Cotton field soil humidity monitoring method, system, equipment, medium and product

Abstract

The invention discloses a cotton field soil humidity monitoring method, system and device, a medium and a product, and relates to the field of soil humidity monitoring. Sentinel No.1 microwave data and Gaofen-1 optical data of a to-be-monitored cotton field area are acquired, and a cotton microwave characteristic factor and an optical characteristic factor are extracted; estimating the soil humidity of the cotton field area to be monitored by utilizing a cotton field soil humidity estimation model according to the two characteristic factors; the cotton field soil humidity estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network by using a training data set; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training the improved two-dimensional convolutional neural network by using a pre-training sample set obtained by simulation of an AIEM-Oh model based on microwave parameters and ground actually measured soil roughness; the improved two-dimensional convolutional neural network comprises a convolutional layer, a depth separable convolutional layer and a full connection layer. The method can accurately estimate the soil humidity of the cotton field area.

Description

Technical Field

[0001] The present application relates to the field of soil moisture monitoring, and in particular to a cotton field soil moisture monitoring method, system, equipment, medium and product. Background Art

[0002] Soil moisture usually refers to the water content of the unsaturated layer of soil, and is an important basis for quantifying the land-atmosphere material and energy exchange and agricultural field management. Cotton is an important strategic material. Xinjiang is China's largest cotton production and export base, and also China's largest inland irrigation area. Since the 1990s, Xinjiang's agricultural water use has become increasingly scarce, lacking intensive management, and wasted seriously. The cotton-growing area located in the economic zone on the northern slope of the Tianshan Mountains is an important part of the Xinjiang cotton area. Remote sensing monitoring of soil moisture in cotton fields has important practical significance for accurately controlling the temporal and spatial changes in soil moisture in cotton fields, and then scientifically guiding quantitative irrigation, rationally using water resources, and promoting cotton water productivity.

[0003] Compared with traditional soil moisture measurement methods, remote sensing technology provides a periodic, global, and multi-temporal means of ground observation, which brings new technical support to soil moisture research. According to the sensor type and electromagnetic wave frequency band, soil moisture remote sensing monitoring mainly includes visible light / near infrared remote sensing and active / passive microwave remote sensing. The soil moisture inversion methods of different remote sensing data have their own characteristics. However, optical remote sensing has low temporal resolution and is affected by weather. It cannot be observed at night, and most estimation methods can only reflect the relative humidity of the soil. Microwave remote sensing is more suitable for soil moisture estimation. At present, global-scale soil moisture remote sensing products are all based on microwave detection data, but the spatial resolution is low and they are easily affected by vegetation and surface roughness. In order to give full play to the advantages of various types of data, the comprehensive use of multi-source remote sensing data such as optical remote sensing and microwave remote sensing to invert soil moisture to improve the inversion accuracy and reliability has become the main direction of soil moisture estimation. The key to improving the accuracy of soil moisture inversion in vegetation-covered areas by using visible light / near infrared and microwaves is how to quantify and eliminate the influence of vegetation. Current research is mostly focused on improving parameters such as vegetation moisture content and roughness in vegetation effect models. At present, the more mature models that consider vegetation scattering effects mainly include the water cloud model and the Michigan microwave vegetation scattering model. The former is relatively simple and practical in crop-covered areas, but it ignores the multiple scattering effects between crops and soil, and has large errors for crops with tall stems, such as corn and sorghum. The Michigan microwave vegetation scattering model can more realistically simulate the influence of tall vegetation on microwave backscattering, but it is difficult to promote because it needs to obtain the leaf and branch parameters of vegetation. Current research is also mostly focused on improving and calibrating vegetation moisture content parameters in vegetation effect models. In addition, different types of sensors reflect soil moisture at different spatial resolutions and depths. Radiometers are usually global-scale products, while visible / near-infrared and radar are generally regional or watershed-scale products. It is necessary to improve the spatial resolution through spatial downscaling of visible / near-infrared and microwave remote sensing, establish the correlation or physical model between passive microwave soil moisture products and auxiliary variables of medium- and high-spatial-resolution visible / near-infrared data, and effectively expand the application space of soil moisture products. Most of these methods are aimed at specific spatiotemporal research areas, and use "true value" data sets to simplify theoretical models. They have good inversion accuracy and applicability. Among them, the water-cloud model, which focuses on soil moisture estimation in crop growing areas, can simultaneously invert soil moisture and crop biomass and other related parameters, and is widely used. In particular, since field soil moisture is restricted by multiple factors such as meteorological conditions, soil properties, crop types and phenological periods, it is essential to select the optimal data source; at the same time, the contribution of the data source will also change over time, resulting in great differences in the scope of application of inversion methods and models. Even if some aspects of them have advantages and characteristics, they inevitably have certain limitations.Studies have shown that for planting areas, as crops grow and develop, their vegetation coverage, water consumption, soil and crop evapotranspiration also change accordingly, making it particularly difficult to estimate soil moisture and obtain spatiotemporal variation characteristics. In addition, due to the different observation mechanisms, spectral or band distribution, spatiotemporal resolution, transit time, data quantification level, and electromagnetic radiation detection depth of different sensors, most remote sensing surface magnetic sources (usually referred to as the surface layer) have visible light / near infrared and thermal bands of only a few millimeters, and near-surface magnetic sources have microwaves of X, C or L frequencies of only a few centimeters. Therefore, multi-load data fusion leads to the range of soil moisture estimation results that is difficult to unify and the quantitative relationship is very complex. The condition of crop water stress is often difficult to be truly reflected, and the inversion results usually have large uncertainties. Generally, when using empirical, semi-empirical function methods and physical models to describe soil moisture, there are clear restrictions on the specific characteristics of soil type, roughness and structure, while nonlinear learning methods can obtain better fitting effects, alleviate or even solve this problem. In order to simulate complex nonlinear quantitative relationships and improve the ability to measure the complexity of various factors, machine learning algorithms are a good choice for estimating surface soil moisture at high spatial and temporal scales. Although this method cannot obtain an explicit expression of the inversion model like empirical, semi-empirical and theoretical models, once it is successfully trained, it can be quickly applied to thematic mapping of surface soil moisture, and can be adapted to rough surfaces and vegetation-covered areas, with strong practical value. Even though the machine learning algorithm is computationally complex, it can intelligently mine the deep intrinsic relationships of different remote sensing data. Its application has an important role in improving the accuracy of surface soil moisture inversion, especially with the enrichment of multi-angle, multi-band and multi-temporal visible light / near infrared and microwave remote sensing data, the deep mining of visible light / near infrared and microwave based on machine learning, and the quantitative method of soil roughness and vegetation canopy influence are also gradually improving. However, machine learning methods also have problems such as confusing definition of weight adjustment, fuzzy hidden layer nodes, overfitting local optimum, and slow convergence. At present, deep learning technology has completely automated the key steps of shallow machine learning technology, namely feature engineering, which is more user-friendly; more importantly, deep learning technology can continuously learn online and learn all features at once, with the ability to perform real-time operations. Convolutional neural networks achieve the effects of local association and parameter sharing, and can reduce human intervention. They can find information features to the greatest extent, thereby enhancing the accuracy of estimation, and are gradually being used in soil moisture prediction and inversion. Summary of the invention

[0004] The purpose of this application is to provide a cotton field soil moisture monitoring method, system, equipment, medium and product to improve the accuracy of cotton field soil moisture monitoring.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a cotton field soil moisture monitoring method, comprising:

