Soil humidity high-precision prediction method based on space-time convolutional neural network coupling model

Abstract

The invention relates to the technical field of soil humidity prediction, and discloses a high-precision soil humidity prediction method based on a space-time convolutional neural network coupling model, which comprises the following steps of: deploying soil humidity detection equipment with a GPS (Global Position System) function to obtain measured soil humidity data of different depths of a target area, and obtaining a satellite remote sensing image of the target area; after the image is corrected, humidity data of the whole area is obtained, and errors of remote sensing humidity data are corrected through sensors and earth surface factors. According to the regional soil humidity dynamic monitoring method based on multi-source data fusion and deep learning, limitation caused by a single data source is avoided by fusing a soil humidity sensor carrying a GPS and satellite remote sensing data, and the monitoring precision of soil humidity is improved by introducing earth surface factors to correct remote sensing data. The result shows that the early warning forecast accuracy is not lower than 90%. Utilizing a machine learning model to expand measured data of the humidity sensor, and predicting humidity change of a certain time window in the future.

Description

Technical Field

[0001] The present invention belongs to the technical field of soil moisture prediction and forecasting, and in particular, relates to a high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model. Background Art

[0002] Soil moisture is one of the important factors affecting the occurrence of landslides. Changes in soil moisture directly affect the physical and mechanical properties of the soil. In particular, when precipitation or snowmelt causes the soil water content to increase, the cohesion and friction of the soil will decrease, threatening the stability of the slope. Excessive soil moisture can lead to soil saturation, thereby increasing the pore water pressure, weakening the adhesion between soil particles, and ultimately causing landslides. In actual monitoring, soil moisture is regarded as one of the important indicators of landslide precursors. By monitoring changes in soil moisture, potential landslide risks can be identified in a timely manner. Therefore, in the landslide early warning system, reasonable monitoring and analysis of soil moisture is the key to improving the early warning and protection capabilities of landslide disasters. Combining meteorological data and geological conditions, establishing a comprehensive landslide early warning model can more accurately predict the occurrence of landslides and provide a scientific basis for decision-making departments.

[0003] Common monitoring methods include burying soil moisture sensors and using satellite remote sensing technology. In particular, the development of remote sensing technology in recent years has made soil moisture monitoring over a large area more feasible, effectively improving the accuracy of landslide warning. First, although buried soil moisture sensors can provide local high-precision monitoring data, the distribution density of sensors is limited by terrain conditions and economic costs, and it is difficult to cover the entire landslide area. The deployment and maintenance of sensors are also restricted by factors such as complex terrain and inconvenient transportation. In addition, the long-term stability of sensor data is also affected by changes in soil physical and chemical properties, which requires regular calibration and maintenance of sensors. On the other hand, although remote sensing technology can obtain soil moisture information over a large area, the soil moisture data obtained by remote sensing is difficult to achieve the same accuracy as sensors because the data accuracy is limited by spatial resolution and temporal resolution. At the same time, the monitoring results of remote sensing technology are subject to external environmental conditions, such as vegetation coverage, surface disturbance, and atmospheric conditions. For example, the shielding effect of vegetation may lead to deviations in monitoring data, and atmospheric water vapor and clouds will also affect the inversion accuracy of soil moisture. This makes the fusion of remote sensing data and ground sensor data an important research direction in soil moisture monitoring.

[0004] In future research, how to combine the advantages of sensors and remote sensing technology, use data assimilation and multi-source information fusion technology, and organically combine ground monitoring with remote sensing observations is the key to solving the current problems of soil moisture monitoring accuracy and spatial coverage. Through this fusion method, the soil moisture change process before the landslide can be more comprehensively grasped, thereby further improving the monitoring and early warning capabilities of landslide disasters.

[0005] In view of this, the present invention is proposed. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art. The present invention provides a high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model, which has the advantage of high precision in landslide deformation monitoring and solves the problem of single data of GPS ground measured data and InSAR data.

