A geographic information driven meteorological monitoring method and system

Abstract

The present application provides a kind of meteorological monitoring method and system based on geographic information driving.Collect the time series meteorological data of monitoring area and the multi-source geographic grid data such as elevation, slope, underlying surface type, generate static geographic feature vector by grid coding, and build the model containing parameter generation module and main prediction network.Parameter generation module fuses static geographic features and local meteorological state at the last time, outputs the shape control parameters used to regulate composite activation function, including periodic basis function amplitude, frequency and non-periodic basis function scaling factor.Main network extracts grid meteorological features by combining spatial convolution and time series processing unit, and carries out adaptive activation regulation on hidden state output link based on keeping long short-term memory network gate mechanism unchanged.In the training stage, the gradient is enhanced according to the grid function complexity, and the time smoothing regular term is introduced to constrain the parameter mutation;In the inference stage, the refined meteorological monitoring result data is output.

Description

Technical Field

[0001] This application belongs to the field of meteorological monitoring, and in particular relates to a meteorological monitoring method and system based on geographic information. Background Technology

[0002] Meteorological monitoring is a crucial foundation for disaster prevention and mitigation, agricultural production, and sustainable socio-economic development. Its core lies in the analysis and prediction of massive amounts of spatiotemporal meteorological data. With advancements in observation technology, meteorological monitoring has gradually evolved into a problem of spatiotemporal series processing. In the natural environment, the evolution of meteorological elements depends not only on their own historical changes but also on geographical factors such as topography, elevation, slope, and underlying surface type, resulting in significant spatial heterogeneity in local microclimates and meteorological conditions. However, existing meteorological monitoring models often focus on the temporal correlation of meteorological sequences or simple spatial distance topology, making it difficult to effectively integrate multi-source geospatial information with meteorological evolution patterns. This leads to insufficient identification of spatiotemporal differences caused by topography and surface attributes, limiting monitoring accuracy in complex environments.

[0003] Furthermore, meteorological systems are highly nonlinear and time-nonstationary, with local meteorological conditions exhibiting both long-term trend changes and periodic fluctuations driven by natural laws. Existing deep learning-based meteorological monitoring methods typically employ a uniform, fixed network architecture and static nonlinear computational mechanisms to process data from different regions. This makes it difficult to flexibly adjust the model response based on the static geographical features and real-time meteorological conditions of specific spatial locations, limiting the ability to express meteorological patterns across different regions and times. In training with high-dimensional, highly heterogeneous spatiotemporal data, the lack of constraints on the complexity of the model's internal state and smooth control over its continuous temporal evolution can easily lead to gradient backpropagation instability and abrupt changes in monitoring results. Therefore, there is an urgent need for a meteorological monitoring method driven by multi-source geographic information that can adjust the model's internal processing mechanisms according to specific spatiotemporal environments to improve monitoring accuracy and reliability. Summary of the Invention

[0004] According to one aspect of the present invention, a geographic information-driven meteorological monitoring method is proposed to improve the accuracy and reliability of monitoring, comprising the following steps: Acquire time-series meteorological data and multi-source geographic information data including elevation, slope and underlying surface type for the area to be monitored. Extract multi-source geographic information data and encode it for each geographic location raster, generate a static geographic feature vector corresponding to the raster, and construct a model including a parameter generation module and a backbone prediction network. The parameter generation module receives static geographic feature vectors and local meteorological state features from the previous time step, and outputs the activation function morphological control parameters for the current time step. The activation function is a composite function determined by the morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions. The time-series meteorological data is input into the backbone prediction network for forward computation. In the time series processing unit, the gating activation mechanism of the Long Short-Term Memory network remains unchanged, and the nonlinear activation function of the hidden state output stage is instantiated and configured using the morphological control parameters for the current time step. The spatial convolution and time series processing unit are used to extract the local meteorological state features of each grid at the current time step. During the model training phase, a local functional complexity index is calculated for each grid based on the morphological control parameters. The index is used to adjust the feature backpropagation gradient of the corresponding spatial location. A time smoothing regularization term is incorporated into the loss function to penalize the changes in morphological control parameters generated in consecutive time steps for the same grid. During the inference phase, meteorological monitoring results are output through the backbone prediction network.

[0005] Optionally, the step of extracting and encoding multi-source geographic information data for each geographic location raster to generate a static geographic feature vector corresponding to the raster includes: Continuous features are obtained by standardizing the range of the raster elevation and slope data; The underlying surface type data is converted into discrete feature vectors using one-hot encoding; The continuous features and discrete feature vectors are concatenated along the feature dimension to form an initial feature concatenation vector; The initial feature concatenation vector is input into a multilayer perceptron containing a nonlinear activation function. The nonlinear mapping transformation is performed through a fully connected layer, and the hidden layer representation with fixed dimensions is output as the static geographic feature vector corresponding to the grid.