[0007] Obtain the Sentinel-1 and Gaofen-1 remote sensing images of the cotton field area to be monitored;

[0008] According to the Sentinel-1 microwave data and the GF-1 optical data, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the cotton field area to be monitored are extracted; the microwave characteristic factors include the radar backscatter coefficient under vertical polarization mode, the radar backscatter coefficient under vertical-horizontal polarization mode, the combination of the radar backscatter coefficient under vertical polarization mode and the radar backscatter coefficient under vertical-horizontal polarization mode, the incident angle, the polarization entropy, the average scattering angle, the anisotropy degree and the surface combination roughness; the optical characteristic factors include the full color spectrum, the blue spectrum, the green spectrum, the red spectrum, the normalized vegetation index, the difference vegetation index, the ratio vegetation index, the soil adjusted vegetation index and the enhanced vegetation index;

[0009] According to the microwave characteristic factors and optical characteristic factors of the cotton field area to be monitored, the soil moisture of the cotton field area to be monitored is estimated by using a cotton field soil moisture estimation model; wherein the cotton field soil moisture estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network using a training data set; the training data set includes microwave backscattering signals of the training cotton field area after removing cotton canopy interference, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods, and corresponding soil moisture labels; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training an improved two-dimensional convolutional neural network using a pre-training sample set; the pre-training sample set is simulated using the AIEM-Oh model based on microwave parameters and ground measured soil roughness; the improved two-dimensional convolutional neural network includes a 1×1 convolutional layer connected in sequence, a 3×3 depth-separable convolutional layer based on Dropout regularization technology, and a fully connected layer.

[0010] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0011] The present application provides a cotton field soil moisture monitoring method, system, equipment, medium and product, which obtains the Sentinel-1 microwave data and the High-resolution-1 optical data of the cotton field area to be monitored and extracts the microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the cotton field area to be monitored; according to the microwave characteristic factors and optical characteristic factors of the cotton field area to be monitored, the cotton field soil moisture estimation model is used to determine the soil moisture of the cotton field area to be monitored; wherein the cotton field soil moisture estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network using a training data set; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training an improved two-dimensional convolutional neural network using a pre-training sample set; the pre-training sample set is simulated by the AIEM-Oh model based on the microwave's own parameters and the measured soil roughness on the ground; the improved two-dimensional convolutional neural network includes a 1×1 convolutional layer, a 3×3 depth-separable convolutional layer and a fully connected layer connected in sequence. The present application uses the constructed cotton field soil moisture estimation model to accurately estimate the soil moisture in the cotton field area. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A schematic diagram of a cotton field soil moisture monitoring method provided in an embodiment of the present application;

[0013] Figure 2 Schematic diagram of the process of generating high-precision DEM data for sample points;

[0014] Figure 3 It is the response curve of soil moisture and VV backscattering coefficient under different incident angles;

[0015] Figure 4 It is the response curve of soil moisture and VV backscattering coefficient at different RMS heights;

[0016] Figure 5 It is the response curve of soil moisture and VV backscattering coefficient under different correlation lengths;

[0017] Figure 6 Estimating VWC for the model RVI Scatter plot of values ​​and synchronous measured sample data;

[0018] Figure 7 For VWC RVI Schematic diagram of the distribution of estimated values;

[0019] Figure 8 It is a two-dimensional convolutional neural network model structure diagram;

[0020] Fig. 9 This is a structural diagram of the improved two-dimensional convolutional neural network model;

[0021] Fig.10Set the corresponding Loss change curve for the training set and test set Epoch;

[0022] Fig.11 The curve chart of the test set Epoch setting and the corresponding MSE change;

[0023] Fig.12 Set the corresponding R for the test set Epoch 2 Change curve chart;

[0024] Fig.13 The MSE change curve corresponding to the test set Epoch setting when training different models;

[0025] Fig.14 Set the corresponding R for the test set Epoch when training different models 2 Change curve chart. DETAILED DESCRIPTION

[0026] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0027] In an exemplary embodiment, Figure 1 As shown, a cotton field soil moisture monitoring method is provided, comprising the following steps:

[0028] S1: Obtain Sentinel-1 microwave data and GF-1 optical data of the cotton field area to be monitored.

[0029] S2: Based on the Sentinel-1 microwave data and the GF-1 optical data, the microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the cotton field area to be monitored are extracted; the microwave characteristic factors include the radar backscattering coefficient under vertical polarization, the radar backscattering coefficient under vertical-horizontal polarization, the combination of the radar backscattering coefficient under vertical polarization and the radar backscattering coefficient under vertical-horizontal polarization, the incident angle, polarization entropy, average scattering angle, anisotropy and surface combination roughness; the optical characteristic factors include the full color spectrum, blue spectrum, green spectrum, red spectrum, normalized vegetation index, difference vegetation index, ratio vegetation index, soil adjusted vegetation index and enhanced vegetation index.

[0030] S3: Estimate the soil moisture of the cotton field area to be monitored according to the microwave characteristic factors and optical characteristic factors of the cotton field area to be monitored and use the cotton field soil moisture estimation model; wherein the cotton field soil moisture estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network using a training data set; the training data set includes microwave backscattering signals of the training cotton field area after removing cotton canopy interference, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods, and corresponding soil moisture labels; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training an improved two-dimensional convolutional neural network using a pre-training sample set; the pre-training sample set is simulated using the AIEM-Oh model based on microwave parameters and ground measured soil roughness; the improved two-dimensional convolutional neural network includes sequentially connected 1×1 convolutional layers, 3×3 depth-separable convolutional layers based on Dropout regularization technology, and fully connected layers.

[0031] In the typical Shihezi-Shawan-Karamay cotton region in northern Xinjiang, an improved two-dimensional convolutional neural network (CNN-2D model) was constructed based on the two-dimensional convolutional neural network (CNN-2D model) using microwave data from IW (Interferometric Wideswath) of Sentinel-1, optical data from WFV (16m resolution multispectral camera) of GF-1 (GF-1), and synchronous / quasi-synchronous soil moisture measured data. In order to reduce sample annotation and improve model efficiency and accuracy, the AIEM-Oh mechanism model, microwave parameters, and ground-measured soil roughness were first used to determine the value range of the mechanism model, and a large number of relatively rough data samples were fitted to pre-train the improved two-dimensional convolutional neural network, thereby embedding physical mechanisms in the data-driven scheme. Then, the interference of cotton canopy on scattering was collaboratively eliminated using multi-source, multi-temporal remote sensing satellite data and ground-measured data through a semi-empirical model. Finally, a more accurate small data set of synchronous / quasi-synchronous ground measurements is input into the model, and the pre-trained improved two-dimensional convolutional neural network is fine-tuned to accurately estimate the soil moisture in the cotton field.

[0032] The improved two-dimensional convolutional neural network is pre-trained using the pre-training sample set, i.e., step 1.

[0033] Step 1: Use the AIEM-Oh mechanism model, microwave parameters and ground-measured soil roughness to determine the value range of the mechanism model, and fit a large number of relatively rough data samples for deep learning model pre-training, so as to embed the physical mechanism in the data-driven solution.

[0034] The improved two-dimensional convolutional neural network is pre-trained using the pre-training sample set, including:

[0035] The microwave parameters themselves and the measured soil roughness on the ground are obtained; the microwave parameters themselves include wavelength, incident angle and polarization mode.