[0007] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0008] A high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model comprises the following steps:

[0009] S1: Deploy soil moisture detection equipment equipped with GPS function to the target area to obtain soil moisture measurement data at different depths;

[0010] S2: Obtain remote sensing satellite images of the target area, import the surface factors into the deep learning model for training, use the trained learning model to correct the remote sensing images and divide different types of terrain; the specific steps of constructing remote sensing image optimization and extracting features are:

[0011] S2.1: Define the three-dimensional coordinate matrix C = [xyz] of the humidity sensor in the x, y, and z directions, where C is the three-dimensional coordinate quantity, and xyz is the coordinate value of the sensor in the north-south, east-west, and vertical directions respectively;

[0012] S2.2: Define the surface factor matrix E=[SAV] around the humidity sensor deployment point, where E is the surface factor, and SAV is the slope, slope aspect, and vegetation coverage around the deployment point;

[0013] S2.3: In the input layer, the coordinate matrix of the known points and the terrain data are input into the convolutional neural network as an independent feature channel. Construct the remote sensing image optimization matrix;

[0014] I'=[I,C,E]

[0015] Among them, I' is the enhanced image matrix, I is the original remote sensing image, C is the three-dimensional coordinate matrix of the known points, and E is the terrain information matrix, etc.;

[0016] S2.4: Use CNN to extract the spatial features of the target remote sensing image. For the input image I', the feature map is obtained after the convolution operation of the CNN layer. The expression is:

[0017]

[0018] Among them, l represents the number of convolution layers, k represents the number of convolution kernels, is the weight of the kth convolution kernel, * represents the convolution operation, is the bias term, σ is the activation function;

[0019] S2.5: Combine the surface factor matrix E and the spatial feature matrix Fusion, fusion module X t The expression is:

[0020]

[0021] Among them, α t is the attention weight of the surface factor;

[0022] S2.6: Build an LSTM model and transform X t Input the LSTM model to extract the temporal features of the target remote sensing image. The LSTM core expression is:

[0023] Determine which information in the memory unit needs to be forgotten. The forget gate expression is:

[0024] f t =ρ(W f ·[h t-1 ,Q t ]+d f )

[0025] Among them, f t is the forgotten information, ρ is the activation function, W f is the weight matrix of the forget gate, h t-1 represents the hidden state at the previous moment, Q t represents the input at the current moment, d f is the bias term of the forget gate;

[0026] Determine what new information needs to be stored in the memory cell. The input gate expression is:

[0027] i t =ρ(W i ·[h t-1 ,Q t ]+d i )

[0028] Among them, Wi is the weight matrix of the input gate, d i is the bias term of the input gate;

[0029] Candidate memory cell states, The expression is:

[0030]

[0031] Among them, W u is the weight matrix of the candidate state, d u is the bias term, and tanh is the activation function;

[0032] Update the memory unit state, U t The expression is:

[0033]

[0034] Among them, U t-1 Indicates the state of the memory unit at the previous moment;

[0035] Determine the hidden state O of the output t , the output gate expression is:

[0036] O t =ε*W O ·[h t-1 ,Q i ]+d O

[0037] Among them, ε represents the activation function of the output gate, W O is the weight matrix of the output gate, d O is the bias term of the output gate;

[0038] h t =O t tanh+U t

[0039] Among them, h t Represents the hidden state output at the current moment;

[0040] S3: Construct improved MPDI and NDVI formulas and use surface factors to correct the soil moisture inversion value of the area; the specific steps of the improved NDVI and MPDI are:

[0041] S3.1: Introduce the soil adjustment coefficient L and calculate the improved remote sensing image index SAVI, which is expressed as:

[0042]

[0043] Among them, L is the coefficient obtained by analyzing soil type and texture through field soil sampling, NIR is the reflectance of the near-infrared band, and Red is the reflectance of the red light band;

[0044] S3.2: Introduce the slope S and the slope direction A, and calculate the incident angle cos(θ), the expression is:

[0045] cos(θ)=cos(μ)·cos(S)+sin(μ)·sin(S)·cos(A)

[0046] Among them, μ is the solar zenith angle, A is the slope direction, and S is the slope;

[0047] S3.3: According to the incident angle value cos(θ), calculate the improved remote sensing image index TSNDVI, the expression is:

[0048] TSNDVI=SAVI×cos(θ)

[0049] S3.4: Improve the soil moisture status MPDI1 based on the water retention and thermal properties of the soil, expressed as:

[0050]

