Remote sensing rainfall inversion method and system based on multiple environmental factors and physical constraint convolution long short-term memory network

Abstract

The invention discloses a remote sensing rainfall inversion method and a remote sensing rainfall inversion system based on multiple environmental factors and a physical constraint convolution long short-term memory network. Secondly, static factor dynamic features are extracted for dynamic factors including remote sensing rainfall products and meteorological factors and static factors including terrain factors and earth surface feature factors through a Gaussian kernel function; then constructing a remote sensing rainfall inversion network based on a convolutional long-short-term memory network, and introducing a water vapor balance constraint and a surface energy conservation constraint at the same time; static factor dynamic features are used as input, corresponding ground observation rainfall data are used as output to train the network, a remote sensing rainfall inversion model is obtained, and during training, a loss function composed of water vapor balance constraint, surface energy conservation constraint and mean square error is introduced; and finally, a high-precision remote sensing rainfall inversion result is obtained by using the sensing rainfall inversion model. The method can effectively improve the temporal-spatial resolution and inversion precision of remote sensing rainfall data.

Description

Technical Field

[0001] The present invention belongs to the field of high-resolution precipitation spatial distribution prediction, and specifically relates to a remote sensing precipitation inversion method and system based on multiple environmental factors and physically constrained convolutional long short-term memory networks. Background Art

[0002] Precipitation is a key variable in ecohydrology, agricultural production, and climate change research. Its spatial distribution and dynamics directly influence regional water resource distribution, surface runoff, soil moisture, and evaporation. However, traditional precipitation prediction methods (such as inverse distance weighting and kriging) are limited in modeling complex terrain and heterogeneous environmental factors, making them inadequate for high-resolution precipitation prediction.

[0003] In recent years, deep learning techniques, particularly convolutional long short-term memory (ConvLSTM) networks, have demonstrated remarkable performance in spatiotemporal series forecasting. However, existing deep learning models are typically "black box" models that lack physical constraints, making it difficult to ensure that predictions conform to actual physical laws. Shandong Province, a major agricultural production region in China, has a significant impact on agricultural production and water resource management due to its spatiotemporal distribution. However, traditional precipitation interpolation methods struggle to effectively capture the high-resolution distribution characteristics of regional precipitation. Summary of the Invention

[0004] The present invention addresses the problem in the prior art that traditional precipitation interpolation methods at ground observation stations cannot effectively capture the impact of complex terrain and dynamic environmental factors on precipitation distribution, i.e., the accuracy and spatiotemporal resolution of remote sensing products are insufficient. The present invention provides a remote sensing precipitation inversion method and system based on multiple environmental factors and physically constrained convolutional long short-term memory networks. This method can achieve high-precision remote sensing precipitation inversion under limited observation data conditions, providing a scientific basis for runoff simulation, regional water resources management, and climate change research.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a remote sensing precipitation inversion method based on multiple environmental factors and a physically constrained convolutional long short-term memory network, comprising the following steps:

[0006] S1. Obtain remote sensing precipitation data and corresponding ground observation precipitation data in the target area, perform multi-scale spatiotemporal fusion and physical consistency correction on the remote sensing precipitation data, and generate dynamic factors;

[0007] S2. Using an adaptive kernel function to extract dynamic features of static factors from dynamic factors and static factors; the dynamic factors include remote sensing precipitation products and meteorological factors; the static factors include terrain factors and surface characteristic factors;

[0008] S3. Construct and train a remote sensing precipitation inversion network using the dynamic characteristics of static factors as input and corresponding ground-based precipitation observation data as output to obtain a remote sensing precipitation inversion model. When constructing the remote sensing precipitation inversion network, physical constraints are introduced, including water vapor balance constraints and surface energy conservation constraints. When training the remote sensing precipitation inversion network, a loss function consisting of water vapor balance principle constraints, surface energy conservation constraints, and mean square error is referenced.

[0009] S4, execute steps S1 to S2 for the remote sensing precipitation data of the area to be predicted and the corresponding ground observation precipitation data, and then

[0010] The remote sensing precipitation inversion model will be used to obtain high-precision remote sensing precipitation inversion results.

