain correlation with soil moisture. Based on the multispectral characteristics of GF-6WFV data and the unique red edge index characteristics, this experiment finally selected 16 vegetation indices commonly used in vegetation information inversion research, including normalized difference vegetation index, normalized difference spectral index 1, normalized difference spectral index 2, normalized difference spectral index 3, enhanced vegetation index, enhanced spectral index 1, enhanced spectral index 2, enhanced spectral index 3, soil adjusted vegetation index, soil adjusted vegetation index 1, soil adjusted vegetation index 2, soil adjusted vegetation index 3, ratio vegetation index, ratio vegetation index 1, ratio vegetation index 2, ratio vegetation index 3, and the calculation formula is shown in Table 1.

[0085] Table 1 Spectral vegetation index data calculation formula

[0086]

[0087]

[0088] ρ 04 , ρ 05 , ρ 06 , ρ R and ρ B They are respectively the reflectance of the near infrared band, red edge I band, red edge II band, infrared band and blue band of the Gaofen 6 image. L is the soil adjustment coefficient, and according to the vegetation coverage of the study area, the present invention takes L=0.5.

[0089] The Sentinel-1SAR data and GF-6WFV data were preprocessed, and the 29 characteristic parameters mentioned above were calculated through the water cloud model, the surface combined roughness model and the exponential formula of the vegetation index to complete the construction of the sample set.

[0090] Use the feature selection algorithm to select the best feature subset from 29 feature parameters. Set the maximum number of iterations to 300 and the population size to 100. Through the iteration of the genetic algorithm, the best feature subset is gradually found to improve the performance and generalization ability of the model. By evaluating the adaptability and genetic operations, the feature subset can be searched and optimized in the feature space to find the most distinctive and predictive feature combination, thereby improving the effectiveness of the machine learning model. The results of the model ranking of importance are shown in the figure below. Figure 2-3 shown.

[0091] In the experiment, it was found that when 12 features with importance greater than 0.045 were selected from the SAR features, and 21 features with importance greater than or equal to 0.02 were selected for the radar + optical feature combination as the sample set, the generalization was good and the model was relatively stable. The selected feature set was used as the sample input and soil moisture was used as the output to construct a soil moisture inversion model to invert soil moisture. The model results constructed with a single SAR data source were compared with the model results after adding optical features.

[0092] Convolutional neural networks (CNN) are a type of feedforward neural networks that include convolution operations and have a deep structure. They are one of the representative algorithms of deep learning. Convolutional neural networks have representation learning capabilities and can abstract input sample information according to its structural information to obtain deeper feature information. The present invention establishes a "convolution-pooling-full connection" convolutional neural network structure to explore the feasibility of convolutional neural networks in soil moisture inversion in cotton fields in arid areas.

[0093] The present invention uses 325 measured soil volume moisture content data to construct a soil moisture inversion model. Among them, 70% of the data is used for training, and 30% is used to test and evaluate the model accuracy. We use a convolutional neural network as the inversion model, set the number of input layer neurons to 21, and the number of output layer neurons to 1. The model contains 5 convolutional layers, 5 pooling layers and one fully connected layer. The convolution layer extracts deep abstract features through a convolution kernel with a kernel size of 1×3. To avoid overfitting, we add MA normalization for regularization during training, and select LeakyRelu as the activation function. The maximum pooling parameter step size is set to 2 to reduce the number of parameters and remove redundant information, thereby simplifying the complexity of the network. The model structure is as follows Figure 4 shown.