[0006] Optionally, the step of extracting the local meteorological state features of each grid cell at the current moment using spatial convolution and time series processing units includes: Using a two-dimensional convolutional layer with a convolutional kernel of a preset size, a sliding window operation is performed on the grid map of the area to be monitored to extract the spatial proximity meteorological feature matrix of the central grid and its neighborhood. Extract the channel features corresponding to the target grid in the spatial proximity meteorological feature matrix, flatten the channel features into a one-dimensional spatial feature vector, combine it with the hidden state vector of the long short-term memory network at the previous time, and input it into the long short-term memory network at the current time. By using the forget gate, input gate, and output gate of a long short-term memory network to perform nonlinear fusion of historical meteorological memory features and current spatial features, the output hidden state vector at the current moment is the local meteorological state feature of the grid.

[0007] Optionally, the parameter generation module receives a static geographic feature vector and the local meteorological state features of the previous time step, and outputs the activation function morphological control parameters for the current time step, including: The static geographic feature vector and the local meteorological state features of the previous moment are concatenated along the feature dimension to construct a joint context vector; The joint context vector is input into a multi-layer feedforward neural network, the last layer of which is a grouped linear fully connected layer. The grouped linear fully connected layer divides the output vector into multiple independent parameter subsets according to a preset dimension. The parameter subsets correspond to the amplitude coefficient and frequency coefficient of the periodic basis function and the scaling factor of the non-periodic basis function, which together constitute the activation function shape control parameters.

[0008] Optionally, the activation function is a composite function determined by morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions, including: The aperiodic basis functions are defined as linear units with leakage correction, and the periodic basis functions are defined as sine functions; Substitute the input variables into the aperiodic basis function and the periodic basis function respectively; The scaling factor in the shape control parameters is used as a weight to multiply the output value of the linear unit with leakage correction, and the amplitude coefficient and frequency coefficient in the shape control parameters are used as the amplitude and angular frequency of the sine function, respectively, to calculate the output value of the periodic term. Add the scaled aperiodic term result to the periodic term output value to obtain the output response value of the activation function of the instantiated configuration.

[0009] Optionally, during the model training phase, calculating a local functional complexity index for each grid based on morphological control parameters, and using the index to adjust the feature backpropagation gradient at the corresponding spatial location, includes: Extract the amplitude coefficient and frequency coefficient of the periodic basis function from the shape control parameters corresponding to the current grid. Calculate the product of the absolute value of the amplitude coefficient and the absolute value of the frequency coefficient, and add a preset smoothing constant to the product result to obtain the local functional complexity index. When updating the weights of the backbone prediction network through backpropagation, a bounded spatial enhancement coefficient is generated based on the local functional complexity index, and the bounded spatial enhancement coefficient is multiplied into the feature backpropagation gradient of the corresponding spatial location, so that the grid region with higher local functional complexity contributes more to the update of the shared weights within a preset upper limit.

[0010] Optionally, incorporating a time-smoothing regularization term into the loss function to penalize changes in shape control parameters generated at consecutive time steps for the same raster includes: Obtain the shape control parameter vector output by the same geographic location raster at the current time step, and the historical shape control parameter vector output at the previous consecutive time step; Calculate the difference vector between the current time step and the shape control parameter vector of the previous time step; Solve for the square of the L2 norm of the difference vector, and use the square of the L2 norm of the difference vector as a penalty term for a single time step; The penalty terms for all monitored grids and all time steps are averaged and multiplied by a preset regularization hyperparameter to obtain the time smoothing regularization term. The time smoothing regularization term is added to the meteorological prediction error loss to form the total loss function used for joint optimization of model parameters.

[0011] According to another aspect of the present invention, a geographic information-driven meteorological monitoring system is proposed, comprising the following modules: The module is used to acquire time-series meteorological data and multi-source geographic information data including elevation, slope and underlying surface type of the area to be monitored. It extracts multi-source geographic information data and encodes it for each geographic location raster, generates a static geographic feature vector corresponding to the raster, and constructs a model including a parameter generation module and a backbone prediction network. The calculation module receives static geographic feature vectors and local meteorological state features from the previous time step from the parameter generation module, and outputs the activation function morphological control parameters for the current time step. The activation function is a composite function determined by the morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions. The time-series meteorological data is input into the backbone prediction network for forward computation. In the time series processing unit, the gating activation mechanism of the Long Short-Term Memory network remains unchanged, and the nonlinear activation function of the hidden state output stage is instantiated and configured by the morphological control parameters for the current time step. The spatial convolution and time series processing unit are used to extract the local meteorological state features of each grid at the current time step. The output module is used to calculate the local functional complexity index for each grid based on the morphological control parameters during the model training phase, adjust the feature backpropagation gradient of the corresponding spatial location using the index, incorporate a time smoothing regularization term into the loss function, and penalize the changes in morphological control parameters generated in consecutive time steps for the same grid; during the inference phase, it outputs meteorological monitoring results through the backbone prediction network.

[0012] Preferably, the step of extracting and encoding multi-source geographic information data for each geographic location raster to generate a static geographic feature vector corresponding to the raster includes: Continuous features are obtained by standardizing the range of the raster elevation and slope data; The underlying surface type data is converted into discrete feature vectors using one-hot encoding; The continuous features and discrete feature vectors are concatenated along the feature dimension to form an initial feature concatenation vector; The initial feature concatenation vector is input into a multilayer perceptron containing a nonlinear activation function. The nonlinear mapping transformation is performed through a fully connected layer, and the hidden layer representation with fixed dimensions is output as the static geographic feature vector corresponding to the grid.

