A soil moisture multi-source remote sensing monitoring method, device and storage medium

Abstract

The present application relates to the field of remote sensing monitoring, and discloses a kind of soil humidity multi-source remote sensing monitoring method, equipment and storage medium, method includes steps: obtaining multi-source satellite remote sensing data;Multi-source satellite remote sensing data is preprocessed, and the data after preprocessing is obtained;According to the data after preprocessing, data set is constructed;Three-dimensional joint convolutional neural network is constructed, and three-dimensional joint convolutional neural network is trained based on data set, and the network after training is completed is obtained;Soil humidity is predicted using the network after training is completed.The beneficial effects of the present application are: effectively solve the problem of multi-source remote sensing data fusion and information complementary deficiency in the prior art, realize the information complementary and feature fusion of multi-source remote sensing data, enhance the robustness of the model, improve the prediction accuracy and stability;It has the characteristics of simple and effective, without the need for research and analysis of complex remote sensing mechanism, more accurately approximates the complex nonlinear function between remote sensing information and ground object parameters.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring, and in particular to a method, equipment and storage medium for multi-source remote sensing monitoring of soil moisture. Background Technology

[0002] Soil, as an important natural resource on the Earth's surface, has parameters (such as humidity, temperature, and organic matter content) that are of great significance to agricultural production, ecological environmental protection, and water resource management. Traditional methods for measuring soil parameters, such as field sampling and laboratory analysis, can provide relatively accurate data, but they are cumbersome, time-consuming, and labor-intensive, and it is difficult to achieve large-scale, high-frequency real-time monitoring.

[0003] For decades, remote sensing technology has been used for soil moisture estimation. For example, Hassan-Esfahani et al. developed a neural network method that uses high spatial resolution UAV imagery to provide reasonably accurate soil moisture estimates for irrigated farmland, with a root mean square error of less than 0.3 m. 3 / m 3 Thermal infrared remote sensing inversion; Sadeghi and Babaeian et al. proposed a novel optical trapezoidal model to replace the traditional triangle / trapezoidal method, which can estimate soil moisture from reflectance data in the Sentinel-2, Landsat-8, and MODIS shortwave infrared bands. This method is applicable to new high-resolution optical satellites that do not have thermal infrared bands. Microwave remote sensing technology further obtains the dielectric properties of water, air, and solids by capturing the emissivity and backscattering information of microwaves on the soil surface. Le Hégarat-Mascle et al. proposed a soil moisture monitoring method based on ERS / SAR data, based on the prior information of the linear relationship between soil moisture and SAR signals. The results show that multi-source remote sensing technology provides a very powerful means for monitoring near-surface soil moisture.

[0004] However, remote sensing inversion of soil moisture involves the fusion of multi-source data and information complementarity, and few studies have considered the impact of multi-source data on soil moisture inversion. Summary of the Invention

[0005] The purpose of this invention is to propose a multi-source remote sensing monitoring method, equipment, and storage medium for soil moisture, thereby solving the technical problem of inaccurate soil moisture inversion results caused by the lack of multi-source information fusion in current soil moisture remote sensing inversion research.

[0006] This invention provides a multi-source remote sensing monitoring method for soil moisture, comprising the following steps:

[0007] S1. Acquire multi-source satellite remote sensing data;

[0008] S2. Preprocess the multi-source satellite remote sensing data to obtain preprocessed data;

[0009] S3. Construct a dataset based on the preprocessed data;

[0010] S4. Construct a three-dimensional joint convolutional neural network and train it based on the dataset to obtain the trained network.

[0011] S5. Use the trained network to predict soil moisture.

[0012] A storage medium that stores instructions and data for implementing a multi-source remote sensing monitoring method for soil moisture.

[0013] A multi-source remote sensing monitoring device for soil moisture includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a multi-source remote sensing monitoring method for soil moisture.