[0036] Based on the microwave parameters and the measured soil roughness on the ground, the AIEM-Oh model is used to simulate and obtain a pre-training sample set.

[0037] The improved two-dimensional convolutional neural network is pre-trained using the pre-training sample set to obtain an improved two-dimensional convolutional neural network that is similar to the trained one.

[0038] In practical applications, the representation of radar scattering coefficient on radar images is the grayscale of the image, which is generally related to the parameters of the radar system itself and the surface characteristic parameters, and can be simply expressed by the following formula:

[0039] σ o =f(λ,θ,P)×g(Veg,ε s ,Zs)(1)

[0040] σ o is the radar scattering coefficient; its own parameters mainly include wavelength λ (frequency f), incident angle θ and polarization mode P; surface parameters include vegetation parameters Veg, soil dielectric constant ε s , soil roughness Zs; f, g represent the functional relationship. Therefore, the difference in soil moisture can be reflected by the radar backscatter coefficient when other parameters are determined. 2 / l represents the combined roughness of measured data, so Zs can be used to well characterize the roughness of natural surfaces. A smaller Zs corresponds to a smaller s and a larger l. In this case, the surface characterized by Zs is relatively smooth, and vice versa, the surface is relatively rough. Among them, s represents the standard deviation of the height of a surface, and l represents a measure of the similarity between two height points. It is a benchmark for estimating the independence of two points. That is, if the distance between two points in the horizontal direction is greater than l, then the two points are considered to be approximately independent. Its expression is as follows:

[0041]

[0042] n is the number of profile feature points, h i is the elevation of the ith point, For all profile feature points h i l represents the similarity between the elevation of point x1 on the profile and the elevation of another point x2 deviating from it, which can be obtained by the normalized autocorrelation function of the distance x' = (j-1) Δx (j is a natural number).

[0043]

[0044] ρ(x') is the normalized autocorrelation function, which is used to characterize the similarity of elevation values; h j+i-1 It represents the elevation value of the j+i-1th point; when ρ(x') = 1 / e, the x' value represents the correlation length l, the smaller it is, the greater the surface roughness.

[0045] The AIEM model is used to simulate pre-training samples, thereby obtaining a large number of data samples for deep learning model pre-training. However, the AIEM model can only simulate co-polarized backscattering. For cross-polarized samples, according to previous studies, the improved Oh model can simulate the backscattering of the soil surface more accurately, and the simulation effect of VH polarization (vertical-horizontal polarization) is better than that of VV polarization (vertical polarization). Therefore, the Oh model is used to simulate the VH polarization soil surface backscattering coefficient, and the AIEM model is used to obtain a large number of data samples (pre-training sample set) for pre-training of the improved two-dimensional convolutional neural network.

[0046] Oh model uses M v Instead of dielectric constant, it improves accuracy and reduces input parameters. The applicable parameter range is: 10°<θ<70°, 0.1 <ks<6.0,2.6<kl<19.7,9%<M v <31%; q is the cross-polarization ratio.

[0047] In radar remote sensing, the local angle of incidence refers to the angle formed by the incident beam and the surface normal. When the ground surface is parallel to the horizontal plane, the angle of incidence is the same as the local angle of incidence.

[0048] In addition, the polarization scattering matrix is ​​a matrix used to describe the polarization characteristics of electromagnetic waves in the scattering process, which is divided into a scattering matrix and a polarization matrix. The scattering matrix reveals the changing relationship between the amplitude and phase of the incident wave after it is scattered by an object. Its size is related to the scattering cross section of the object, and it describes the scattering in the vertical and horizontal directions. The polarization matrix reveals the polarization relationship between the incident wave and the scattered wave. Its size is related to the polarization characteristics of the object, and it describes the polarization in the vertical and horizontal directions. Commonly used polarization decomposition methods include Cloude-Pottier decomposition (also known as H / a / A decomposition), which is a classification method based on polarization entropy H (entropy), average scattering angle a (alpha), and anisotropy A (anisotropy) parameters. They can cover all random scattering mechanisms, and H / a / A decomposition is considered to be an effective polarization parameter feature extraction method that is not limited to specific applications. The three parameters are defined as:

[0049]

[0050] Formula 2-Formula 4, a i represents the type of the i-th scattering mechanism; P iRepresents the probability of occurrence of each eigenvector, and its definition formula is as follows:

[0051]

[0052] In the formula, λ i represents the eigenvalue of the polarization coherence matrix C, which can be expressed as:

[0053]

[0054] In the formula <c>represents the geometric mean of C; T represents the conjugate transpose; ∑ represents the characteristic matrix, which is expressed as follows:

[0055]

[0056] λ1, λ2, and λ3 are the plane scattering component, dihedral scattering component, and volume scattering component, respectively.

[0057] U is a unitary matrix, which is a matrix composed of eigenvectors. The formula is as follows:

[0058]

[0059] Among them, α1, α2, and α3 represent specific angles, which are angles related to the eigenvectors; φ1, φ2, and φ3 represent corresponding phase angles; β1, β2, and β3 represent angles related to the sine function; δ1, δ2, and δ3 are phase offsets; γ1, γ2, and γ3 are another set of phase offsets or correction angles.

[0060] Existing measurement methods include pin-type roughness measurement plates, profile plate method, laser profiler and photogrammetry. Photogrammetry can obtain the ground DEM (Digital Elevation Model) raster data through modeling, select contours from any direction, extract roughness parameters, and reduce the cumbersome operations of pin-type and profile plate measurement methods in multi-directional measurements. The shooting position and direction of the camera device and the geometric shape of the target object in the photo can be inferred through the parallax of corresponding pixels in different photos, thereby obtaining the DEM of the target surface. The screenshot of the process is shown below. Figure 2 As shown in the figure, a represents the schematic diagram of the sparse point cloud model of the surface, b represents the schematic diagram of the dense point cloud model of the surface, c represents the schematic diagram of the established DEM model, and d represents the cross-sectional diagram in eight directions. The photographs taken are processed, and the coordinates are marked, the positions are corrected, and a relative three-dimensional coordinate system is established to establish a dense point cloud model of the surface, and finally generate high-precision DEM data of the sample point. The combined roughness range of the surface after mechanical leveling and before film covering is selected as the value interval of the model simulation value.

[0061] The main points that need to be paid attention to during the whole process are: (1) The longitude and latitude of the location where the photo was taken need to be converted into relative positions in meters; (2) The density of the markers should not be too large, which may easily cause mutual occlusion and hinder the subsequent placement of markers and input of scales; (3) The elevation data is measured in 8 directions, with the north direction as 0° and the clockwise direction, every 22.5°. Each measurement takes a length of 50 cm and an interval of 1 mm, and the elevation value of each profile is extracted. Save the data, then calculate the root mean square height and surface correlation length of each profile of the discrete data, and finally average the values ​​of these 8 profiles to obtain the roughness parameters of the sample point. The maximum, minimum and average values ​​of soil s and l of the profiles of the three sample points are shown in Table 1.

[0062] Table 1 Statistical table of root mean square height and correlation length of the surface of various plots

[0063]

[0064]

[0065] In the mechanical drip irrigation cultivation mode, the periodic furrow and ridge structure of the cotton field surface roughness is less prominent than that in the flood irrigation cultivation mode, and the land is leveled before sowing. Therefore, when measuring the soil roughness on site, it is only necessary to determine the orientation but not the row direction.