[0013] Preferably, the step of extracting the local meteorological state features of each grid cell at the current moment using spatial convolution and time series processing units includes: Using a two-dimensional convolutional layer with a convolutional kernel of a preset size, a sliding window operation is performed on the grid map of the area to be monitored to extract the spatial proximity meteorological feature matrix of the central grid and its neighborhood. Extract the channel features corresponding to the target grid in the spatial proximity meteorological feature matrix, flatten the channel features into a one-dimensional spatial feature vector, combine it with the hidden state vector of the long short-term memory network at the previous time, and input it into the long short-term memory network at the current time. By using the forget gate, input gate, and output gate of a long short-term memory network to perform nonlinear fusion of historical meteorological memory features and current spatial features, the output hidden state vector at the current moment is the local meteorological state feature of the grid.

[0014] Preferably, the parameter generation module receives a static geographic feature vector and the local meteorological state features of the previous time step, and outputs the activation function morphological control parameters for the current time step, including: The static geographic feature vector and the local meteorological state features of the previous moment are concatenated along the feature dimension to construct a joint context vector; The joint context vector is input into a multi-layer feedforward neural network, the last layer of which is a grouped linear fully connected layer. The grouped linear fully connected layer divides the output vector into multiple independent parameter subsets according to a preset dimension. The parameter subsets correspond to the amplitude coefficient and frequency coefficient of the periodic basis function and the scaling factor of the non-periodic basis function, which together constitute the activation function shape control parameters.

[0015] Preferably, the activation function is a composite function determined by morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions, including: The aperiodic basis functions are defined as linear units with leakage correction, and the periodic basis functions are defined as sine functions; Substitute the input variables into the aperiodic basis function and the periodic basis function respectively; The scaling factor in the shape control parameters is used as a weight to multiply the output value of the linear unit with leakage correction, and the amplitude coefficient and frequency coefficient in the shape control parameters are used as the amplitude and angular frequency of the sine function, respectively, to calculate the output value of the periodic term. Add the scaled aperiodic term result to the periodic term output value to obtain the output response value of the activation function of the instantiated configuration.

[0016] Preferably, during the model training phase, calculating a local functional complexity index for each grid based on morphological control parameters, and using the index to adjust the feature backpropagation gradient at the corresponding spatial location, includes: Extract the amplitude coefficient and frequency coefficient of the periodic basis function from the shape control parameters corresponding to the current grid. Calculate the product of the absolute value of the amplitude coefficient and the absolute value of the frequency coefficient, and add a preset smoothing constant to the product result to obtain the local functional complexity index. When updating the weights of the backbone prediction network through backpropagation, a bounded spatial enhancement coefficient is generated based on the local functional complexity index, and the bounded spatial enhancement coefficient is multiplied into the feature backpropagation gradient of the corresponding spatial location, so that the grid region with higher local functional complexity contributes more to the update of the shared weights within a preset upper limit.

[0017] Preferably, incorporating a time-smoothing regularization term into the loss function to penalize changes in shape control parameters generated at consecutive time steps within the same raster includes: Obtain the shape control parameter vector output by the same geographic location raster at the current time step, and the historical shape control parameter vector output at the previous consecutive time step; Calculate the difference vector between the current time step and the shape control parameter vector of the previous time step; Solve for the square of the L2 norm of the difference vector, and use the square of the L2 norm of the difference vector as a penalty term for a single time step; The penalty terms for all monitored grids and all time steps are averaged and multiplied by a preset regularization hyperparameter to obtain the time smoothing regularization term. The time smoothing regularization term is added to the meteorological prediction error loss to form the total loss function used for joint optimization of model parameters.

[0018] This invention integrates time-series meteorological data with multi-source geographic information such as elevation, slope, and underlying surface type to construct static geographic feature vectors for each geographic location raster, enhancing the model's ability to perceive spatial differences in complex terrain. By generating morphological control parameters from the static geographic features and preceding meteorological state input parameters, these parameters are used to instantiate and configure the composite activation function in the hidden state output stage of the time-series processing unit. This mechanism combines periodic and non-periodic basis functions to adapt to the data distribution characteristics of different regions and time periods, enhancing the model's nonlinear expressive power. During training, a local functional complexity index is calculated based on the morphological control parameters to adjust the backpropagation gradient, and a time smoothing regularization term is used to constrain parameter changes across consecutive time steps, ensuring a smooth transition in the activation function's morphology, thereby improving the spatial precision and temporal continuity of meteorological monitoring results. Attached Figure Description

[0019] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram of the response curve for a composite activation node; Figure 3 This is a schematic diagram showing the temporal continuity comparison curves under regularization loss constraints. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] In the first embodiment, the present invention proposes a meteorological monitoring method based on geographic information, such as... Figure 1 It includes the following steps: S1. Acquire time-series meteorological data and multi-source geographic information data including elevation, slope and underlying surface type for the area to be monitored. Extract multi-source geographic information data and encode it for each geographic location raster, generate a static geographic feature vector corresponding to the raster, and construct a model including a parameter generation module and a backbone prediction network.