[0014] The beneficial effects provided by this invention are as follows: By using a parallel structure of spatial spectral feature extraction blocks and complex-valued three-dimensional convolutional blocks, the spatial spectral structure of multispectral remote sensing images and polarized complex images is extracted. This not only includes the contextual information of the remote sensing images but also ensures the existence of spectral and phase information, effectively solving the problem of insufficient multi-source remote sensing data fusion and information complementarity in existing technologies. It realizes information complementarity and feature fusion of multi-source remote sensing data, enhances the robustness of the model, and improves prediction accuracy and stability. At the same time, this invention is based on convolutional neural network technology, which has the characteristics of simplicity and effectiveness. It does not require research and analysis of complex remote sensing mechanisms and can more accurately approximate the complex nonlinear functions between remote sensing information and ground object parameters. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0016] Figure 2 This is a structural diagram of the 3D-Unified-CNN model;

[0017] Figure 3 This is a structural diagram of the spatial spectral feature extraction block;

[0018] Figure 4 This is a structural diagram of a complex-valued 3D convolutional block;

[0019] Figure 5 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0021] Before formally describing the present invention, the relevant technical terms used in the present invention will be explained. In addition, the solution of the present invention will be given a general description first to facilitate understanding.

[0022] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the method of the present invention.

[0023] This invention provides a method for multi-source remote sensing monitoring of soil moisture, comprising the following steps:

[0024] S1. Acquire multi-source satellite remote sensing data;

[0025] It should be noted that step S1 is as follows:

[0026] S11. Obtain experimental data from multiple soil samples;

[0027] S12. Obtain radar complex images (IMG) from the corresponding radar data source based on soil sample experimental data. CV and multispectral band images IMG MSSI .

[0028] As one embodiment, the present invention selects experimental data from n soil samples, with the sample collection location being P. i =(x i y i ), i = 1, 2, ..., n, the corresponding soil moisture data is Y = {Y1, Y2, ..., Yn} n The collection dates are D = {D1, D2, ..., D}. n}

[0029] Based on the collection date D of each sample i With geographical location P i D, obtained from the same radar data source i Multipolar / Allpolarized Radar Complex Images of the Day (IMG) CV Simultaneously, distance D from the same optical data source was obtained. i Recent multispectral band image (IMG) MSSI Among them, IMG CV It must contain amplitude and phase information for at least two of the polarization channels: HH, HV, and VV. (IMG) MSSI It contains information from at least four bands: blue, green, red, and near-infrared.

[0030] S2. Preprocess the multi-source satellite remote sensing data to obtain preprocessed data;

[0031] It should be noted that step S2 specifically involves:

[0032] S21, Regarding radar complex image IMG CV Radiometric calibration, multi-view processing, speckle filtering, and geocoding are performed sequentially to obtain the processed image IMG'. MSSI ;

[0033] S22, For multispectral band images IMG MSSI Radiometric calibration, atmospheric correction, and orthorectification were performed sequentially to obtain the processed image IMG'. MSSI .

[0034] As one embodiment, this invention uses ENVI 5.3 or other professional remote sensing software to perform radiometric calibration, atmospheric correction, and orthorectification on multispectral images; and uses ESA SNAP or other radar image processing software to perform radiometric calibration, multi-view processing, speckle filtering, and geocoding on polarimetric radar images.

[0035] S3. Construct a dataset based on the preprocessed data:

[0036] It should be noted that the process of constructing the dataset in step S3 is as follows:

[0037] S31. Using the latitude and longitude coordinates P of the i-th data point in the soil sample experimental data. i =(x i y i ), where i = 1, 2, ..., n are the center points, and the multispectral band image IMG' is processed in step S2. MSSI Cut out the data X from the H1×W1 window MSSI (H1, W1, D1), where D1≥4, H1=W1=2k+1, k∈Z;

[0038] S32, with P i =(x i y i ), i = 1, 2, ..., n are the centers of the complex radar image IMG' CV Crop the data X from the H2×W2 window CV (H2, W2, D2, 2), where D2≥2, H2=W2=2k+1, k∈Z, and the last dimension 2 represents the real and imaginary parts of the complex number;

[0039] S33. Taking the center of the window (H2×W2) as the origin, set the width and height of each pixel to 1. Then, based on the geographical coordinates (x, y, y) of the four corner points of the center pixel... left x right y up y down Calculate P for each sampling point i Relative position in the window The calculation formula is as follows:

[0040]

[0041] Finally, we obtain n datasets A = [A1, A2, ..., A...]. n};

[0042] in The dataset is randomly divided into a training set and a validation set according to a predetermined ratio. Preferably, the predetermined ratio is set to 8:2.

[0043] In one embodiment, step S3 requires finding the corresponding pixel location in the image based on the geographical location of the soil sample, and extracting data of a specified window size centered on this location. Assuming all sampling points are distributed within an image range, and the latitude and longitude information of the sampling points has been stored in a CSV file, the present invention provides an example of some pseudocode as follows:

[0044]

[0045]

[0046] The above is for illustrative purposes only.