[0066] The parameters that have a greater impact on the backscattering simulation results in the model are selected, that is, different radar system parameters (θ, f, polarization mode) and surface characteristic parameters (s, l, Mv) are used to simulate the radar backscattering coefficient. The value range is shown in Table 2. A simulation database for studying the surface backscattering characteristics is established and used as the input of the pre-training model for estimating soil moisture in cotton fields in the early stage. According to the soil data set used, the soil texture structure of the surface 0cm-30cm in the study area is 45% sandy, 15% clay, and the remaining 40%; the soil volume is 1.24g / cm 3 ; Use the exponential correlation function to simulate the backscatter coefficients of the surface VV polarization and VH polarization when the frequency is 5.4GHz and the incident angle range is 10°~50°; The surface characteristic parameters are specifically set as follows: The surface temperature takes the average value of the measured values ​​of the sample points during the entire sampling process, 32°C. The range of soil volume moisture content is 4%~40%, and the interval is set to 1%; the range of surface root mean square height is set to 0.2cm~3.6cm, and the interval is set to 0.2cm; the range of correlation length setting is 2.5cm~35cm, and the interval is set to 2.5cm, as shown in Table 2. Other parameters in the AIEM model use measured values, and assume that the roughness of the surface remains unchanged throughout the growth process of the vegetation; Generate a database of backscatter coefficients for bare surfaces. According to the improved first-order coherent scattering model suitable for crops, the vegetation backscattering characteristics are simulated to form the total backscattering characteristics under different incident angles, polarization modes and vegetation moisture content, and establish the surface and soil backscattering characteristic models.

[0067] Table 2 Radar system parameters and surface feature parameter setting ranges

[0068]

[0069] The simulation process is implemented through programs such as Figure 3 The relationship curve between VV polarization backscattering coefficient and soil moisture under the conditions of incident frequency of 5.4GH, p of 1.5, exponential autocorrelation function of the surface autocorrelation function, root mean square height of 2cm, correlation length of 25cm, and incident angles of 15°, 25°, 35°, and 45°. Figure 3 It can be seen that with the increase of soil moisture, the backscattering coefficient shows an increasing trend, but the increasing trend gradually weakens. The backscattering coefficient with a small incident angle is higher than the backscattering coefficient with a large incident angle, and with the increase of the incident angle, the distance between two adjacent curves decreases, indicating that with the increase of the incident angle, its influence on the relationship between soil moisture and backscattering coefficient gradually decreases.

[0070] Given an incident frequency of 5.4 GH, p of 1.5, an incident angle of 32°, a correlation length of 25 cm, and root mean square heights of 0.2 cm, 0.9 cm, 1.8 cm, and 3.6 cm, the relationship curve between the VV polarization backscattering coefficient and soil moisture is as follows: Figure 4 shown.

[0071] from Figure 4 It can be seen that with the increase of soil moisture, the backscattering coefficient shows an increasing trend, but the increasing trend gradually weakens. The backscattering coefficient with a small RMS height is lower than the backscattering coefficient with a large RMS height, and with the increase of the RMS height, the distance between two adjacent curves decreases, indicating that with the increase of the RMS height, its influence on the relationship between soil moisture and backscattering coefficient gradually decreases.

[0072] Given an incident frequency of 5.4 GH, p of 1.5, an exponential autocorrelation function for the surface autocorrelation function, an incident angle of 32°, a root mean square height of 2 cm, and correlation lengths of 5 cm, 15 cm, 25 cm, and 35 cm, the relationship curve between the VV polarization backscattering coefficient and soil moisture is as follows: Figure 5 shown.

[0073] Through the characteristics of spectral reflectance, canopy moisture content, soil moisture and other characteristics, the characteristic laws are analyzed from the mechanism, and based on remote sensing data, B1, B2, B3, B4, NDVI (Normalized Difference Vegetation Index), DVI (Difference Vegetation Index), RVI (Radar Vegetation Index), SAVI (Soil-adjusted Vegetation Index), EVI (Enhanced Vegetation Index) of cotton in different phenological periods are extracted. The feature set for soil moisture estimation is composed of 9 optical characteristic factors, including VV, VH, VV+VH, VH-VV, VH / VV, VV×VH, incident angles, including cosθ, sinθ and θ, polarization decomposition, including Alpha, Anisotropy and Entropy, and surface roughness Zs.

[0074] In order to obtain deep feature relationships while reducing the reliance on samples and confirm the reliability of the method of this application, in addition to parallel comparisons with machine learning models (RF model (Random Forest model), SVM (Support Vector Machine), XGBoost model (eXtreme Gradient Boosting model)), different deep learning model combinations are designed to analyze model accuracy. The specific models and feature sets used are shown in Table 3.

[0075] Table 3 Statistics of model and feature set combinations

[0076]

[0077]

[0078] Step 2: Use a semi-empirical model to collaboratively eliminate the interference of cotton canopy on scattering by using multi-source, multi-temporal remote sensing satellite data and ground-based measured data.

[0079] In areas with bare surfaces and sparse vegetation, microwave inversion of soil moisture has a certain theoretical basis. The integral equation model (IEM) and advanced integral equation model (AIEM) based on the electromagnetic spectrum and radiation transmission mechanism can better simulate the backscattering coefficient of the surface and invert soil moisture. However, the theoretical model is complex and requires many surface parameters. In practical applications, semi-empirical models are often used. In the model, the canopy extinction coefficient (τ) is often used to represent the strength of absorption to characterize the absorption and scattering characteristics of vegetation, so as to reasonably use the scattering model to remove this part of the noise and achieve the purpose of improving the accuracy of the underlying surface soil moisture estimation.

[0080] The water cloud model is a classic semi-empirical vegetation image correction model proposed based on the radiation transfer model. It combines radar and surface parameters and has been widely used in recent years to estimate the backscatter of surface vegetation and soil in vegetation-covered areas. It assumes that "cloud" represents the vegetation layer, which is uniformly distributed in the entire vegetation space by isotropic particles of the same size, similar to water molecules, and is regarded as a layer of water cloud. It has a certain relationship with the crop planting density, height and type, and these three factors are related to the water content of the vegetation canopy. The water cloud model simplifies the complex absorption and scattering process of microwaves in the vegetation layer and between the vegetation layer and the underlying surface, ignoring the multiple scattering between them, only considering single scattering, and classifying the backscattering as the superposition of the two radar signals of the vegetation scattering contribution component and the soil scattering contribution component. For a given incident angle, the water cloud model expression is as follows:

[0081]

[0082] τ 2 (θ) = exp(-2BVsecθ) (13)

[0083] Where: is the total backscatter coefficient of the cotton field surface, and is the scattering component of the cotton canopy and bare soil scattering component (greatly affected by surface soil moisture and roughness); τ 2 is the attenuation factor of radar waves penetrating the cotton canopy; is the contribution of bare soil backscatter after absorption and attenuation by cotton canopy; θ is the radar incident angle; V is the descriptive parameter of cotton canopy, which can be determined by vegetation index, leaf area index or vegetation moisture content in practical applications, depending on the specific situation; A and B are the calibration coefficients of water cloud model, which have different values ​​for different vegetation and are mainly related to vegetation type. There is no common theoretical basis for calculating these two parameters. Generally, a large amount of field observation data and satellite incident angle parameters are brought into the coefficient calibration equation, and the least squares method is used to solve the calibration coefficients A and B in the water cloud model. To ensure the accuracy of the coefficients, the cost function constructed is:

[0084]

[0085] In the formula, σ WCM The backscatter coefficient of the vegetation-covered area is obtained by bringing the bare soil scattering component into the water cloud model. It is a function of vegetation water content, soil volume water content, surface root mean square height, and radar incident angle; σ SAT is the backscatter coefficient obtained by the radar, δσ is the uncertainty parameter, and its value is related to the uncertainty of the radar observation value and the model value. The purpose is to expand the difference between the model simulation value and the radar observation value. The value selection principle is based on the numerical magnitude of the backscatter coefficient of the radar observation value in linear units.