[0022] Multi-source geographic information data is obtained by reading the TIFF format file of the digital elevation model (DEM) of the monitored area and land cover classification remote sensing image data, and time-series meteorological data is obtained by reading historical meteorological datasets. The slope value of each geographic raster cell is calculated using a terrain attribute calculation function on the DEM data. Discrete underlying surface type labels are processed by one-hot encoding, and continuous elevation and slope values ​​are normalized to the range of 0 to 1 using a maximum-minimum normalization method. In the channel dimension, the normalized elevation matrix, slope matrix, and one-hot encoded underlying surface tensor are concatenated to generate an initial feature concatenation vector for each geographic raster. This initial feature concatenation vector is then input into a multilayer perceptron for nonlinear mapping transformation, outputting a static geographic feature vector with fixed dimensions. An artificial intelligence model is constructed, and a multilayer perceptron is built as a parameter generation module. A backbone prediction network is constructed using a combination of two-dimensional convolutional layer operators and long short-term memory network units.

[0023] In an optional embodiment, the step of extracting and encoding multi-source geographic information data for each geographic location raster to generate a static geographic feature vector corresponding to the raster includes: Continuous features are obtained by standardizing the range of the raster elevation and slope data; The underlying surface type data is converted into discrete feature vectors using one-hot encoding; The continuous features and discrete feature vectors are concatenated along the feature dimension to form an initial feature concatenation vector; The initial feature concatenation vector is input into a multilayer perceptron containing a nonlinear activation function. The nonlinear mapping transformation is performed through a fully connected layer, and the hidden layer representation with fixed dimensions is output as the static geographic feature vector corresponding to the grid.

[0024] Elevation and slope data from multi-source geographic information are mapped to a range of 0 to 1 using maximum and minimum value scaling. Taking an elevation data range of 5000 meters and 0 meters as an example, when the actual elevation of a certain raster is 2500 meters, the standardized feature value is 0.5. Slope data is linearly transformed within the range of 0 to 90 degrees to generate a 2D continuous feature. For underlying surface type data, if the area contains 10 categories including water, vegetation, buildings, and bare soil, these 10 categories are converted into 10-dimensional one-hot encoded vectors. For example, the vegetation category is encoded as 0, 1, 0, 0, 0, 0, 0, 0, 0, 0. The 2D continuous feature and the 10D discrete feature are concatenated along the feature channel dimension to obtain an initial feature concatenation vector of length 12. The input to the multilayer perceptron network is an initial feature concatenation vector of length 12. The network structure is a fully connected network with three hidden layers. The input layer dimension is 12, and the number of neurons in the three hidden layers is set to 64, 32, and 16 respectively. Nonlinear mapping is achieved through fully connected layers that first increase the dimension of the expression and then compress it layer by layer, finally outputting a 16-dimensional hidden layer representation. After linear calculation at each layer, a linear rectified function or a Gaussian error linear unit is used as a nonlinear activation function, and a random deactivation layer with a neuron retention probability of 0.9 is input after the activation operation to prevent overfitting. The multilayer perceptron outputs a one-dimensional hidden layer representation vector with a fixed dimension of 16. The 16-dimensional vector does not contain time-varying sequence information and serves as a static geographic feature vector corresponding to and bound to the current grid position, which is input to the subsequent parameter generation module.

[0025] S2, the parameter generation module receives the static geographic feature vector and the local meteorological state features of the previous time step, and outputs the activation function morphological control parameters for the current time step. The activation function is a composite function determined by the morphological control parameters, which is implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions. The time-series meteorological data is input into the backbone prediction network for forward computation. In the time series processing unit, the gating activation mechanism of the Long Short-Term Memory network remains unchanged, and the nonlinear activation function of the hidden state output stage is instantiated and configured by the morphological control parameters for the current time step. The spatial convolution and time series processing unit are used to extract the local meteorological state features of each grid at the current time step.

[0026] The static geographic feature vector is concatenated with the local meteorological state features output by the Long Short-Term Memory network at the previous time step along the feature channel dimension. The local meteorological state features are either the hidden state vector of the corresponding raster or a hidden state tensor composed of the hidden states of each raster. For the first time step of the sequence, the local meteorological state features from the previous time step are initialized as a zero vector or a learnable initial state vector. The concatenated feature tensor is input into the multilayer perceptron parameter generation module. After processing with fully connected linear mapping and grouped activation mapping, three feature map tensors with the same resolution as the spatial raster at the current time step are output, serving as the aperiodic basis function scaling factor tensor, the periodic basis function amplitude parameter tensor, and the frequency parameter tensor in the morphological control parameters, respectively. A composite activation function algorithm module is constructed, setting the aperiodic basis function as a linear unit with leakage correction and the periodic basis function as a sine function. The scaling factor tensor is multiplied by the output of the linear unit with leakage correction, and then the amplitude parameter tensor is multiplied by the sine function calculated by scaling the independent variable using the frequency parameter tensor, thus achieving the composite activation transformation. In the forward time-step iteration loop of the backbone prediction network, the memory gating mechanism is maintained using the original Sigmoid and Tanh functions to ensure the numerical stability of the long-term memory information flow. The nonlinear activation function in the long short-term memory network cell state to hidden state feature output transformation is replaced with the aforementioned composite activation function instantiated and configured with the current morphological control parameters. That is, the original activation mechanisms of the forget gate, input gate, output gate, and candidate state generation function remain unchanged, and only the nonlinear response of the hidden state output stage is morphologically adaptively controlled. In each sequence time step, the spatial topological features of the meteorological input data are extracted. Combined with the addition and multiplication logic of the cell states in the time series unit, the updated hidden state vector is generated. The hidden feature tensor is composed of the hidden state vectors of each grid, which serves as the local meteorological state feature extracted for each grid.