[0011] Furthermore, the aforementioned step S1 includes the following sub-steps:

[0012] S1.1. Obtain ground rain gauge data, various remote sensing precipitation data, wind speed, relative humidity, evaporation, surface temperature, digital elevation model (DEM), coastline vector data, surface roughness, and soil moisture data within the preset area;

[0013] S1.2. Preprocess the remote sensing precipitation data, including calculating slope and aspect based on the digital elevation model (DEM), calculating the shortest distance from the grid center to the nearest coastline based on the DEM and coastline vector data, unifying the temporal and spatial resolution of all data, terrain correction, cloud cover processing, and surface feature change correction;

[0014] S1.3. Using the characteristics of typical precipitation clouds in the selected area, combined with wind speed and cloud height, calculate the precipitation range and estimate the number of remote sensing grids as follows:

[0015]

[0016] Where H is the average height of precipitation clouds; V is the wind speed; R is the precipitation impact range, and the number of grids n is determined according to R.

[0017] Furthermore, in the aforementioned step S2, the adaptive kernel function adopts a Gaussian kernel function, as shown in the following formula:

[0018]

[0019] Where K(x,y) is the Gaussian kernel function; x and y are the dynamic factor and static factor respectively; σ is the bandwidth parameter of the kernel function, which controls the smoothness of the Gaussian distribution;

[0020] Furthermore, in the aforementioned step S3, the water vapor balance constraint is as follows:

[0021] P=RH·(ET+Δq)

[0022] Where P is precipitation; RH is relative humidity; ET is evaporation; and Δq is the change in water vapor flux.

[0023] The surface energy conservation constraint is as follows:

[0024]

[0025] Where, e s (T) is the saturated water vapor pressure; T is the temperature; T0 is the reference temperature; L is the latent heat of water vapor; R v is the gas constant of water vapor.

[0026] Furthermore, in the aforementioned step S3, the loss function is as follows:

[0027]

[0028] Where, Represents the loss function, dynamically adjusting the weights λ1 and λ2 by changing the constraint error ratio. represents the water vapor balance constraint, represents the physical constraint based on surface energy conservation, is the data-driven mean squared error loss.

[0029] Furthermore, the aforementioned water vapor balance constraint is as follows:

[0030]

[0031] Where, E t,i is the evaporation amount of sample i at time t; Δq t.i is the water vapor flux change of sample i at time t; RH t,i is the relative humidity of sample i at time t;

[0032] The physical constraint based on surface energy conservation is as follows:

[0033]

[0034] Where C is a constant, ΔH is the change in water vapor flux, ΔP is the change in precipitation; L is the latent heat of water vapor; R v is the water vapor gas constant; T0 is the reference temperature; T t,i is the surface temperature of sample i at time t; RH t,i is the relative humidity of sample i at time t; P t,i is the assimilated value of the remote sensing precipitation model for the i-th sample at time t.

[0035] Furthermore, the aforementioned remote sensing precipitation inversion network is constructed based on a convolutional long short-term memory neural network. Training the remote sensing precipitation inversion network includes the following sub-steps:

[0036] S3.1. Preset the time series length T and determine the shape of the input tensor through partial autocorrelation analysis of several precipitation series: (T, H, W, C) = (T, n, n, x); where n is the number of remote sensing grids corresponding to a site in step 1; x is the number of dynamic factors after processing by the adaptive function;

[0037] S3.2. Perform a two-dimensional convolution operation on the hidden state of the previous time step and the current input data, and calculate it as follows:

[0038]

[0039] Where, O (i，j) Represents the pixel value of the coordinate position in the output image O; x and y are the horizontal and vertical coordinates in the input image respectively; ix and jy are the positions of the convolution kernel on the input image respectively; h t-1 represents the output of the model at the previous moment, i.e., time t-1; x t is the current input data;

[0040] S3.3, update the current unit state through the forget gate, input gate and output gate, selectively remember or forget historical information, specifically calculated by the following method: the forget gate matrix f at time t t The calculation formula is:

[0041] f t =sigmod(W f O+b f )

[0042] Where O is the matrix after two-dimensional convolution; W f is the parameter matrix; b f For bias top;

[0043] The input gate of the model is calculated by the following formula:

[0044] i t =sigmod(W i O+b i )

[0045]

[0046] Where i t represents the input gate matrix; W i Represents the parameter matrix of this step; b i is the bias term; Indicates the current memory; W CThe weight matrix representing the input information; b C is the bias term; C t-1 Represents the information at time t-1; C t Represents the current state; the output of each time step is determined by the current unit state and the output gate:

[0047] The new cell state is passed through the tanh network layer and multiplied by the part to be output through the output gate to obtain the output result. The calculation formula is as follows:

[0048] o t =sigmod(W o ·[P t-1 ,x t ]+b o )

[0049] P t =o t *tanh(C t )

[0050] In the formula, o t represents the output gate matrix; W o Represents the parameter matrix of this step; b o is the bias term; P t Represents the output at time t, that is, the precipitation error value at that moment.

[0051] Another aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the present invention when executing the computer program.

[0052] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any one of the methods described in the present invention when executed by a processor.

[0053] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0054] 1. Fusion optimization of dynamic and static factors

[0055] By using a Gaussian kernel function to differentially integrate dynamic factors with static factors (such as DEM terrain data), this method effectively overcomes the model training bias caused by the long-term invariance of static factors in traditional methods. This mechanism ensures that static geographic information remains effectively involved in the model learning process, preventing it from being misjudged as irrelevant and ignored, significantly enhancing the accuracy of modeling the impact of terrain elements on precipitation processes.

[0056] 2. Coupling of multi-dimensional environmental factors

[0057] The innovative integration of underlying surface parameters (such as land cover type and vegetation index) alongside meteorological factors into the model inputs breaks through the limitations of traditional precipitation inversion, which relies solely on precipitation products. By integrating the spatial heterogeneity of multiple environmental factors, the model accurately captures the modulation of topographic relief and surface properties on the spatial distribution of precipitation.

[0058] 3. Physics-enhanced model architecture

[0059] By embedding hydrological and meteorological physical constraint equations within a neural network, a hybrid model is constructed that combines data learning capabilities with physical regularity. Compared to purely data-driven models, this architecture not only improves generalization for extreme precipitation events but also accelerates model convergence by approximately 2 times. Furthermore, it enhances model interpretability by visualizing intermediate physical fields.

[0060] 4. Advantages of Spatiotemporal Collaborative Modeling: This method uses a convolutional long short-term memory (ConvLSTM) network to simultaneously capture the spatiotemporal evolution of precipitation processes. Compared to traditional machine learning methods and kriging interpolation techniques, it can achieve refined inversion of hourly precipitation fields at a spatial resolution of 1 km. This method is particularly adept at capturing precipitation variations caused by spatial heterogeneity factors such as urban heat islands and topographic uplift, with a spatial correlation coefficient of over 0.85 between the inversion results and measured data. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the remote sensing precipitation inversion process of the present invention.

[0062] Figure 2 It is a schematic diagram of the unit structure of the convolutional long short-term memory neural network of the present invention. DETAILED DESCRIPTION

[0063] In order to better understand the technical content of the present invention, specific embodiments are given and described below with reference to the accompanying drawings.

[0064] Various aspects of the present invention are described herein with reference to the accompanying drawings, which show a number of illustrative embodiments. The embodiments of the present invention are not limited to those described in the accompanying drawings. It should be understood that the present invention can be implemented by any of the various concepts and embodiments described above, as well as the concepts and implementations described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. In addition, some aspects disclosed herein may be used alone or in any appropriate combination with other aspects disclosed herein.

[0065] This example uses the Daqing River system in the Haihe River Basin (plain terrain), the Yihe River Basin in Shandong Province (hilly terrain), and the Tunxi River Basin in Zhejiang Province (mountainous terrain) as test areas to verify the applicability of the present invention under different terrain and meteorological conditions. The selected areas have significant spatial heterogeneity in precipitation and complex terrain characteristics, providing comprehensive verification of the model's performance.