[0094] Random forest (RF) is a self-learning machine learning algorithm that combines and integrates multiple trees based on the idea of ​​ensemble learning. The basic component unit is a series of decision trees that follow binary rules, also called classification and regression trees (CART), which determines that the RF model can be applied to classification and regression fields. In order to obtain higher model prediction accuracy, the RF model combines multiple independent decision numbers. Compared with other traditional regression models, the RF model can better deal with noise and outliers, has higher computational efficiency, self-learns multidimensional nonlinear mapping, and has better fitting effect. The RF model can find nonlinear relationships between multiple independent variables and dependent variables in complex environmental variables, and fit them to achieve the best estimation effect. The training set is divided into 70% of the total number of samples and the test set is 30%. During the construction process, the present invention tests the number of decision numbers, the number of features of a single decision tree, the tree depth and other parameters in a cyclic iterative manner to seek the optimal parameters of the model to achieve the best estimation effect. For regression problems, the prediction function of random forest can be expressed as:

[0095]

[0096] in, is the prediction output of the random forest, N is the number of decision trees in the random forest, and f i (x) is the prediction result of the i-th decision tree for the input sample x.

[0097] Support Vector Machine (SVM) achieves the effect of data classification and prediction by finding a partitioning hyperplane that can correctly separate two types of data on both sides. Different support vector machines can be constructed by different kernel functions. Common kernel functions for constructing support vector machines include Sigmoid function, polynomial function, linear function, radial basis function (RBF), etc. When applied to regression problems, SVM is called Support Vector Regression (SVR). When the present invention is applied to the regression of soil moisture, the commonly used RBF kernel function is selected. The reason is that the processing accuracy of the RBF kernel function is generally higher than that of the polynomial and Sigmoid functions, it can handle nonlinear problems, and the RBF kernel function requires fewer parameters to be determined. Similarly, 70% of the data is randomly divided as a training set, and the remaining 30% is a test set. The nonlinear problem between the measured soil moisture and the constructed feature set independent variables is converted into a prior problem in a high-dimensional feature space. The regression function of the constructed support vector machine is:

[0098]

[0099] Where: α i , (i=1,2,3…,n) is the lagrange multiplier; x is the predictor variable; x i is the sample factor vector of the support vector machine; σ is the kernel parameter, b is a constant, b∈R.

[0100] eXtreme Gradient Boosting XGBoost is an ensemble learning algorithm based on the Boosting concept. It is an optimized ensemble tree algorithm generated by optimizing and upgrading the GBDT (gradient boosted tree) algorithm. The central idea is to continuously split features to grow into trees. Each newly generated tree is a new function used to fit the last residual. Finally, the calculated value of each leaf node is added to obtain the final regression prediction value. The maximum tree depth of the model is set to 7, 300 decision trees are set to participate in the construction of the model, the objective function is set to a linear regression function, and the minimum sum of sample weights in the child nodes is 1 to prevent overfitting. The regression formula is as follows:

[0101]

[0102] Where: is the predicted value of the model, K is the number of trees, F is the set space of regression trees (also known as CART), X i represents the eigenvector of the ith data point. Where F = {f(X) = w q(x)}(q:R m→T,w∈R T ),q represents the structure of each tree that maps examples to corresponding leaves; the number of leaves in T tree; f k The structure q and leaf weight w of the k-th independent tree. Each regression tree will contain a continuous score on each leaf, applicable to w i To represent the score on the i-th leaf.

[0103] The coefficient of determination (R 2 ), root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the inversion accuracy of the model. 2 It is used to evaluate the degree of fit between the inversion result of soil moisture and the measured result of soil moisture, while RMSE can measure the deviation between the predicted value and the measured value. The smaller the value, the closer the inversion result of soil moisture is to the measured soil moisture value. MAE represents the mean absolute difference between the predicted value and the measured value of soil moisture, the average size of the measurement error, but does not consider the direction. Figure 5-Figure 8 This is a scatter plot comparing the inversion accuracy of the four models after inputting SAR data. Figure 9-12 The scatter plot is a comparison of the inversion accuracy of the four models after adding the optical data, and Table 2 is a comparison of the accuracy of the eight experimental schemes.