[0027] In an optional embodiment, the extraction of local meteorological state features for each grid cell at the current moment using spatial convolution and time series processing units includes: Using a two-dimensional convolutional layer with a convolutional kernel of a preset size, a sliding window operation is performed on the grid map of the area to be monitored to extract the spatial proximity meteorological feature matrix of the central grid and its neighborhood. Extract the channel features corresponding to the target grid in the spatial proximity meteorological feature matrix, flatten the channel features into a one-dimensional spatial feature vector, combine it with the hidden state vector of the long short-term memory network at the previous time, and input it into the long short-term memory network at the current time. By using the forget gate, input gate, and output gate of a long short-term memory network to perform nonlinear fusion of historical meteorological memory features and current spatial features, the output hidden state vector at the current moment is the local meteorological state feature of the grid.

[0028] The input to the spatial convolutional network is a four-dimensional spatiotemporal meteorological data tensor of the area to be monitored, with dimensions set to time step × height × width × 4. The four input channels correspond to temperature, humidity, wind speed, and air pressure. At each time slice, the meteorological feature map is organized according to the channel dimensions and then input into a two-dimensional convolutional layer for processing. The network uses a 3×3 or 5×5 two-dimensional convolutional kernel to perform a sliding window operation on the raster image of each time slice, with a convolution stride set to 1. When the kernel size is 3×3, a zero-padding strategy is used at the four edges; when the kernel size is 5×5... At 5 o'clock, a two-layer zero-padding strategy is adopted at the four edges to ensure that the spatial resolution of the output feature map is consistent with that of the original raster image. After the dot product calculation of 32 independent filters, when the convolution kernel size is 3×3, the spatial interaction response features of the center raster and the surrounding 8 adjacent grids of each location are extracted; when the convolution kernel size is 5×5, the spatial interaction response features of the center raster and the surrounding 24 adjacent grids of each location are extracted. The output of this spatial convolutional network is a three-dimensional spatial proximity meteorological feature matrix with the number of channels expanded to 32, and the size is height×width×32. For a specific spatial coordinate in this matrix, the feature channel of length 32 is extracted and flattened into a one-dimensional feature vector. The input of the Long Short-Term Memory network is the flattened 32-dimensional spatial feature vector and the hidden state vector of dimension 64 from the previous time step. Within the network structure, the forget gate outputs a retention probability vector matrix with values ​​ranging from 0 to 1, used to attenuate redundant historical memory features. The input gate generates state candidate values ​​from the current 32-dimensional spatial features and weights them, fusing them into the backbone memory. The output gate performs weighted calculations based on the updated cell state combined with a composite activation function instantiated from the current morphological control parameters. The output of this Long Short-Term Memory network is a 64-dimensional hidden state vector at the current time step. This vector aggregates the short-term meteorological change patterns and spatial proximity characteristics of the local coordinate grid, and saves and transmits them as local meteorological state characteristics.

[0029] In an optional embodiment, the parameter generation module receives a static geographic feature vector and local meteorological state features from the previous time step, and outputs activation function morphological control parameters for the current time step, including: The static geographic feature vector and the local meteorological state features of the previous moment are concatenated along the feature dimension to construct a joint context vector; The joint context vector is input into a multi-layer feedforward neural network, the last layer of which is a grouped linear fully connected layer. The grouped linear fully connected layer divides the output vector into multiple independent parameter subsets according to a preset dimension. The parameter subsets correspond to the amplitude coefficient and frequency coefficient of the periodic basis function and the scaling factor of the non-periodic basis function, which together constitute the activation function shape control parameters.