[0066] refer to Figure 1 The present invention provides a remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network, comprising the following steps:

[0067] S1. Obtain remote sensing precipitation data and corresponding ground observation precipitation data in the target area, perform multi-scale spatiotemporal fusion and physical consistency correction, and generate dynamic factors;

[0068] S2. Using an adaptive kernel function to extract dynamic features of static factors from dynamic factors and static factors; the dynamic factors include remote sensing precipitation products and meteorological factors; the static factors include terrain factors and surface characteristic factors;

[0069] S3. Build and train a remote sensing precipitation inversion network using the dynamic characteristics of static factors as input and corresponding ground-based precipitation observation data as output to obtain a remote sensing precipitation inversion model. When building the remote sensing precipitation inversion network, introduce physical constraints, including water vapor balance constraints and surface energy conservation constraints. When training the remote sensing precipitation inversion network, reference a loss function consisting of the water vapor balance principle constraints, the surface energy conservation constraints, and the mean square error.

[0070] S4. Execute steps S1 to S2 for the remote sensing precipitation data of the area to be predicted and the corresponding ground observation precipitation data, and then use the remote sensing precipitation inversion model to obtain a high-precision remote sensing precipitation inversion result.

[0071] As a preferred embodiment of the present invention, in step S1, remote sensing precipitation data and corresponding ground observation precipitation data are obtained for the Daqing River system in the Haihe River Basin, the Yihe River Basin in Shandong Province, and the Tunxi River Basin in Zhejiang Province. The remote sensing precipitation data is obtained through satellite remote sensing equipment (GSMaP, PERSIANN, GPM IMERG Early Run, CMORPH, Sentinel-1, GPM IMERGFinal Run, and ERA5), and the ground observation precipitation data is obtained through ground observation stations in the basin. The remote sensing precipitation data is preprocessed to filter out hourly precipitation data for days with precipitation in the preset area. Specifically, step S1 includes the following sub-steps:

[0072] S1.1. Obtain ground rain gauge data, various remote sensing precipitation data, wind speed, relative humidity, evaporation, surface temperature, digital elevation model (DEM), coastline vector data, surface roughness, and soil moisture data within the preset area; remote sensing precipitation data can be obtained through remote sensing equipment such as satellites and radars, while ground observation precipitation data can be obtained through ground observation stations.

[0073] S1.2. Preprocess the remote sensing precipitation data, including calculating slope and aspect based on the digital elevation model (DEM), calculating the shortest distance from the grid center to the nearest coastline based on the DEM and coastline vector data, unifying the temporal and spatial resolution of all data, terrain correction, cloud cover processing, and surface feature change correction;

[0074] S1.3. Using the characteristics of typical precipitation clouds in the selected area, combined with wind speed and cloud height, calculate the precipitation range and estimate the number of remote sensing grids as follows:

[0075]

[0076] Where H is the average height of precipitation clouds; V is the wind speed; R is the precipitation impact range, and the number of grids n is determined according to R.

[0077] As a preferred embodiment of the present invention, step S2 uses an adaptive kernel function to extract the interaction relationship between the dynamic factor and the static factor, and calculates the following formula:

[0078]

[0079] Where K(x,y) is the Gaussian kernel function; x and y are the dynamic factor and static factor respectively; σ is the bandwidth parameter of the kernel function, which controls the smoothness of the Gaussian distribution;

[0080] As a preferred embodiment of the present invention, during the model construction process in step S3, two physical constraints, water vapor balance constraint and surface energy conservation, are introduced. At the same time, the model is operated using processed dynamic input data and ground-based precipitation data.

[0081] The water vapor balance constraint is as follows:

[0082] P=RH·(ET+Δq)

[0083] Where P is precipitation; RH is relative humidity; ET is evaporation; and Δq is the change in water vapor flux.

[0084] The surface energy conservation constraint is as follows:

[0085]

[0086] Where, e s(T) is the saturated water vapor pressure; T is the temperature; T0 is the reference temperature (273.15K); L is the latent heat of water vapor (about 2.5×10 6 J / kg); R v is the gas constant of water vapor (461 J / kg·K).

[0087] As a preferred embodiment of the present invention, in step S3, the loss function is as follows:

[0088]

[0089] Where, Represents the loss function, dynamically adjusting the weights λ1 and λ2 by changing the constraint error ratio. represents the water vapor balance constraint, represents the physical constraint based on surface energy conservation, is the data-driven mean squared error loss.