[0051] Among them, γ represents the influence weight of soil texture, NIR is the reflectance of near infrared band, SWIR is the reflectance of short wave infrared band, and Tq is the soil texture coefficient;

[0052] S3.5: Improve the soil moisture status MPDI2 based on the scattering and reflection of light on the surface. The expression is:

[0053] MPDI2=MPDI1·(1+δ·Rou)

[0054] Among them, δ represents the influence weight of surface roughness, and Rou represents the degree of surface roughness;

[0055] S3.6: Since the structure of the vegetation canopy (such as canopy height and density) changes the reflective properties of water, the soil moisture status MPDI3 is improved, expressed as:

[0056] MPDI3=MPDI2·(1+C·Vs)

[0057] Among them, C represents the influence weight of the structure of the vegetation canopy, and Vs represents the canopy structure coefficient;

[0058] S4: Using the measured data of ground sensors to assist the remote sensing data humidity inversion, and obtain the soil moisture change in the target monitoring area; wherein, using the measured data of ground sensors to assist the remote sensing data humidity inversion specific steps are:

[0059] S4.1: Replace the humidity value of the pixel point corresponding to the measured point in the remote sensing image with the actual humidity value of the measured point;

[0060] S4.2: Construct a regression model based on the measured point humidity data and surface factors. The predicted value of the model output is M, which is expressed as:

[0061] M i =φ(P i ,E i )+τ i

[0062] Among them, P i is the humidity at the known measured point, E i is the surface factor matrix, τ i is the error term;

[0063] S4.3: Use the humidity of the nearest measured point for interpolation correction to obtain the corresponding interpolation matrix S, which is expressed as:

[0064]

[0065] Among them, S is the humidity value of the non-measured point, n is the number of measured points around the target point, S i is the humidity value corresponding to the measured point, is the calculation weight;

[0066] S4.4: The prediction results of the regression model are weighted and combined with the interpolation results to obtain the final corrected humidity value matrix L, which is expressed as:

[0067] L final =Δ·S+(1-Δ)·M

[0068] Among them, L final is the combined estimated humidity value, Δ is the weighting coefficient used to balance the regression model and the interpolation results, S is the humidity interpolation value at the current point, and M is the humidity prediction value at the current point;

[0069] S5: Construct an LSTM model to predict the humidity change in the target monitoring area at a future time point. The specific steps of using LSTM to predict the humidity change in the target monitoring area at a future time point are as follows:

[0070] S5.1: Input all the attributes of the current point and construct the future input data sequence K, which is expressed as:

[0071] K i ={I,C,E,Tem,Mos}

[0072] Among them, I is the original remote sensing image, C is the three-dimensional coordinate matrix of known points, E is the terrain information matrix, Tem is the meteorological data, and Mos is the current humidity data;

[0073] S5.2: Input the sequence into the LSTM model that has been trained in the second stage, and predict the humidity at each time step through the time recursion mechanism of the LSTM model;

[0074] S5.3: Combine the surface factors to make the prediction more consistent with the actual humidity value G in the area. The expression is:

[0075]

[0076] in, is the weight calculated according to the surface factor, N is the humidity feature obtained by the LSTM model, and E is the surface factor matrix;

[0077] S5.4: For each prediction step, the LSTM model updates the hidden state h t As the input of the next time step, combined with surface factors and meteorological data, the future humidity value Hv is recursively predicted. The prediction formula is:

[0078]

[0079] in, represents the predicted humidity value at time t+1, and ω(·) represents the prediction layer.

[0080] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0081] This high-precision soil moisture prediction method based on the spatiotemporal convolutional neural network coupling model avoids the limitations of a single data source by fusing measured soil moisture data at different depths of the surface with remote sensing data, and uses a deep learning model to improve the image recognition rate. At the same time, the surface factor is introduced to optimize and correct the soil inversion results of the remote sensing image of the target area, and the soil moisture changes in the target area are obtained, which improves the accuracy of soil moisture monitoring. The results show that the early warning and forecast accuracy is not less than 90%. The deep learning model is used to extract the temporal and spatial characteristics of the humidity data in the target area, and the surface humidity data in a certain time window in the future is predicted, so as to construct a spatiotemporal coupling neural network to restore the humidity changes in the target area.