[0030] For any given grid, the parameter generation module obtains a static geographic feature vector of dimension 16 and a hidden feature vector of the previous time-state local weather conditions of dimension 64, and concatenates them on the feature axis to construct a joint context vector of dimension 80. The input to the multilayer feedforward neural network is this joint context vector of dimension 80. The feedforward neural network has a two-layer structure. The first hidden layer contains 128 neurons and combines layer normalization and a sigmoid nonlinear activation function to fuse static terrain features and dynamic meteorological features. In the last layer of the feedforward neural network, a linear fully connected output layer based on group operations is set to separate different parameter spaces. The output of this fully connected layer is a long vector of dimension 192, that is, when the number of data channels in a single layer of the backbone prediction network is C=64, the total output dimension is 3×C. The 192-dimensional long vector is divided into three independent parameter subsets, each 64 dimensions long, based on the channel sequence index: the first 64 dimensions are processed and used as the amplitude coefficient vector of the periodic basis function; the middle 64 dimensions are processed and mapped to a preset interval of 0.1 to 5.0, serving as the frequency coefficient vector; the last 64 dimensions are exponentially processed and limited to a preset upper limit to ensure positive values, serving as the scaling factor vector of the aperiodic basis function. These three parameter subsets constitute the morphological control parameters of the meteorological evolution pattern corresponding to the current characteristic channel.

[0031] In an optional embodiment, the activation function is a composite function determined by morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions, including: The aperiodic basis functions are defined as linear units with leakage correction, and the periodic basis functions are defined as sine functions; Substitute the input variables into the aperiodic basis function and the periodic basis function respectively; The scaling factor in the shape control parameters is used as a weight to multiply the output value of the linear unit with leakage correction, and the amplitude coefficient and frequency coefficient in the shape control parameters are used as the amplitude and angular frequency of the sine function, respectively, to calculate the output value of the periodic term. Add the scaled aperiodic term result to the periodic term output value to obtain the output response value of the activation function of the instantiated configuration.

[0032] The composite activation function comprises two parallel scalar computation branches. The linear unit with leakage correction serves as an aperiodic basis function, with a negative half-axis slope set to 0.01, denoted as y=max(0.01x,x). The sine function serves as a periodic basis function, characterizing the periodic meteorological oscillations modulated by the frequency parameter. For the node input scalar x during forward propagation, this input simultaneously and independently passes through the two function branches to calculate the fundamental response value. The results are fused using the parameter scalars output by the feedforward network. Taking a scaling factor α=1.5, amplitude coefficient A=0.8, and frequency coefficient ω=2.5 as an example: multiplying the result of the linear unit with leakage correction by 1.5 as a scaling weight yields 1.5×max(0.01x,x); using A and ω, 0.8×sin(2.5×x) is calculated as the periodic fluctuation response term. At the end of the computation graph node, the weighted results of the two branches are summed element-wise, i.e., y = 1.5 × max(0.01x, x) + 0.8 × sin(2.5 × x), where x is the dimensionless node input after linear mapping of the network, and the frequency coefficient is used to scale the dimensionless feature domain to keep the input independent variable of the sine function dimensionless. This summation result is used as the final nonlinear response value of the node under the environmental configuration and input into the subsequent computation layers. The layered output superposition characteristics of the composite activation function are as follows: Figure 2 As shown.

[0033] S3. During the model training phase, a local functional complexity index is calculated for each grid based on the morphological control parameters. The index is used to adjust the feature backpropagation gradient of the corresponding spatial location. A time smoothing regularization term is incorporated into the loss function to penalize the changes in morphological control parameters generated in consecutive time steps for the same grid. During the inference phase, meteorological monitoring results are output through the backbone prediction network.

[0034] In the backpropagation optimization phase of the AI ​​model training, the amplitude parameter tensor and frequency parameter tensor corresponding to each geospatial grid location at the current time step are extracted. The product of the absolute values ​​of the amplitude coefficient and the frequency coefficient is calculated, and a preset smoothing constant is superimposed on the product result as a local functional complexity index representing the intensity of nonlinear oscillations. A gradient hook mechanism is used to intercept the gradient tensor of the characteristic partial derivatives of the composite activation function layer. Based on the aforementioned local functional complexity index, a bounded spatial enhancement coefficient is generated and multiplied into the backpropagation gradient of the corresponding spatial location. This allows the spatial grid region with higher complexity to obtain a stronger learning signal within a preset upper limit. The gradient hook mechanism refers to a callback mechanism that intercepts, reads, or modifies the gradient of a specified layer or tensor during the backpropagation process of the neural network, used to adjust the gradient propagation method without changing the forward network structure. When constructing the loss function for joint optimization, the squared error between the model-predicted meteorological numerical tensor and the actual observed tensor is calculated as the basic prediction loss.

[0035] The complete morphological control parameter vectors output from the same spatial raster at two consecutive adjacent time steps are extracted. These vectors include scaling factors, amplitude parameters, and frequency parameters. The L2 norm squared of the parameter difference vectors between adjacent time steps is calculated, and a temporal smoothing regularization term is constructed. This regularization term is then combined with the aforementioned basic prediction loss using a scalar addition operation to obtain the overall loss value. Gradient descent is performed to iteratively update the weights of the neural network model. During the inference monitoring phase, the hidden state feature map at the end of the backbone prediction network is input into a 1×1 convolutional mapping layer to generate a meteorological element prediction array. Finally, the meteorological element prediction data array for the monitoring area is output at the spatiotemporal tensor level to achieve the meteorological monitoring results.