[0090] The water vapor balance constraint is as follows:

[0091]

[0092] Where, E t,i is the evaporation of sample i at time t (obtained from ERA5 data); Δq t.i is the change in water vapor flux of sample i at time t (calculated from ERA5 wind speed and humidity gradient); RH t,i is the relative humidity of sample i at time t;

[0093] The physical constraint based on surface energy conservation is as follows:

[0094]

[0095] Where C is a constant, which is directly related to the water vapor flux and precipitation conversion efficiency per unit area in the region. ΔH is the change in water vapor flux, which can be obtained from ERA5; ΔP is the change in precipitation; L is the latent heat of water vapor (2.26×106 J / kg); R v is the water vapor gas constant (461 J / (kg·K)); T0 is the reference temperature, usually 273.15 K; T t,i is the surface temperature of sample i at time t, which can be obtained from MODIS LST data; RH t,i is the relative humidity of sample i at time t, obtained from ERA5 data; P t,i is the assimilated value of the remote sensing precipitation model for the i-th sample at time t.

[0096] As a preferred embodiment of the present invention, refer to Figure 2,The remote sensing precipitation inversion network is constructed based on the ,convolutional long short term memory neural network. The convolutional long short term memory neural network model ,calculates through the following steps;

[0097] S3.1. Preset the time series length T = 4. Through partial autocorrelation analysis of several precipitation series, determine the shape of the input tensor to be: (T, H, W, C) = (4, n, n, x); where n is the number of remote sensing grids corresponding to a site in step S1; and x is the number of dynamic factors after processing by the adaptive function.

[0098] S3.2. Perform a two-dimensional convolution operation on the hidden state of the previous time step and the current input data, and calculate it as follows:

[0099]

[0100] Where, O (i，j) Represents the pixel value of the coordinate position in the output image O; x and y are the horizontal and vertical coordinates in the input image respectively; ix and jy are the positions of the convolution kernel on the input image respectively; h t-1 represents the output of the model at the previous moment, i.e., time t-1; x t is the current input data;

[0101] S3.3, update the current unit state through the forget gate, input gate and output gate, selectively remember or forget historical information, specifically calculated by the following method: the forget gate of the model forgets information that is likely to lead to wrong predictions, and the forget gate matrix f at time t t The calculation formula is:

[0102] f t =sigmod(W f O+b f )

[0103] Where O is the matrix after two-dimensional convolution; W f is the parameter matrix; b f For bias top;

[0104] The input gate of the model is calculated by the following formula:

[0105] i t =sigmod(W i O+b i )

[0106]

[0107]

[0108] Where i t represents the input gate matrix; W iRepresents the parameter matrix of this step; b i is the bias term; Indicates the current memory; W C The weight matrix representing the input information; b C is the bias term; C t-1 Represents the information at time t-1; C t Represents the current state; the output of each time step is determined by the current unit state and the output gate:

[0109] The new cell state is passed through the tanh network layer and multiplied by the part to be output through the output gate to obtain the output result. The calculation formula is as follows:

[0110] o t =sigmod(W o ·[P t-1 ,x t ]+b o )

[0111] P t =o t *tanh(C t )

[0112] In the formula, o t represents the output gate matrix; W o Represents the parameter matrix of this step; b o is the bias term; P t Represents the output at time t, that is, the precipitation error value at that moment.

[0113] In this embodiment, the convolutional long short-term memory neural network model learning rate is set to 0.001, the maximum number of iterations is 2000 times, and the model saves the best weights by the early stopping method. The early stopping period is 200, ensuring that the model avoids overfitting during training and enhancing the generalization ability of the model. The present invention was tested in the Daqing River system in the Haihe River Basin, the Yi River Basin in Shandong Province, and the Tunxi River Basin in Zhejiang Province. The average root mean square error (RMSE) was reduced by 50.6% compared with the average value of the four kinds of original satellite precipitation data, and the average mean absolute error (MAE) was reduced by 31.2% compared with the average value of the four kinds of original satellite precipitation data. A total of 32 floods in the three basins within 20 years were simulated using the distributed Xin'anjiang model. Compared with the use of the four kinds of original satellite precipitation average values to input the hydrological model, the Nash efficiency coefficient (NSE) was increased by an average of 0.231, especially in the Daqing River system in the Haihe River Basin, the performance was significantly improved, and the Nash efficiency coefficient (NSE) was increased from 0.484 to 0.763.