[0082] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0084] Figure 1 This is a flow chart of a high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model proposed in the present invention.

[0085] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0086] The present invention will now be described in further detail with reference to the accompanying drawings.

[0087] See also Figure 1 As shown, in this embodiment, a high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model is provided, comprising the following steps:

[0088] S1: Deploy soil moisture detection equipment equipped with GPS function to the target area to obtain soil moisture measurement data at different depths;

[0089] S2: Obtain remote sensing satellite images of the target area, import the surface factors into the deep learning model for training, use the trained learning model to correct the remote sensing images and divide different types of terrain; the specific steps of constructing remote sensing image optimization and extracting features are:

[0090] S2.1: Define the three-dimensional coordinate matrix C = [xyz] of the humidity sensor in the x, y, and z directions, where C is the three-dimensional coordinate quantity, and xyz is the coordinate value of the sensor in the north-south, east-west, and vertical directions respectively;

[0091] S2.2: Define the surface factor matrix E=[SAV] around the humidity sensor deployment point, where E is the surface factor, and SAV is the slope, slope aspect, and vegetation coverage around the deployment point;

[0092] S2.3: In the input layer, the coordinate matrix of the known points and the terrain data are input into the convolutional neural network as an independent feature channel. Construct the remote sensing image optimization matrix;

[0093] I'=[I,C,E]

[0094] Among them, I' is the enhanced image matrix, I is the original remote sensing image, C is the three-dimensional coordinate matrix of the known points, and E is the terrain information matrix, etc.;

[0095] S2.4: Use CNN to extract the spatial features of the target remote sensing image. For the input image I', the feature map is obtained after the convolution operation of the CNN layer. The expression is:

[0096]

[0097] Among them, l represents the number of convolution layers, k represents the number of convolution kernels, is the weight of the kth convolution kernel, * represents the convolution operation, is the bias term, σ is the activation function;

[0098] S2.5: Combine the surface factor matrix E and the spatial feature matrix Fusion, fusion module X t The expression is:

[0099]

[0100] Among them, α t is the attention weight of the surface factor;

[0101] S2.6: Build an LSTM model and transform X t Input the LSTM model to extract the temporal features of the target remote sensing image. The LSTM core expression is:

[0102] Determine which information in the memory unit needs to be forgotten. The forget gate expression is:

[0103] f t =ρ(W f ·[h t-1 ,Q t ]+d f )

[0104] Among them, f t is the forgotten information, ρ is the activation function, W f is the weight matrix of the forget gate, h t-1 represents the hidden state at the previous moment, Q t represents the input at the current moment, d f is the bias term of the forget gate;

[0105] Determine what new information needs to be stored in the memory cell. The input gate expression is:

[0106] i t =ρ(W i ·[h t-1 ,Q t ]+d i )

[0107] Among them, W i is the weight matrix of the input gate, d i is the bias term of the input gate;

[0108] Candidate memory cell states, The expression is:

[0109]

[0110] Among them, W u is the weight matrix of the candidate state, d u is the bias term, and tanh is the activation function;

[0111] Update the memory unit state, U t The expression is:

[0112]

[0113] Among them, U t-1 Indicates the state of the memory unit at the previous moment;

[0114] Determine the hidden state O of the output t , the output gate expression is:

[0115] O t =ε*W O ·[h t-1 ,Q i ]+d O

[0116] Among them, ε represents the activation function of the output gate, W O is the weight matrix of the output gate, d O is the bias term of the output gate;

[0117] h t =O t tanh+U t

[0118] Among them, h t Represents the hidden state output at the current moment;

[0119] S3: Construct improved MPDI and NDVI formulas and use surface factors to correct the soil moisture inversion value of the area; the specific steps of the improved NDVI and MPDI are:

[0120] S3.1: Introduce the soil adjustment coefficient L and calculate the improved remote sensing image index SAVI, which is expressed as:

[0121]

[0122] Among them, L is the coefficient obtained by analyzing soil type and texture through field soil sampling, NIR is the reflectance of the near-infrared band, and Red is the reflectance of the red light band;

[0123] S3.2: Introduce the slope S and the slope direction A, and calculate the incident angle cos(θ), the expression is:

[0124] cos(θ)=cos(μ)·cos(S)+sin(μ)·sin(S)·cos(A)