[0047] Similarly, for local data extraction from polarimetric radar images, the same implementation method can be used. After extracting the data for the corresponding region, the matrix can be reshaped using the reshape function.

[0048] When calculating the relative position of a sampling point within a local window of a complex image, the following calculation formula can be used as a reference:

[0049]

[0050] Where (x) left x right y up y down (x) represents the corner coordinates of the center pixel of the local window. i y i ) represents the geographical coordinates of the sampling point.

[0051] After merging all datasets, you can use the scikit-learn library in Python to randomly divide the dataset into training and validation sets proportionally. The pseudocode is as follows:

[0052] train_set, validation_set=train_test_split(dataset_A, test_size=0.2, random_state=42).

[0053] S4. Construct a three-dimensional joint convolutional neural network and train it based on the dataset to obtain the trained network.

[0054] It should be noted that the construction process of the aforementioned three-dimensional joint convolutional neural network is as follows:

[0055] Construct two input layers, each receiving a multispectral image X. MSSI (H1, W1, D1), complex image X CV (H2, W2, D2, 2);

[0056] Construct a spatial spectral feature extraction block and accept X MSSI After processing by the spatial spectral feature extraction block, the spatial spectral feature F is obtained. MSSI (C1, 3, 3, 1), where C1 is the number of channels in the spatial spectral feature.

[0057] Construct complex-valued 3D convolutional blocks, receiving X CV After processing with complex-valued 3D convolutional blocks, the complex-valued feature F is obtained. CV (N2, 3, 3, 1, 2), where C2 is the number of channels for the complex-valued feature.

[0058] Spatial spectral features F MSSI and complex-valued feature F CV With relative position parameter P REL The spatial spectral feature vector I at the target location is obtained after passing through a variable convolutional layer. MSSI (1, C1) and the real eigenvector I CV-real (1, C2), imaginary eigenvector I CV-imag (1, C2);

[0059] Finally, merge the spatial spectral eigenvectors I MSSI and complex-valued eigenvectors I CV Obtain the composite feature vector I COM After passing through three fully connected layers FC1, FC2, and FC3, the soil moisture can be accurately predicted. (1, C1+2×C2)

[0060] As one example, please refer to Figure 2 , Figure 2 The structure diagram of the 3D-Unified-CNN deep learning model proposed in this invention is shown. In establishing the 3D-Unified-CNN deep learning model (S4), two input layers are defined, each receiving a multispectral image X. MSSI (H1, W1, D1), complex image X CV (H2, W2, D2, 2); Construct a spatial spectral feature extraction block and a complex-valued 3D convolution block, and adopt a parallel structure to improve computational efficiency.

[0061] After obtaining the spatial spectral features and complex value features, they are input into a variable convolutional layer to obtain three one-dimensional vectors. These vectors are then merged and input into a fully connected layer. After passing through three fully connected layers FC1, FC2, and FC3, accurate prediction of soil moisture is achieved.

[0062] Further, please refer to Figure 3 , Figure 3 The diagram shows the structure of the spatial spectral feature extraction block in the model, which consists of 3 consecutive convolutional layers, 1 pooling layer, and 2 consecutive convolutional layers. The inputs of the first 3 consecutive convolutional layers are the set of inputs and outputs of the upper layers.

[0063] Further, please refer to Figure 4 , Figure 4 This is a structural diagram of a complex-valued three-dimensional convolutional block.

[0064] Figure 4 (a) in the diagram shows the structure of the complex-valued 3D convolutional block in the model, which is a fully convolutional layer structure. Figure 4 (b) illustrates the steps of complex-valued convolution processing on the input image X. CV The complex-valued 3D convolutional block processed by (H2, W2, D2, 2) is a fully convolutional layer structure. In the complex-valued 3D convolutional processing, the input tensor is first separated into real and imaginary parts, and then 3D convolution operations are performed on the real and imaginary parts respectively to calculate the real and imaginary parts of the output complex number. Finally, the results are merged into a complex tensor.

[0065] In step S4, the activation function for the convolutional layers in the 3D-Unified-CNN deep learning model is ReLU, and the activation function for the first two fully connected layers is Softplus, as shown in the following formula:

[0066] ReLU(x) = max(0, x),

[0067] Softplus(x) = log(1 + e) x ),

[0068] The 3D-Unified-CNN model was trained using the training set. The model's parameters were optimized using the Adam optimizer, and the loss function was the Huber function, as shown in the following formula:

[0069]

[0070] Where Δy is the difference between the predicted value and the true value, and δ is a hyperparameter, which defaults to 1.