[0086] In order to verify the degree of fit between various vegetation indices and the empirical model of cotton crop canopy moisture content, this application selected five common vegetation indices (NDVI, EVI, SAVI, RVI and DVI) to simulate cotton crop canopy moisture content using a quadratic empirical model based on vegetation index (single parameter). The five common vegetation indices were extracted through GF-1 images, and the accuracy of their estimation results was evaluated and analyzed. The Pearson correlation coefficient was used as the verification indicator to evaluate the accuracy of the vegetation moisture content model; the evaluation of the goodness of fit of the vegetation moisture content model prediction results was based on R 2 It is a statistical indicator used to reflect the degree to which the model explains the variability of observed values.

[0087] The Pearson correlation coefficient is an indicator that measures the degree of linear correlation between two variables. Its value is between -1 and 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation. Its formula is as follows:

[0088]

[0089] Among them, ρ x,y is the Pearson correlation coefficient, x i ,y i are two variables; is the average of two variables; n is the number of variables.

[0090] Coefficient of determination (R 2 ) represents the numerical characteristics of the relationship between a random variable and multiple random variables. It can be used to reflect the goodness of fit of the regression model. It can explain the reliability of the dependent variable change, that is, the degree to which the model explains the variability of the observed values. It can be defined as the ratio of the variation of the independent variable explained by all the independent variables in the model to the total variation of the independent variable. 2 It is usually between [0, 1]. The closer it is to 1, the more it can explain that the fitting curve can explain most of the variation in the data and the better the fitting effect. 2 =1, it means that the fitting curve fully explains the variation of the target variable, and the fitting effect is very good; on the contrary, the closer it is to 0, it means that the fitting curve cannot explain the variation of the data, and the fitting effect is poor. 2 =0, it means that the fitting curve cannot explain the variation of the target variable. The calculation formula is:

[0091]

[0092] Among them, a i is the measured value, is the predicted value and n is the number of samples.

[0093] The determination coefficient and accuracy evaluation index of the fitted vegetation moisture content model are shown in Table 4.

[0094] Table 4 Statistical table of various vegetation indexes and vegetation moisture content model parameters and evaluation indicators

[0095]

[0096] Table 4 shows the fitting and accuracy evaluation of the Gaofen-1 image and field synchronously measured vegetation moisture content data during the boll-setting period of cotton on July 27, 2022. The cotton in the sample plots at this time was in the boll-setting growth stage. The overall average height of each sample plot was 97 cm to 138 cm, and the vegetation moisture content was 3.32 kg / m 2 ~4.58kg / m 2 The empirical single-parameter quadratic models constructed using five vegetation indices showed different accuracy, among which VWC RVI The model showed the highest accuracy with a Pearson correlation coefficient value of 0.9188. Model VWC RVI Using GF-1 satellite image data fitting, the actual measured value and estimated value fitting results, Pearson correlation coefficient, R 2 and RMSE as Figure 6 As shown. The VWC RVI The root mean square error (RMSE) of the estimated value is 0.086 kg / m 2 , R 2 It is 0.8442, which is significantly correlated at the 0.01 level.

[0097] RMSE measures the difference between the predicted value and the measured value. The calculation formula is:

[0098]

[0099] Among them, a i is the measured value, is the predicted value and n is the number of samples.

[0100] Using VWC VI The estimated results of the model are used as the input of the data set to carry out soil moisture estimation simulation experiments: first, the high-precision vegetation scattering component in the water cloud model is obtained, and the parameters A and B are fitted according to the field ground test sampling data. RVI The distribution of the estimated values, such as Figure 7 shown.

[0101] Considering that the soil moisture of the underlying surface with a single vegetation cover type in the study area is estimated, the multi-parameter empirical fitting model is no longer used, but the model with the highest accuracy in each phase is selected as the cotton canopy moisture content estimation model, that is, VWC NDVI As the time estimation model for the budding period (June 06 to June 28) and the flowering period (June 29 to July 18), VWC RVI As the temporal estimation model for the boll-setting period (July 19 to August 18) and the boll-breaking period (August 19 to September 5), the vegetation index and vegetation water content model parameters are shown in Table 5.

[0102] Table 5 Parameters and accuracy evaluation of various vegetation index VWC models

[0103]

[0104] As an optional implementation, the pre-trained improved two-dimensional convolutional neural network is trained using the training data set, i.e., step 3.

[0105] Step 3: Using the pre-trained improved two-dimensional convolutional neural network, input the synchronous / quasi-synchronous ground more accurate measured small data set into the model, and fine-tune the pre-trained improved two-dimensional convolutional neural network to accurately estimate the soil moisture in the cotton field.

[0106] The pre-trained improved two-dimensional convolutional neural network is trained using the training data set, including:

[0107] 1. Obtain the training data set, including:

[0108] 1) Obtain Sentinel-1 microwave data and GF-1 optical data of the cotton field area for training.

[0109] 2) Based on the Sentinel-1 microwave data and the GF-1 optical data of the training cotton field area, microwave characteristic factors and optical characteristic factors of cotton in different phenological stages in the training cotton field area are extracted.

[0110] 3) quantifying the interference of the cotton canopy in the training cotton field area on each microwave backscattering coefficient, and obtaining each microwave backscattering signal after removing the interference of the cotton canopy, specifically including:

[0111] Based on multi-source and multi-temporal remote sensing satellite data and ground-based measured data, the quantitative relationships between water content and the normalized difference vegetation index, difference vegetation index, ratio vegetation index, soil adjusted vegetation index and enhanced vegetation index were obtained using a semi-empirical model.

[0112] Based on the quantitative relationship, the interference of the cotton canopy in the training cotton field area on each microwave backscattering coefficient is quantified, and each microwave backscattering coefficient is obtained after the interference of the cotton canopy is removed.

[0113] 4) The microwave characteristic factors and optical characteristic factors of cotton in different phenological stages in the training cotton field area are measured, and soil moisture labels are marked.

[0114] 2. Preprocess the training data set to obtain a processed training data set; the preprocessing includes removing outliers and cropping images.

[0115] 3. Taking the microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the training cotton field area in the processed training data set and the corresponding soil moisture labels as input, and taking the soil moisture estimation values ​​of cotton in different phenological periods in the training cotton field area as output, the pre-trained improved two-dimensional convolutional neural network is trained to obtain a cotton field soil moisture estimation model.

[0116] The two-dimensional convolutional neural network (CNN-2D) processes two-dimensional or three-dimensional arrays. The three-dimensional data may refer to two-dimensional pixels and RGB channels on a plane in image processing. It is mainly composed of several key components: input layer, convolution layer, pooling layer, fully connected layer and output layer. Its construction process mainly includes setting the number and size of convolution layers, selecting appropriate activation functions, determining the method and parameters of the pooling layer, and determining the number of neurons and output size of the fully connected layer.