[0125] Among them, μ is the solar zenith angle, A is the slope direction, and S is the slope;

[0126] S3.3: According to the incident angle value cos(θ), calculate the improved remote sensing image index TSNDVI, the expression is:

[0127] TSNDVI=SAVI×cos(θ)

[0128] S3.4: Improve the soil moisture status MPDI1 based on the water retention and thermal properties of the soil, expressed as:

[0129]

[0130] Among them, γ represents the influence weight of soil texture, NIR is the reflectance of near infrared band, SWIR is the reflectance of short wave infrared band, and Tq is the soil texture coefficient;

[0131] S3.5: Improve the soil moisture status MPDI2 based on the scattering and reflection of light on the surface. The expression is:

[0132] MPDI2=MPDI1·(1+δ·Rou)

[0133] Among them, δ represents the influence weight of surface roughness, and Rou represents the degree of surface roughness;

[0134] S3.6: Since the structure of the vegetation canopy (such as canopy height and density) changes the reflective properties of water, the soil moisture status MPDI3 is improved, expressed as:

[0135] MPDI3=MPDI2·(1+C·Vs)

[0136] Among them, C represents the influence weight of the structure of the vegetation canopy, and Vs represents the canopy structure coefficient;

[0137] S4: Using the measured data of ground sensors to assist the remote sensing data humidity inversion, and obtain the soil moisture change in the target monitoring area; wherein, using the measured data of ground sensors to assist the remote sensing data humidity inversion specific steps are:

[0138] S4.1: Replace the humidity value of the pixel point corresponding to the measured point in the remote sensing image with the actual humidity value of the measured point;

[0139] S4.2: Construct a regression model based on the measured point humidity data and surface factors. The predicted value of the model output is M, which is expressed as:

[0140] M i =φ(P i ,E i )+τ i

[0141] Among them, P i is the humidity at the known measured point, E i is the surface factor matrix, τ i is the error term;

[0142] S4.3: Use the humidity of the nearest measured point for interpolation correction to obtain the corresponding interpolation matrix S, which is expressed as:

[0143]

[0144] Among them, S is the humidity value of the non-measured point, n is the number of measured points around the target point, S i is the humidity value corresponding to the measured point, is the calculation weight;

[0145] S4.4: The prediction results of the regression model are weighted and combined with the interpolation results to obtain the final corrected humidity value matrix L, which is expressed as:

[0146] L final =Δ·S+(1-Δ)·M

[0147] Among them, L final is the combined estimated humidity value, Δ is the weighting coefficient used to balance the regression model and the interpolation results, S is the humidity interpolation value at the current point, and M is the humidity prediction value at the current point;

[0148] S5: Construct an LSTM model to predict the humidity change in the target monitoring area at a future time point. The specific steps of using LSTM to predict the humidity change in the target monitoring area at a future time point are as follows:

[0149] S5.1: Input all the attributes of the current point and construct the future input data sequence K, which is expressed as:

[0150] K i ={I,C,E,Tem,Mos}

[0151] Among them, I is the original remote sensing image, C is the three-dimensional coordinate matrix of known points, E is the terrain information matrix, Tem is the meteorological data, and Mos is the current humidity data;

[0152] S5.2: Input the sequence into the LSTM model that has been trained in the second stage, and predict the humidity at each time step through the time recursion mechanism of the LSTM model;

[0153] S5.3: Combine the surface factors to make the prediction more consistent with the actual humidity value G in the area. The expression is:

[0154]

[0155] in, is the weight calculated according to the surface factor, N is the humidity feature obtained by the LSTM model, and E is the surface factor matrix;

[0156] S5.4: For each prediction step, the LSTM model updates the hidden state h t As the input of the next time step, combined with surface factors and meteorological data, the future humidity value Hv is recursively predicted. The prediction formula is:

[0157]

[0158] in, represents the predicted humidity value at time t+1, and ω(·) represents the prediction layer.