[0071] In predicting and assessing soil moisture (S5), multi-source remote sensing image data from the validation set are input into a trained 3D-Unified-CNN model. The root mean square error (RMSE) of the model's output is used to evaluate the predictive performance, and its calculation formula is as follows:

[0072]

[0073] In the formula, k is the number of test samples, and Y i These are the actual values ​​of the test samples. These are the predicted values ​​for the test samples.

[0074] S5. Use the trained network to predict soil moisture.

[0075] As one embodiment, this invention selects 200 sets of soil moisture sample data and collects multispectral band images (IMG) from the Gaofen-1 satellite based on the spatiotemporal information of the samples. MSSI (Blue band, green band, red band, near-infrared band, 8m resolution), and complex images (IMG) of the HH, HV, and VV polarization channels of the Gaofen-3 satellite's fully polarimetric stripe (QPS I). CV (Amplitude, phase, 8m resolution).

[0076] After preprocessing the Gaofen-1 and Gaofen-3 images using ENVI 5.3 and ESA SNAP respectively, 23x23 windows of data X were cropped from the Gaofen-1 image. MSSI (23, 23, 4), data X is cropped from the Gaofen-3 image into a 7x7 window. CV (7, 7, 6) is then reshaped into X. CV (7, 7, 3, 2).

[0077] And calculate the sampling point P based on the geographic coordinates of the four corner points of the center pixel. i Relative position in the window Finally, we obtain 200 datasets A = {A1, A2, ..., A...} 200},

[0078] in

[0079] Dataset A was randomly divided into a training set and a validation set in an 8:2 ratio. The training set contained 160 samples, and the validation set contained 40 samples.

[0080] When constructing the 3D-Unified-CNN deep learning model, two input layers are defined, each accepting a multispectral image X. MSSI (23, 23, 4), Fully Polarized Complex Image X CV(7, 7, 3, 2); Construct a spatial spectral feature extraction block and a complex-valued three-dimensional convolution block, and adopt a parallel structure to improve computational efficiency.

[0081] X MSSI After processing through multiple extraction blocks, the spatial spectral feature F is obtained. MSSI (64, 3, 3, 1), X CV After processing through multiple convolutional blocks, the complex-valued feature F is obtained. CV (16, 3, 3, 1, 2).

[0082] Based on relative position After passing through a variable convolutional layer, the spatial spectral feature vector I at the target location is obtained. MSSI (1, 64) and the real eigenvector I CV-real (1, 16), Imaginary eigenvector I CV-imag (1, 16); after merging, the eigenvector I is obtained. ALL The (1, 96) input fully connected layer is used to achieve accurate prediction of soil moisture through three fully connected layers FC1, FC2, and FC3. FC1 has 96 neurons, FC2 has 32 neurons, and FC3 has 1 neuron, which is the predicted value of soil moisture.

[0083] The structures of the spatial spectral feature extraction block and the complex-valued 3D convolution block are shown in the table below.

[0084] Table 1. Structure of the spatial spectral feature extraction block and the complex-valued 3D convolution block.

[0085]

[0086] 160 training samples were input into the 3D-Unified-CNN deep learning model for training. The batch size was 16, and the number of iterations was 100. The model was optimized using the Adam optimizer with the Huber loss function. Early stopping was used to stop training when the performance on the validation set no longer improved, thus avoiding overfitting caused by excessive iterations.

[0087] Forty validation samples were input into the trained 3D-Unified-CNN model for prediction, and the root mean square error was used to evaluate the prediction performance. Finally, the model that passed the performance evaluation was used for soil moisture inversion prediction.

[0088] Please see Figure 5 , Figure 5 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a multi-source remote sensing monitoring device for soil moisture 401, a processor 402, and a storage medium 403.

[0089] A multi-source remote sensing monitoring device for soil moisture 401: The multi-source remote sensing monitoring device for soil moisture 401 implements the multi-source remote sensing monitoring method for soil moisture.

[0090] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the multi-source remote sensing monitoring method for soil moisture.

[0091] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the above-mentioned method for multi-source remote sensing monitoring of soil moisture.