[0117] The Dropout layer is a regularization technique widely used in neural networks to avoid overfitting. It reduces the complexity of the model and the risk of overfitting of the neural network by randomly discarding some of the connections of neurons during the training process. It is specifically reflected in the following points: First, when training a neural network, neurons may develop complex co-adaptive relationships, which causes some neurons to be activated only in specific situations, resulting in overfitting on the training data and poor generalization ability on unseen data. The Dropout layer randomly discards the connections of some neurons in each training so that the model does not rely too much on specific neurons and reduces the co-adaptation phenomenon. Secondly, randomness is introduced, and each neuron may be discarded at any time. The Dropout layer forces the model to learn more robust features during the training process, so it needs to learn multiple feature combinations to ensure that it can produce appropriate outputs in different situations to reduce excessive dependence on specific features. Finally, Dropout can be regarded as a method of simulated integration to some extent. Each training iteration is regarded as training in a different sub-network, and the final estimation result is the average or integration of all sub-networks, which helps to reduce parameters.

[0118] The essential task of building the improved CNN-2D model is to learn the intrinsic relationship between the feature set data and the true value data through the microwave features, spectral features and pre-trained models of cotton in different phenological periods constructed by fine training, and the true value of the function output marked on the sample. This relationship can calculate the output of the corresponding model through the newly input features to meet the quantitative relationship between the expression of the training sample and the true value data. At the same time, the original CNN-2D model is replaced with a deep separable network and a Dropout layer is added (i.e., a 3×3 deep separable convolution layer based on the Dropout regularization technology).

[0119] Depthwise separable convolution decomposes the convolution operation into two independent steps: channel-by-channel convolution and point-by-point convolution. It can reduce the number of parameters and computational complexity, while retaining the uniqueness and spatial relationship of features, reducing the risk of overfitting and improving the generalization ability of the model to a certain extent.

[0120] Dropout regularization technology reduces the complexity of the model and reduces the risk of overfitting of the neural network.

[0121] The essence of sample labeling is to provide the neural network with the location information and attribute information of the measured points of the soil moisture on the cotton field surface. Since this application is for the regression problem of continuous variables, it is necessary to ensure that the time and geographic location information of the data are accurate in order to match with the satellite image, that is, the number and position of pixels occupied by the soil moisture information on the corresponding label sample must be consistent. However, although the sampling is carried out as far as possible in the south and north directions, due to the limitation of the row direction, the error between the label data and the pixel is inevitable. Therefore, the labeled data must meet the requirements of "synchronous / quasi-synchronous, accurate, comprehensive, and representative" for the remote sensing data and the measured elements. The specific steps and processes include:

[0122] First, the polarization mode of Sentinel-1 is used as the base map, and the resolution is uniformly sampled to 16m×16m. The soil moisture sampling point information in the same or adjacent time period as the acquisition of this image is selected, and then the geographic coordinates of the sampling points are derived according to the vector map. Then the vector information is imported, and the positions of some sampling points are appropriately adjusted according to the base map so that they fall into a single pixel, and then converted into a raster format. The labeled sample data is the volume water content corrected by the drying method, which represents the actual situation as much as possible. The relative humidity data only retains two decimal places, and the data is standardized to retain four decimal places to improve the training effect and generalization ability of the regression model. Secondly, the characteristic values ​​of the sampling points and their corresponding measured value information are derived from the image in sequence, and each 25 data are taken as a group, and then transposed to n×5×5 (n represents the number of features), so that the image obtained is used as the sample map; the corresponding measured value is obtained by the same method. The 1×5×5 image is used as the annotation map. The above operations are all processed by writing programs. During data measurement, data with deviations exceeding 15% were deleted based on the average value displayed by the instrument.

[0123] The CNN-2D model performs well in automatically extracting data of various types and sizes and learning features from low-level to high-level. The standard CNN-2D model architecture is as follows: Figure 8 shown.

[0124] In order to reduce the dependence on samples and simplify the complexity of the model, the standard two-dimensional convolution layer is replaced by a deep separable network; at the same time, in order to better combine the correlation between input features and soil moisture, the constructed model combines the features extracted by 1×1 convolution and 3×3 convolution. The specific model architecture diagram is as follows: Fig. 9 As shown in the figure, the model consists of three parts. The first part is a convolution layer with a convolution kernel of 1×1. The 1×1 convolution layer can better learn the linear correlation between the input features and soil moisture. The second part is a deep separable network with a convolution kernel of 3×3, which can not only learn the correlation between the input features and soil moisture, but also learn the spatial features between different input samples. The third part is a fully connected layer, which passes the output result of the convolution layer through the fully connected layer to obtain the final soil moisture estimation value.

[0125] Depthwise separable convolution decomposes the convolution operation into two independent steps: channel-by-channel convolution and point-by-point convolution. Although it may have certain limitations when capturing global information and involving complex tasks, it can reduce the number of parameters and computational complexity, while retaining the uniqueness and spatial relationship of features, reducing the risk of overfitting, and improving the generalization ability of the model to a certain extent. Its specific structure is as follows: Fig. 9 This application selects the ReLu function as the activation function, which is placed after each convolutional layer to extract spatial features in the image and perform nonlinear operations; batch normalization is used to accelerate training and improve model stability; mean square error (MSE) is selected as the loss function; the Warm-up attenuation strategy in the single-cycle learning rate attenuation method is used to continuously update the learning rate, and the Dropout regularization technology is used to reduce the complexity of the model and reduce the risk of overfitting of the neural network.

[0126] In the deep separable network with the added Dropout layer determined in this application. The learning rate of the convolutional neural network model is basically controlled between 0.0001 and 0.015; the input pixel size P = 25, the convolution kernel Q = 3 × 3, the step size M = 1, and other parameters remain the default. The learning rate is from 0.0001 to 0.015 for model training and testing. The results of the influence on the convergence speed and overall accuracy are shown in Table 6. When the learning rate is 0.01, the accuracy is optimal, which is the point with the highest accuracy in the learning rate adjustment process of the model, so the learning rate is set to 0.01.

[0127] Table 6 Statistics of the influence of learning rate on convergence speed and overall accuracy

[0128]

[0129]

[0130] Epoch represents the number of iterations after a complete traversal of the training set. In each iteration, the model continuously updates parameters to learn the features in the data. In this application, the parameters are gradually adjusted through prior knowledge and continuous experimental attempts. The changes in the corresponding loss value curve Loss during the training of the model training set and the test set are shown in the figure below. Fig.10 As shown in the figure, train-loss is the corresponding loss value curve during the training of the model training set, val-loss is the corresponding loss value curve during the training of the test set, and the R 2 It can reach 87.33%. In addition, different CNN-2D models are tested in a single direction and compared with this model. The test set Epoch settings and the corresponding MSE and R 2 Changes such as Fig.11 and Fig.12 shown. Fig.11 and Fig.12 1×1CNN refers to a two-dimensional convolutional neural network with a pre-trained convolution kernel of 1×1, 2DCNN refers to a convolutional neural network with a standard 3×3 convolution kernel, and DSN refers to a depthwise separable convolutional neural network. Although all models can reach a convergent state, the improved model has higher accuracy than other models, and its MSE is smaller than that of 1×1CNN, DSN, and 2DCNN. The order of MSE of each model from large to small is 2DCNN, DSN, 1×1CNN, and MyModel, and their highest accuracy values ​​are 0.0426, 0.0380, 0.0184, and 0.0138, respectively. The R of each model is 2 The order from largest to smallest is MyModel, 1×1CNN, DSN, 2DCNN, and their highest accuracy values ​​are 87.33%, 83.15%, 65.20%, and 60.96%, respectively. The improved model used performs relatively well. When the Epoch is around 1200, the model is in a good state, but there is still room for improvement in the later stage, so it is set to 1500.