[0036] In an optional embodiment, the step of calculating a local functional complexity index for each grid based on morphological control parameters during the model training phase, and adjusting the feature backpropagation gradient at the corresponding spatial location using the index, includes: Extract the amplitude coefficient and frequency coefficient of the periodic basis function from the shape control parameters corresponding to the current grid. Calculate the product of the absolute value of the amplitude coefficient and the absolute value of the frequency coefficient, and add a preset smoothing constant to the product result to obtain the local functional complexity index. When updating the weights of the backbone prediction network through backpropagation, a bounded spatial enhancement coefficient is generated based on the local functional complexity index, and the bounded spatial enhancement coefficient is multiplied into the feature backpropagation gradient of the corresponding spatial location, so that the grid region with higher local functional complexity contributes more to the update of the shared weights within a preset upper limit.

[0037] During the model backpropagation phase, the local functional complexity calculation module extracts the amplitude coefficients at spatial coordinates i,j and feature channel k. With frequency coefficient Take the absolute values ​​of both, and multiply the absolute values ​​point by point to obtain a non-negative estimate of the oscillation amplitude. To avoid local functional complexity metrics being zero and to improve the numerical stability of subsequent proportional calculations, a very small positive smoothing constant of preset size 1e-5 is superimposed. Constructing a local functional complexity index If the product of a node is 5.0, then the local function complexity index at that point is calculated to be 5.00001. When the model executes the backpropagation algorithm to calculate the global update gradient matrix, the bounded space augmentation coefficient is calculated based on the aforementioned local function complexity index, for example, by setting... Where ρ is the preset maximum enhancement amplitude and τ is the preset stability constant, so that It always lies between 1 and 1+ρ; then, by applying element-wise multiplication-based weight scaling at spatial coordinates i,j,k, it will... Multiply by the local backward gradient tensor propagated at that location. (Corresponding to the above...) In the case where ρ = 5.00001, if ρ = 1 and τ = 5, then The value is approximately 1.50. This spatial gradient enhancement mechanism enables regions of complex meteorological signals with high amplitude and high frequency to obtain a larger gradient contribution within a bounded range, thereby improving the backbone network's ability to learn about regions with complex weather changes. At the same time, it avoids gradient runaway through upper limit constraints.

[0038] In an optional embodiment, incorporating a temporal smoothing regularization term into the loss function to penalize changes in morphological control parameters generated at consecutive time steps within the same raster includes: Obtain the shape control parameter vector output by the same geographic location raster at the current time step, and the historical shape control parameter vector output at the previous consecutive time step; Calculate the difference vector between the current time step and the shape control parameter vector of the previous time step; Solve for the square of the L2 norm of the difference vector, and use the square of the L2 norm of the difference vector as a penalty term for a single time step; The penalty terms for all monitored grids and all time steps are averaged and multiplied by a preset regularization hyperparameter to obtain the time smoothing regularization term. The time smoothing regularization term is added to the meteorological prediction error loss to form the total loss function used for joint optimization of model parameters.

[0039] For the forward propagation data stream in the training batch, a buffer queue is established to record the shape control parameter vector output by the raster with spatial coordinates i,j at each time step. Let the parameter vector output at the current time step t, which includes scaling, amplitude, and frequency dimensions, be... Retrieve the parameter vector corresponding to the previous consecutive time step t-1 at the same position from the cache. The scaling factor, amplitude coefficient, and frequency coefficient are normalized according to their respective preset value ranges and then concatenated to obtain a normalized shape control parameter vector. Element-wise subtraction is then performed on these two normalized 192-dimensional tensors to generate the corresponding difference variation vector. The algorithm calculates the sum of squares of each element in the 192-dimensional difference vector, which is the square of the L2 norm of the difference vector, to form a non-negative scalar penalty score. The algorithm spans a preset time window, for example, 24 time steps corresponding to 24 hours, and performs summation and averaging operations on the penalty scores of M×N raster nodes across the entire image.

[0040] The calculated expected mean is denoted as the unweighted smoothing penalty base value. and will Multiplying by a preset regularization decay hyperparameter λ, for example, setting λ to 0.05, yields the time smoothing regularization term. =λ× When constructing the total loss function, the task master loss, representing the prediction error, will be used. With time smoothing regularization term Add them together, the formula is as follows Backpropagation using this combined loss mechanism can constrain the fluctuations of activation network parameters over continuous time series during model optimization, improving the temporal smoothness of parameter changes. The effect of the time smoothing regularization term on controlling temporal parameter fluctuations is as follows: Figure 3 As shown.