[0104] Table 2 Comparison of model accuracy of different schemes

[0105]

[0106]

[0107] Compared with the eight experimental schemes in Table 2, when using single synthetic aperture radar data to invert soil moisture based on feature selection, the CNN model has the highest estimation accuracy, R 2 is 0.614, and the RMSE is 0.0341 cm 3 / cm 3 , MAE is 0.0258cm 3 / cm 3 XGBoost's estimation effect is second, R 2 is 0.609, and RMSE is 0.0349cm 3 / cm 3 , MAE is 0.0269cm 3 / cm 3 . The RF model is the second worst, and the SVR model has the worst estimation effect. 2 Only 0.236, RMSE is 0.0476cm 3 / cm3 , MAE is 0.0364cm 3 / cm3. After adding optical data, the estimation accuracy of the four models is generally improved. Among them, the R 2 It increases to 0.789, and the RMSE decreases by 0.0092cm 3 / cm 3 , MAE decreased by 0.006cm 3 / cm 3 The estimation accuracy of RF increased to 0.709 and the RMSE decreased by 0.0059 cm. 3 / cm 3 , MAE is 0.0219cm 3 / cm 3 . SVR R 2 It rose to 0.277, and the RMSE dropped by 0.0015cm 3 / cm 3 , MAE decreased by 0.0023cm3 / cm3. After adding optical data, XGBoost’s R 2 Increased to 0.648, RMSE decreased by 0.0648cm 3 / cm 3 , MAE is only 0.0226cm 3 / cm 3 .

[0108] Experimental results show that the convolutional neural network model has higher accuracy and stronger robustness than the random forest, support vector machine and XG-Boost models, and can better cope with the soil moisture inversion problem in arid and semi-arid areas.

[0109] Fig.13 This is the result of soil moisture inversion using a convolutional neural network combining microwave and optical remote sensing data. In order to better display the soil moisture content in farmland, we mask out non-farmland areas such as villages, roads, and lakes. As can be seen from the figure, the soil moisture content ranges from 0.05-0.30cm 3 / cm 3 The inversion results are basically consistent with the actual measured values.

[0110] Based on Sentinel-1A active microwave remote sensing data and GF-6WFV multispectral data, this paper constructs a sample set related to soil moisture through a water cloud model and a surface combined roughness model, and establishes a convolutional neural network soil moisture inversion model of the study area in combination with measured data. The results are compared with those of the random forest, support vector machine and XG-Boost inversion models, and the active microwave remote sensing and optical remote sensing collaborative inversion of cotton field soil moisture in arid and semi-arid areas is realized. The main conclusions are as follows:

[0111] (1) Vegetation coverage will have a significant impact on the inversion results of soil moisture. The present invention calculates the moisture content of cotton field vegetation through the Jackson model and combines it with the water cloud model to effectively reduce the impact of vegetation moisture content on soil moisture and achieve good results.

[0112] (2) Compared with single microwave remote sensing data, the addition of vegetation index features extracted by optical remote sensing can improve the inversion accuracy of soil moisture. Among the vegetation indices commonly used in current soil moisture inversion research, NDVI is more common, while this paper finds that vegetation indices such as RVI and ARVI are more important for soil moisture inversion in the study area.

[0113] (3) Convolutional neural network, random forest and support vector machine models were introduced for soil moisture inversion. The results showed that whether it was a single data source or combined microwave and optical remote sensing data, the inversion effect of convolutional neural network was the best, random forest was second in the estimation effect when microwave data was combined with optical data, XG-Boost was second in the estimation effect when single microwave data was used, and support vector machine was the worst in the two data sources. 2 ) can reach 0.789, and the RMSE is 0.0249cm 3 / cm 3 , MAE is 0.0198cm 3 / cm 3 .

[0114] (4) The three parameters decomposed by H / A / Alpha are less important to soil moisture, or because the cotton was in the boll opening period when the measurement was made, the vegetation coverage was high, which had a greater impact on polarization decomposition, thus affecting the importance of polarization decomposition parameters to soil moisture.