[0131] In order to determine the test set Epoch and the corresponding MSE and R after adding different features 2 First, different feature sets are added to the improved CNN-2D model to test the accuracy of the estimation results. The test set Epoch settings and the corresponding MSE and R 2 Changes such as Fig.13 and Fig.14 shown. Fig.13 and Fig.14 Multispectral+SAR refers to the feature set constructed by inputting both optical satellite image data and microwave image data, SAR refers to the feature set constructed by inputting only microwave data, and Multispectral refers to the feature set constructed by inputting only optical data. Different features and their combinations were input to the best performing model. Although they all converged, it can be found that the accuracy is higher when the input feature is Multispectral+SAR than other models. The order of MSE from large to small is Multispectral, SAR, and Multispectral+SAR, and the highest accuracy values ​​are 0.001065, 0.000187, and 0.000138, respectively. R of each model 2 The order from large to small is Multispectral+SAR, SAR, and Multispectral, with the highest accuracy values ​​of 87.33%, 82.86%, and 2.44%, respectively. The performance of Multispectral+SAR and SAR as input features is relatively good, indicating that radar features play a decisive role in estimating soil moisture in cotton fields, and the accuracy of soil moisture estimation in cotton fields can be improved by 4.47% after adding optical features.

[0132] In order to evaluate the effectiveness of the method, some mainstream algorithms were selected for experimental comparative analysis. The main machine learning methods (SVM, RF, XGboost) and deep learning models (CNN-1D, CNN-2D, pre-trained CNN-2D, pre-trained deep separable CNN-2D model) were used to compare with the proposed algorithm to demonstrate the performance and accuracy of the model. The model performance was evaluated from three aspects: precision, accuracy, and stability. The main indicators include the determination coefficient R of the training set and the validation set. 2 , RMSE, mean absolute error MAE:

[0133]

[0134] In the formula, y true and pred are the observed value and the predicted value, respectively, and n is the number of samples. The larger the MAE, the greater the deviation between the observed value and the predicted value, and the less ideal the prediction result, and vice versa.

[0135] The data generated by the mechanism model were simulated, and a 5×5 sliding window was used to generate 1500 samples for each soil moisture level. For 35 soil moisture integer values, a total of 52500 samples were generated. In each training, 80% of the samples were randomly selected for training, and the hyperparameters of the training were determined. The model performance and whether it was overfitting were evaluated to obtain the best pre-trained model. The remaining 20% ​​of the samples were tested to preliminarily determine the estimation accuracy, so as to embed the physical mechanism in the data-driven scheme. Then, the interference of cotton vegetation canopy on scattering was synergistically eliminated by using multi-source remote sensing data and ground-truth data through a semi-empirical model. Then, using ground synchronous / quasi-synchronous measured data, 420 5×5 images were used for training, and the pre-trained improved two-dimensional convolutional neural network was fine-tuned on a more accurate ground-truth small data set to more accurately estimate cotton field soil moisture. And 100 images were randomly selected from each test data set to verify the regression network.

[0136] Compare the fitting results of different machine learning models and the pre-trained deep separable CNN-2D model combination with different feature sets. The verification indicators RMSE, R 2 The results of R and MAE are shown in Table 7. It is found that when the semi-empirical model uses optical remote sensing satellite data and ground-based measured data to remove the scattering interference of cotton vegetation canopy, the standard 3×3 convolution kernel CNN-2D (Type4) without mechanism model simulation samples for pre-training is compared with the deep separable network CNN-2D (Type5), the CNN-2D model with mechanism model simulation samples for pre-training (Type6) and the deep separable CNN-2D network with mechanism model simulation samples for pre-training (Type9). The R of Type4 is 2 The RMSE of Type4 is 0.0012, 0.0071 and 0.0089 higher than that of Type5, Type6 and Type9, respectively. The MAE of Type4 is 0.0006, 0.0062 and 0.0079 higher than that of Type5, Type6 and Type9, respectively. 2 R in machine learning models (RF, SVM, XGBoost) 2 The lowest SVM model is still 0.022 less. The main reason is that the CNN-2D model requires a large number of samples, and the measured sample size of this application is far from meeting its sample size requirements in the two-dimensional convolution process. Therefore, the estimation accuracy is significantly low. When pre-training is added, the physical mechanism is embedded in the data-driven scheme, and then a more accurate small data set of synchronous / quasi-synchronous ground measurements is input into the model, the pre-trained improved CNN-2D model is fine-tuned, and on this basis, the improved CNN-2D with added deep separable network is used for simulation, thereby improving the accuracy of Type5, Type6 and Type9.

[0137] In addition, by inputting image data with different feature sets as input items, and inputting them into the pre-trained deep separable CNN-2D model (Type7~9), it is found that after the same previous processing, the model using only the microwave feature set as input item (Type8) performs much better than the model using only the optical feature set as input item (Type7). Its accuracy index R 2 The model with microwave and optical remote sensing image data feature sets (Type 9) has a significant improvement in accuracy. 2 It can reach 0.873, and the RMSE and MAE are 0.0118 and 0.0084 respectively.

[0138] Table 7 Accuracy indicators RMSE, R under different models and different feature set combinations 2 And MAE statistics table

[0139]

[0140] Through the measured soil roughness data, as well as the cotton canopy moisture content and soil moisture data of different phenological periods, the value range of the pre-trained sample set for mechanism model simulation is determined; according to the empirical and semi-empirical models, the vegetation effect in the microwave backscattering coefficient is fitted according to the different phenological periods, and its interference is quantitatively removed; the microwave characteristic factor set and the optical characteristic factor set are constructed according to the different phenological periods, and the model and characteristic factor combination scheme are designed based on the characteristic source.

[0141] The mechanism is embedded in the data-driven scheme, and the nonlinear fitting ability of the two-dimensional convolutional neural network is used. The characteristic factor combination of different phenological periods is used as the input feature to fit the surface soil moisture products of different phenological periods, and the influence of different feature combinations and different phenological periods on soil moisture estimation is explored. In order to reduce sample annotation and improve model efficiency and accuracy, the two-dimensional convolutional neural network model is adjusted using the ground measured data set to construct a remote sensing estimation model of cotton field soil moisture in key phenological periods. The results of this remote sensing estimation model are cross-validated with the exponential model, machine learning model and other deep learning models to compare and evaluate the model accuracy.

[0142] In an exemplary embodiment, a cotton field soil moisture monitoring system is provided, comprising:

[0143] The data acquisition module is used to obtain the Sentinel-1 microwave data and the GF-1 optical data of the cotton field area to be monitored.

[0144] The feature extraction module is used to extract microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the cotton field area to be monitored based on the Sentinel-1 microwave data and the Gaofen-1 optical data; the microwave characteristic factors include the radar backscatter coefficient under vertical polarization mode, the radar backscatter coefficient under vertical-horizontal polarization mode, the combination of the radar backscatter coefficient under vertical polarization mode and the radar backscatter coefficient under vertical-horizontal polarization mode, the incident angle, polarization entropy, average scattering angle, anisotropy and surface combination roughness. The optical characteristic factors include the full color spectrum, blue spectrum, green spectrum, red spectrum, normalized vegetation index, difference vegetation index, ratio vegetation index, soil adjusted vegetation index and enhanced vegetation index.