[0092] The beneficial effects of this invention are as follows: By using a parallel structure of spatial spectral feature extraction blocks and complex-valued three-dimensional convolutional blocks, the spatial spectral structure of multispectral remote sensing images and polarized complex images is extracted. This not only includes the contextual information of the remote sensing images but also ensures the existence of spectral and phase information, effectively solving the problem of insufficient multi-source remote sensing data fusion and information complementarity in existing technologies. It realizes information complementarity and feature fusion of multi-source remote sensing data, enhances the robustness of the model, and improves prediction accuracy and stability. At the same time, this invention is based on convolutional neural network technology, which has the characteristics of simplicity and effectiveness. It does not require research and analysis of complex remote sensing mechanisms and can more accurately approximate the complex nonlinear functions between remote sensing information and ground object parameters.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-source remote sensing monitoring of soil moisture, characterized in that: The method includes the following steps: S1. Acquire multi-source satellite remote sensing data; S2. Preprocess the multi-source satellite remote sensing data to obtain preprocessed data; S3. Construct a dataset based on the preprocessed data; S4. Construct a three-dimensional joint convolutional neural network and train it based on the dataset to obtain the trained network. S5. Predict soil moisture using the trained network; Step S1 is as follows: S11. Obtain experimental data from multiple soil samples; S12. Obtain complex radar images from corresponding radar data sources based on soil sample experimental data. and multispectral band images ; Step S2 specifically includes: S21, Regarding radar complex images The image is obtained by sequentially performing radiometric calibration, multi-view processing, speckle filtering, and geocoding. ; S22, Multispectral images The image is obtained by sequentially performing radiometric calibration, atmospheric correction, and orthorectification. ; The process of constructing the dataset in step S3 is as follows: S31, based on the experimental data of the soil samples, the first... i Latitude and longitude coordinates of each data point Centered on the multispectral band image processed in step S2, Cut out from the middle Window data ,in ; S32, with Centered on radar complex images Cut out from the middle Window data ,in The last dimension, 2, represents the real and imaginary parts of the complex number; S33, using a window With the center as the origin, each pixel is set to a width and height of 1, and the coordinates of the four corner points of the center pixel are used as the basis for the calculation. Calculate each sampling point Relative position in the window The calculation formula is as follows: , Finally, n datasets are obtained. ; in The dataset is randomly divided into a training set and a validation set according to a predetermined ratio. The construction process of the three-dimensional joint convolutional neural network is as follows: Construct two input layers, each receiving multispectral band images. Complex images ; Construct a spatial spectral feature extraction block and accept... After processing by the spatial spectral feature extraction block, spatial spectral features are obtained. ,in The number of channels for the empty spectral features; Construct complex-valued 3D convolutional blocks and accept... After processing with complex-valued 3D convolutional blocks, complex-valued features are obtained. ,in The number of channels for the complex-valued feature; spatial spectral features and complex value characteristics With relative position parameters The spatial spectral feature vector at the target location is obtained after passing through a variable convolutional layer. and real eigenvectors Imaginary eigenvectors ; Finally, merge the spatial spectral eigenvectors. and complex-valued eigenvectors Obtain the composite feature vector After passing through three fully connected layers To achieve accurate prediction of soil moisture; For the input image The processed complex-valued 3D convolutional blocks are fully convolutional layer structures. In the complex-valued 3D convolution processing, the input tensor is first separated into real and imaginary parts, and then 3D convolution operations are performed on the real and imaginary parts respectively to calculate the real and imaginary parts of the output complex number. Finally, the results are merged into a complex tensor.

2. The method for multi-source remote sensing monitoring of soil moisture as described in claim 1, characterized in that: The 3D joint convolutional neural network is trained using the training set. The model's parameters are optimized using the Adam optimizer, and the loss function is the Huber function, as shown in the following formula: ， in, It is the difference between the predicted value and the actual value. This is a hyperparameter, defaulting to 1; after training, it is applied from the fully connected layer. The obtained activation values ​​are the soil moisture monitoring results obtained by the three-dimensional joint convolutional neural network based on the extracted image features.

3. A storage medium, characterized in that: The storage medium stores instructions and data to implement the multi-source remote sensing monitoring method for soil moisture as described in any one of claims 1 to 2.

4. A multi-source remote sensing monitoring device for soil moisture, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the multi-source remote sensing monitoring method for soil moisture as described in any one of claims 1 to 2.

Images