[0115] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0116] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present invention 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 in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A cotton field soil moisture inversion method combining microwave and optical data, characterized in that: include: Acquiring satellite data, wherein the satellite data includes radar satellite data and optical satellite data; The water cloud model is used to remove the influence of vegetation on the radar satellite data to obtain radar characteristic parameters; polarization decomposition is performed on the optical satellite data to obtain polarization decomposition characteristic parameters, and vegetation index of the optical satellite data is obtained at the same time; the characteristic parameters include the radar characteristic parameters, polarization decomposition characteristic parameters, and vegetation index; Using a feature selection algorithm to screen the feature parameters, presetting an importance threshold, and selecting feature parameters greater than the importance threshold as a sample set; The neural network is trained by the sample set, the trained neural network is used as an inversion model, and the satellite data to be inverted is input to perform soil moisture inversion.

2. The cotton field soil moisture inversion method combining microwave and optical data according to claim 1 is characterized in that: It also includes preprocessing the radar satellite data and the optical satellite data, specifically including: the optical satellite data preprocessing is performed in ENVI, and the main preprocessing process includes radiation calibration, atmospheric correction, and orthorectification; the radar satellite data preprocessing is performed in SNAP 7.0 software, and the preprocessing process mainly includes orbit correction, thermal noise removal, multi-view processing, alignment, radiation calibration, filtering processing, and terrain correction.

3. The cotton field soil moisture inversion method combining microwave and optical data according to claim 2 is characterized in that: The method also includes decibelizing the pre-processed radar satellite data, specifically including: dB=log 10 (P / P0); Among them, P is the target value and P0 is the reference value.

4. The cotton field soil moisture inversion method combining microwave and optical data according to claim 1 is characterized in that: The obtaining of radar characteristic parameters specifically includes: extracting VH polarization backscattering coefficient, VV polarization backscattering coefficient, and incident angle by calculating surface roughness and soil moisture content; Among them, m veg is the vegetation moisture content, NDVI is the normalized difference vegetation index, is the total radar backscatter coefficient under the vegetation-covered surface, is the direct backscattering coefficient of the vegetation layer, is the direct backscattering coefficient of the soil, γ 2 is the double-layer attenuation factor of the crop layer, A and B are parameters that depend on the vegetation type; m veg is the vegetation water content, Z S is the combined roughness, A V and B V is a coefficient that depends only on the angle of incidence, is the VH polarization backscattering coefficient, is the VV polarization backscattering coefficient.

5. The cotton field soil moisture inversion method combining microwave and optical data according to claim 1 is characterized in that: The polarization decomposition characteristic parameters specifically include the target object polarization entropy, the average scattering angle and the inverse entropy of the polarization entropy.

6. The cotton field soil moisture inversion method combining microwave and optical data according to claim 1 is characterized in that: It also includes dividing the sample set into a training set and a validation set, wherein the data of the training set is used to train a neural network, and the data of the validation set is used to test and evaluate the accuracy of the model.

7. The cotton field soil moisture inversion method combining microwave and optical data according to claim 1 is characterized in that: The inversion model includes a convolution layer, a pooling layer, and a fully connected layer. The convolution layer, the pooling layer, and the fully connected layer are connected in sequence. The number of neurons in the input layer is set to 21, and the number of neurons in the output layer is set to 1. It includes 5 convolution layers, 5 pooling layers, and one fully connected layer. The convolution kernel size is 1×3, MA normalization is added for regularization, Leaky Relu is selected as the activation function, and the maximum pooling parameter step size is 2.

8. A cotton field soil moisture inversion system combining microwave and optical data, characterized in that: include: A data acquisition module, which acquires satellite data, wherein the satellite data includes radar satellite data and optical satellite data; A characteristic parameter acquisition module, which uses a water cloud model to remove vegetation effects on the radar satellite data to obtain radar characteristic parameters; performs polarization decomposition on the optical satellite data to obtain polarization decomposition characteristic parameters, and simultaneously obtains a vegetation index of the optical satellite data; the characteristic parameters include the radar characteristic parameters, polarization decomposition characteristic parameters, and vegetation index; A parameter selection module is used to sort the characteristic parameters by importance, preset an importance threshold, and select characteristic parameters greater than the importance threshold as a sample set; The inversion module trains a neural network through the sample set, uses the trained neural network as an inversion model, and inputs the satellite data to be inverted to perform soil moisture inversion.

Images