[0114] Specifically, in this embodiment, the convolutional long short-term memory neural network model learning rate is set to 0.001, the maximum number of iterations is 2000 times, and the model saves the best weights by the early stopping method. The early stopping period is 200, ensuring that the model avoids overfitting during training and enhancing the generalization ability of the model. The present invention was tested in the Daqing River system in the Haihe River Basin, the Yi River Basin in Shandong Province, and the Tunxi River Basin in Zhejiang Province. The average root mean square error (RMSE) was reduced by 50.6% compared with the average value of the four kinds of original satellite precipitation data, and the average mean absolute error (MAE) was reduced by 31.2% compared with the average value of the four kinds of original satellite precipitation data. A total of 32 floods in the three basins within 20 years were simulated using the distributed Xin'anjiang model. Compared with the use of the four kinds of original satellite precipitation average values to input the hydrological model, the Nash efficiency coefficient (NSE) was increased by an average of 0.231, especially in the Daqing River system in the Haihe River Basin, the performance was significantly improved, and the Nash efficiency coefficient (NSE) was increased from 0.484 to 0.763.

[0115] Another aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the present invention when executing the computer program.

[0116] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any one of the methods described in the present invention when executed by a processor.

[0117] The present invention proposes a remote sensing precipitation inversion method and system based on a convolutional long short-term memory network with multiple environmental factors and physical constraints. By introducing surface energy conservation constraints and water vapor balance constraints, the model is made to conform to actual physical laws. At the same time, the prediction accuracy is improved by utilizing multi-source remote sensing data and environmental factors, effectively improving the inversion accuracy and spatiotemporal resolution of remote sensing precipitation data. This method is particularly suitable for precipitation monitoring and prediction in complex terrain areas such as the Haihe River Basin and the Yihe River Basin in Shandong Province, providing strong data support for agricultural production, regional hydrological management, and climate change research. The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention, and these changes and improvements fall within the scope of the invention claimed for protection. The scope of protection claimed for the present invention is defined by the appended claims and their equivalents.

Claims

1. A remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network, characterized by: The following steps are involved: S1. Obtain remote sensing precipitation data and corresponding ground observation precipitation data in the target area, perform multi-scale spatiotemporal fusion and physical consistency correction on the remote sensing precipitation data, and generate dynamic factors; S2. Using an adaptive kernel function to extract dynamic features of static factors from dynamic factors and static factors; the dynamic factors include remote sensing precipitation products and meteorological factors; the static factors include terrain factors and surface characteristic factors; S3. Construct and train a remote sensing precipitation inversion network using the dynamic characteristics of static factors as input and corresponding ground-based precipitation observation data as output to obtain a remote sensing precipitation inversion model. When constructing the remote sensing precipitation inversion network, physical constraints are introduced, including water vapor balance constraints and surface energy conservation constraints. When training the remote sensing precipitation inversion network, a loss function consisting of water vapor balance principle constraints, surface energy conservation constraints, and mean square error is referenced. S4. Execute steps S1 to S2 for the remote sensing precipitation data of the area to be predicted and the corresponding ground observation precipitation data, and then use the remote sensing precipitation inversion model to obtain a high-precision remote sensing precipitation inversion result.

2. The remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network according to claim 1 is characterized in that: Step S1 includes the following sub-steps: S1.

1. Obtain ground rain gauge data, various remote sensing precipitation data, wind speed, relative humidity, evaporation, surface temperature, digital elevation model (DEM), coastline vector data, surface roughness, and soil moisture data within the preset area; S1.

2. Preprocess the remote sensing precipitation data, including calculating slope and aspect based on the digital elevation model (DEM), calculating the shortest distance from the grid center to the nearest coastline based on the DEM and coastline vector data, unifying the temporal and spatial resolution of all data, terrain correction, cloud cover processing, and surface feature change correction; S1.