[0159] This high-precision soil moisture prediction method based on the spatiotemporal convolutional neural network coupling model avoids the limitations of a single data source by fusing measured soil moisture data at different depths of the surface with remote sensing data, and uses a deep learning model to improve the image recognition rate. At the same time, the surface factor is introduced to optimize and correct the soil inversion results of the remote sensing image of the target area, and the soil moisture changes in the target area are obtained, which improves the accuracy of soil moisture monitoring. The results show that the early warning and forecast accuracy is not less than 90%. The deep learning model is used to extract the temporal and spatial characteristics of the humidity data in the target area, and the surface humidity data in a certain time window in the future is predicted, so as to construct a spatiotemporal coupling neural network to restore the humidity changes in the target area.

[0160] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model, characterized in that: The following steps are involved: S1: Deploy soil moisture detection equipment equipped with GPS function to the target area to obtain soil moisture measurement data at different depths; S2: Obtain remote sensing satellite images of the target area, import surface factors into the deep learning model for training, use the trained learning model to correct the remote sensing images and divide different types of terrain; S3: Construct improved MPDI and NDVI formulas and use surface factors to correct the soil moisture inversion value of the area; S4: Use the measured data from ground sensors to assist the remote sensing data inversion to obtain the soil moisture change in the target monitoring area; S5: Construct an LSTM model to predict the humidity change in the target monitoring area at future time points.

2. According to claim 1, a high-precision soil moisture prediction method based on a spatiotemporal convolutional neural network coupling model is characterized in that: The learning model in step S2 is used to construct remote sensing image optimization and extract features. The specific steps are: S2.1: Define the three-dimensional coordinate matrix C = [xyz] of the humidity sensor in the x, y, and z directions, where C is the three-dimensional coordinate quantity, and xyz is the coordinate value of the sensor in the north-south, east-west, and vertical directions respectively; S2.2: Define the surface factor matrix E=[SAV] around the humidity sensor deployment point, where E is the surface factor, and SAV are the slope, slope aspect and vegetation coverage around the deployment point respectively; S2.3: In the input layer, the coordinate matrix of the known points and the terrain data are input into the convolutional neural network as an independent feature channel, and the remote sensing image optimization matrix I' is constructed; S2.4: Use CNN to extract the spatial features of the target remote sensing image. For the input image I', the feature map is obtained after the convolution operation of the CNN layer. S2.5: Combine the surface factor matrix E and the spatial feature matrix Fusion is performed to generate fusion module X t ; S2.6: Build an LSTM model and transform X t Input the LSTM model to extract the temporal features of the target remote sensing image.

3. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 2 is characterized in that: The LSTM model consists of: Forget gate module: used to determine which information in the memory unit needs to be forgotten; Input gate module: used to determine what new information needs to be stored in the memory cell; Candidate memory cell state module; Update memory unit status module; Output gate module, used to determine the hidden state of the output; Hidden state output module at the current moment.

4. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 3 is characterized in that: The LSTM core expression is: Determine which information in the memory unit needs to be forgotten. The forget gate expression is: f t =ρ(W f ·[h t-1 ,Q t ]+d f ) Among them, f t is the forgotten information, ρ is the activation function, W f is the weight matrix of the forget gate, h t-1 represents the hidden state at the previous moment, Q t represents the input at the current moment, d f is the bias term of the forget gate; Determine what new information needs to be stored in the memory cell. The input gate expression is: i t =ρ(W i ·[h t-1 ,Q t ]+d i ) Among them, W i is the weight matrix of the input gate, d i is the bias term of the input gate; Candidate memory cell states, The expression is: Among them, W u is the weight matrix of the candidate state, d u is the bias term, and tanh is the activation function; Update the memory unit state, U t The expression is: Among them, U t-1 Indicates the state of the memory unit at the previous moment; Determine the hidden state O of the output t , the output gate expression is: O t =ε*W O ·[h t-1 ,Q i ]+d O Among them, ε represents the activation function of the output gate, W O is the weight matrix of the output gate, d O is the bias term of the output gate; h t =O t ·tanh+U t Among them, h t Represents the hidden state output at the current moment.

5. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 1 is characterized in that: In step S3, the specific steps of improving NDVI and MPDI are: S3.1: Introduce the soil adjustment coefficient L and calculate the improved remote sensing image index SAVI; S3.2: Introduce the slope S and the slope direction A, and calculate the incident angle value cos(θ); S3.3: Calculate the improved remote sensing image index TSNDVI according to the incident angle value cos(θ); S3.4: Improve soil moisture status MPDI1 based on soil moisture retention and thermal properties; S3.5: Improve soil moisture status MPDI2 based on the scattering and reflection of light on the surface; S3.6: Improve soil moisture status MPDI3 because the structure of the vegetation canopy (such as canopy height and density) changes the reflective properties of water.

6. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 5 is characterized in that: The improved remote sensing image index SAVI is expressed as: Among them, L is the coefficient obtained by analyzing soil type and texture through field soil sampling, NIR is the reflectance of the near-infrared band, and Red is the reflectance of the red light band; The incident angle value cos(θ) is expressed as: cos(θ)=cos(μ)·cos(S)+sin(μ)·sin(S)·cos(A) Among them, μ is the solar zenith angle, A is the slope direction, and S is the slope; The improved remote sensing image index TSNDVI is expressed as: TSNDVI=SAVI×cos(θ) Improve the soil moisture status MPDI1, the expression is: Among them, γ represents the influence weight of soil texture, NIR is the reflectance of near infrared band, SWIR is the reflectance of short wave infrared band, and Tq is the soil texture coefficient; Improve the soil moisture status MPDI2, the expression is: MPDI2=MPDI1·(1+δ·Rou) Among them, δ represents the influence weight of surface roughness, and Rou represents the degree of surface roughness; Improve the soil moisture status MPDI3, the expression is: MPDI3=MPDI2·(1+C·Vs) Among them, C represents the influence weight of the structure of the vegetation canopy, and Vs represents the canopy structure coefficient.

7. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 1 is characterized in that: In step S4, the specific steps of using the measured data of the ground sensor to assist the humidity inversion of the remote sensing data are as follows: S4.1: Replace the humidity value of the pixel point corresponding to the measured point in the remote sensing image with the actual humidity value of the measured point; S4.2: Construct a regression model based on the measured point humidity data and surface factors, and the predicted value of the model output is M; S4.3: Use the humidity of the nearest measured point to perform interpolation correction and obtain the corresponding interpolation matrix S; S4.4: The prediction results of the regression model are weightedly combined with the interpolation results to obtain the final corrected humidity value matrix L.

8. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 7 is characterized in that: The predicted value of the model output is M, which is expressed as: M i =φ(P i ,E i )+τ i Among them, P i is the humidity at the known measured point, E i is the surface factor matrix, τ i is the error term; The interpolation matrix S is expressed as: Among them, S is the humidity value of the non-measured point, n is the number of measured points around the target point, S i is the humidity value corresponding to the measured point, is the calculation weight; The final corrected humidity value matrix L is expressed as: L final =Δ·S+(1-Δ)·M Among them, L final is the combined estimated humidity value, Δ is the weighting coefficient used to balance the regression model and the interpolation results, S is the humidity interpolation value of the current point, and M is the humidity prediction value of the current point.

9. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 1 is characterized in that: In step S5, the specific steps of using LSTM to predict the humidity change amount of the target monitoring area at a future time point are: S5.1: Input all attributes of the current point and construct the future input data sequence K; S5.2: Input the sequence into the LSTM model that has been trained in the second stage, and predict the humidity at each time step through the time recursion mechanism of the LSTM model; S5.3: Combine surface factors to make the prediction more consistent with the actual humidity value G in the area; S5.4: For each prediction step, the LSTM model updates the hidden state h t As the input of the next time step, and combined with surface factors and meteorological data, the future humidity value Hv is recursively predicted.

10. The method for high-precision soil moisture prediction based on a spatiotemporal convolutional neural network coupling model according to claim 9, characterized in that: The future input data sequence K is expressed as: K i ={I,C,E,Have,Mos} Among them, I is the original remote sensing image, C is the three-dimensional coordinate matrix of known points, E is the terrain information matrix, Tem is the meteorological data, and Mos is the current humidity data; The actual humidity value G is expressed as: G true =θ·N+(1-θ)·E Among them, θ is the weight calculated according to the surface factor, N is the humidity feature obtained by the LSTM model, and E is the surface factor matrix; Predict the future humidity value Hv, the prediction formula is: in, represents the predicted humidity value at time t+1, and ω(·) represents the prediction layer.

Images