[0145] A soil moisture detection module is used to estimate the soil moisture of the cotton field area to be monitored based on the microwave characteristic factors and optical characteristic factors of the cotton field area to be monitored and using a cotton field soil moisture estimation model; wherein the cotton field soil moisture estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network using a training data set; the training data set includes microwave backscattering signals of the training cotton field area after removing cotton canopy interference, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods, and corresponding soil moisture labels; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training the improved two-dimensional convolutional neural network using a pre-training sample set; the pre-training sample set is simulated using the AIEM-Oh model based on microwave parameters and ground measured soil roughness; the improved two-dimensional convolutional neural network includes a 1×1 convolutional layer connected in sequence, a 3×3 depth-separable convolutional layer based on Dropout regularization technology, and a fully connected layer.

[0146] In an exemplary embodiment, a computer device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the cotton field soil moisture monitoring method.

[0147] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the cotton field soil moisture monitoring method when executed by a processor.

[0148] In an exemplary embodiment, a computer program product is provided, including a computer program, which implements the above-mentioned cotton field soil moisture monitoring method when executed by a processor.< / c>

Claims

1. A cotton field soil moisture monitoring method, characterized in that: include: Acquire Sentinel-1 microwave data and GF-1 optical data of the cotton field area to be monitored; According to the Sentinel-1 microwave data and the GF-1 optical data, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the cotton field area to be monitored are extracted; the microwave characteristic factors include the radar backscatter coefficient under vertical polarization mode, the radar backscatter coefficient under vertical-horizontal polarization mode, the combination of the radar backscatter coefficient under vertical polarization mode and the radar backscatter coefficient under vertical-horizontal polarization mode, the incident angle, the polarization entropy, the average scattering angle, the anisotropy degree and the surface combination roughness; the optical characteristic factors include the full color spectrum, the blue spectrum, the green spectrum, the red spectrum, the normalized vegetation index, the difference vegetation index, the ratio vegetation index, the soil adjusted vegetation index and the enhanced vegetation index; According to the microwave characteristic factors and optical characteristic factors of the cotton field area to be monitored, the soil moisture of the cotton field area to be monitored is estimated by using a cotton field soil moisture estimation model; wherein the cotton field soil moisture estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network using a training data set; the training data set includes microwave backscattering signals of the training cotton field area after removing cotton canopy interference, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods, and corresponding soil moisture labels; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training an improved two-dimensional convolutional neural network using a pre-training sample set; the pre-training sample set is simulated using the AIEM-Oh model based on microwave parameters and ground measured soil roughness; the improved two-dimensional convolutional neural network includes a 1×1 convolutional layer connected in sequence, a 3×3 depth-separable convolutional layer based on Dropout regularization technology, and a fully connected layer.

2. The cotton field soil moisture monitoring method according to claim 1, characterized in that: The improved two-dimensional convolutional neural network is pre-trained using the pre-training sample set, including: Obtaining microwave parameters and ground-measured soil roughness; the microwave parameters include wavelength, incident angle and polarization mode; Based on the microwave parameters and the measured soil roughness on the ground, a pre-training sample set is obtained by simulating the AIEM-Oh model; The improved two-dimensional convolutional neural network is pre-trained using the pre-trained sample set to obtain a pre-trained improved two-dimensional convolutional neural network.

3. The cotton field soil moisture monitoring method according to claim 1, characterized in that: The pre-trained improved two-dimensional convolutional neural network is trained using the training data set, including: Get the training dataset; Preprocessing the training data set to obtain a processed training data set; the preprocessing includes removing outliers and cropping images; The microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the training cotton field area in the processed training data set and the corresponding soil moisture labels are used as input, and the soil moisture estimation values ​​of cotton in different phenological periods in the training cotton field area are used as output, and the pre-trained improved two-dimensional convolutional neural network is trained to obtain a cotton field soil moisture estimation model.

4. The cotton field soil moisture estimation method according to claim 3, characterized in that: Obtain the training data set, including: Acquire Sentinel-1 microwave data and Gaofen-1 optical data of the cotton field area used for training; Extracting microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the training cotton field area based on the Sentinel-1 microwave data and the Gaofen-1 optical data of the training cotton field area; quantifying the interference of the cotton canopy in the training cotton field area on each microwave backscattering coefficient, and obtaining each microwave backscattering signal after removing the interference of the cotton canopy; Soil moisture labels are marked on microwave characteristic factors and optical characteristic factors of cotton in different phenological stages in the training cotton field area.

5. The cotton field soil moisture monitoring method according to claim 4, characterized in that: Quantifying the interference of the cotton canopy in the training cotton field area on each microwave backscattering coefficient, and obtaining each microwave backscattering coefficient after removing the interference of the cotton canopy, specifically includes: Based on multi-source and multi-temporal remote sensing satellite data and ground-measured soil roughness, the semi-empirical model was used to obtain the quantitative relationship between water content and the normalized vegetation index, difference vegetation index, ratio vegetation index, soil-adjusted vegetation index and enhanced vegetation index. Based on the quantitative relationship, the interference of the cotton canopy in the training cotton field area on each microwave backscattering coefficient is quantified, and each microwave backscattering coefficient is obtained after the interference of the cotton canopy is removed.

6. A cotton field soil moisture monitoring system, characterized in that: include: A data acquisition module, used to acquire the Sentinel-1 microwave data and the GF-1 optical data of the cotton field area to be monitored; A feature extraction module is used to extract microwave characteristic factors and optical characteristic factors of cotton in different phenological periods in the cotton field area to be monitored according to the Sentinel-1 microwave data and the Gaofen-1 optical data; the microwave characteristic factors include the radar backscatter coefficient under vertical polarization mode, the radar backscatter coefficient under vertical-horizontal polarization mode, the combination of the radar backscatter coefficient under vertical polarization mode and the radar backscatter coefficient under vertical-horizontal polarization mode, the incident angle, polarization entropy, average scattering angle, anisotropy and surface combination roughness; the optical characteristic factors include the full color spectrum, blue spectrum, green spectrum, red spectrum, normalized vegetation index, difference vegetation index, ratio vegetation index, soil adjusted vegetation index and enhanced vegetation index; A soil moisture detection module is used to estimate the soil moisture of the cotton field area to be monitored based on the microwave characteristic factors and optical characteristic factors of the cotton field area to be monitored and using a cotton field soil moisture estimation model; wherein the cotton field soil moisture estimation model is obtained by training a pre-trained improved two-dimensional convolutional neural network using a training data set; the training data set includes microwave backscattering signals of the training cotton field area after removing cotton canopy interference, microwave characteristic factors and optical characteristic factors of cotton in different phenological periods, and corresponding soil moisture labels; the pre-trained improved two-dimensional convolutional neural network is obtained by pre-training the improved two-dimensional convolutional neural network using a pre-training sample set; the pre-training sample set is simulated using the AIEM-Oh model based on microwave parameters and ground measured soil roughness; the improved two-dimensional convolutional neural network includes a 1×1 convolutional layer connected in sequence, a 3×3 depth-separable convolutional layer based on Dropout regularization technology, and a fully connected layer.

7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the cotton field soil moisture monitoring method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the cotton field soil moisture monitoring method described in any one of claims 1 to 5 is implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the cotton field soil moisture monitoring method described in any one of claims 1 to 5 is implemented.

Images