3. Using the characteristics of typical precipitation clouds in the selected area, combined with wind speed and cloud height, calculate the precipitation range and estimate the number of remote sensing grids as follows: Where H is the average height of precipitation clouds; V is the wind speed; R is the precipitation impact range, and the number of grids n is determined according to R.

3. The remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network according to claim 1 is characterized in that: In step S2, the adaptive kernel function adopts the Gaussian kernel function, which is as follows: Where K(x,y) is the Gaussian kernel function; x and y are the dynamic factor and static factor respectively; σ is the bandwidth parameter of the kernel function, which controls the smoothness of the Gaussian distribution.

4. According to the remote sensing precipitation inversion method based on multiple environmental factors and physical constraints convolutional long short-term memory network according to claim 1, in step S3, the water vapor balance constraint is as follows: P=RH·(ET+Δq) Where P is precipitation; RH is relative humidity; ET is evaporation; and Δq is the change in water vapor flux. The surface energy conservation constraint is as follows: Where, e s (T) is the saturated water vapor pressure; T is the temperature; T0 is the reference temperature; L is the latent heat of water vapor; R v is the gas constant of water vapor.

5. The remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network according to claim 1 is characterized in that: In step S3, the loss function is as follows: Where, Represents the loss function, dynamically adjusting the weights λ1 and λ2 by changing the constraint error ratio. represents the water vapor balance constraint, represents the physical constraint based on surface energy conservation, is the data-driven mean squared error loss.

6. The remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network according to claim 5 is characterized in that: The water vapor balance constraint is as follows: Where, E t,i is the evaporation amount of sample i at time t; Δq t.i is the water vapor flux change of sample i at time t; RH t,i is the relative humidity of sample i at time t; The physical constraint term based on surface energy conservation is as follows: Where C is a constant, ΔH is the change in water vapor flux, ΔP is the change in precipitation; L is the latent heat of water vapor; R v is the water vapor gas constant; T0 is the reference temperature; T t,i is the surface temperature of sample i at time t; RH t,i is the relative humidity of sample i at time t; P t,i is the assimilated value of the remote sensing precipitation model for the i-th sample at time t.

7. The remote sensing precipitation inversion method based on multiple environmental factors and physical constraint convolutional long short-term memory network according to claim 1 is characterized in that: The remote sensing precipitation inversion network is built based on a convolutional long short-term memory neural network. Training the remote sensing precipitation inversion network includes the following sub-steps: S3.

1. Preset the time series length T and determine the shape of the input tensor through partial autocorrelation analysis of several precipitation series: (T, H, W, C) = (T, n, n, x); where n is the number of remote sensing grids corresponding to a site in step 1; x is the number of dynamic factors after processing by the adaptive function; S3.

2. Perform a two-dimensional convolution operation on the hidden state of the previous time step and the current input data, and calculate it as follows: Where, O (i，j) Represents the pixel value of the coordinate position in the output image O; x and y are the horizontal and vertical coordinates in the input image respectively; ix and jy are the positions of the convolution kernel on the input image respectively; h t-1 represents the output of the model at the previous moment, i.e., time t-1; x t is the current input data; S3.3, update the current unit state through the forget gate, input gate and output gate, selectively remember or forget historical information, specifically calculated by the following method: the forget gate matrix f at time t t The calculation formula is: f t =sigmod(W f ·O+b f ) Where O is the matrix after two-dimensional convolution; W f is the parameter matrix; b f For bias top; The input gate of the model is calculated by the following formula: i t =sigmod(W i ·O+b i ) Where i t represents the input gate matrix; W i Represents the parameter matrix of this step; b i is the bias term; Indicates the current memory; W C The weight matrix representing the input information; b C is the bias term; C t-1 Represents the information at time t-1; C t Indicates the current state; The output of each time step is determined by the current cell state and the output gate: The new cell state is passed through the tanh network layer and multiplied by the part to be output through the output gate to obtain the output result. The calculation formula is as follows: o t =sigmod(W o ·[P t-1 ,x t ]+b o ) P t =o t *tanh(C t ) In the formula, o t represents the output gate matrix; W o Represents the parameter matrix of this step; b o is the bias term; P t Represents the output at time t, that is, the precipitation error value at that moment.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images