[0041] In the second embodiment, the present invention also proposes a geographic information-driven meteorological monitoring system, comprising the following modules: The module is used to acquire time-series meteorological data and multi-source geographic information data including elevation, slope and underlying surface type of the area to be monitored. It extracts multi-source geographic information data and encodes it for each geographic location raster, generates a static geographic feature vector corresponding to the raster, and constructs a model including a parameter generation module and a backbone prediction network. The calculation module receives static geographic feature vectors and local meteorological state features from the previous time step from the parameter generation module, and outputs the activation function morphological control parameters for the current time step. The activation function is a composite function determined by the morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions. The time-series meteorological data is input into the backbone prediction network for forward computation. In the time series processing unit, the gating activation mechanism of the Long Short-Term Memory network remains unchanged, and the nonlinear activation function of the hidden state output stage is instantiated and configured by the morphological control parameters for the current time step. The spatial convolution and time series processing unit are used to extract the local meteorological state features of each grid at the current time step. The output module is used to calculate the local functional complexity index for each grid based on the morphological control parameters during the model training phase, adjust the feature backpropagation gradient of the corresponding spatial location using the index, incorporate a time smoothing regularization term into the loss function, and penalize the changes in morphological control parameters generated in consecutive time steps for the same grid; during the inference phase, it outputs meteorological monitoring results through the backbone prediction network.

[0042] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A meteorological monitoring method based on geographic information, characterized in that, include: Acquire time-series meteorological data and multi-source geographic information data including elevation, slope and underlying surface type for the area to be monitored. Extract multi-source geographic information data and encode it for each geographic location raster, generate a static geographic feature vector corresponding to the raster, and construct a model including a parameter generation module and a backbone prediction network. The parameter generation module receives static geographic feature vectors and local meteorological state features from the previous time step, and outputs the activation function morphological control parameters for the current time step. The activation function is a composite function determined by the morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions. The time-series meteorological data is input into the backbone prediction network for forward computation. In the time series processing unit, the gating activation mechanism of the Long Short-Term Memory network remains unchanged, and the nonlinear activation function of the hidden state output stage is instantiated and configured using the morphological control parameters for the current time step. The spatial convolution and time series processing unit are used to extract the local meteorological state features of each grid at the current time step. During the model training phase, a local functional complexity index is calculated for each grid based on the morphological control parameters. The index is used to adjust the feature backpropagation gradient of the corresponding spatial location. A time smoothing regularization term is incorporated into the loss function to penalize the changes in morphological control parameters generated in consecutive time steps for the same grid. During the inference phase, meteorological monitoring results are output through the backbone prediction network.

2. The method according to claim 1, characterized in that, The step of extracting and encoding multi-source geographic information data for each geographic location raster to generate a static geographic feature vector corresponding to the raster includes: Continuous features are obtained by standardizing the range of the raster elevation and slope data; The underlying surface type data is converted into discrete feature vectors using one-hot encoding; The continuous features and discrete feature vectors are concatenated along the feature dimension to form an initial feature concatenation vector; The initial feature concatenation vector is input into a multilayer perceptron containing a nonlinear activation function. The nonlinear mapping transformation is performed through a fully connected layer, and the hidden layer representation with fixed dimensions is output as the static geographic feature vector corresponding to the grid.

3. The method according to claim 2, characterized in that, The extraction of local meteorological state features for each grid cell at the current moment using spatial convolution and time series processing units includes: Using a two-dimensional convolutional layer with a convolutional kernel of a preset size, a sliding window operation is performed on the grid map of the area to be monitored to extract the spatial proximity meteorological feature matrix of the central grid and its neighborhood. Extract the channel features corresponding to the target grid in the spatial proximity meteorological feature matrix, flatten the channel features into a one-dimensional spatial feature vector, combine it with the hidden state vector of the long short-term memory network at the previous time, and input it into the long short-term memory network at the current time. By using the forget gate, input gate, and output gate of a long short-term memory network to perform nonlinear fusion of historical meteorological memory features and current spatial features, the output hidden state vector at the current moment is the local meteorological state feature of the grid.

4. The method according to claim 1, characterized in that, The parameter generation module receives a static geographic feature vector and the local meteorological state features from the previous time step, and outputs the activation function morphological control parameters for the current time step, including: The static geographic feature vector and the local meteorological state features of the previous moment are concatenated along the feature dimension to construct a joint context vector; The joint context vector is input into a multi-layer feedforward neural network, the last layer of which is a grouped linear fully connected layer. The grouped linear fully connected layer divides the output vector into multiple independent parameter subsets according to a preset dimension. The parameter subsets correspond to the amplitude coefficient and frequency coefficient of the periodic basis function and the scaling factor of the non-periodic basis function, which together constitute the activation function shape control parameters.

5. The method according to claim 1 or 4, characterized in that, The activation function is a composite function determined by morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions, including: The aperiodic basis functions are defined as linear units with leakage correction, and the periodic basis functions are defined as sine functions; Substitute the input variables into the aperiodic basis function and the periodic basis function respectively; The scaling factor in the shape control parameters is used as a weight to multiply the output value of the linear unit with leakage correction, and the amplitude coefficient and frequency coefficient in the shape control parameters are used as the amplitude and angular frequency of the sine function, respectively, to calculate the output value of the periodic term. Add the scaled aperiodic term result to the periodic term output value to obtain the output response value of the activation function of the instantiated configuration.

6. The method according to claim 1, characterized in that, During the model training phase, a local functional complexity index is calculated for each grid based on morphological control parameters. This index is then used to adjust the feature backpropagation gradient at the corresponding spatial location, including: Extract the amplitude coefficient and frequency coefficient of the periodic basis function from the shape control parameters corresponding to the current grid. Calculate the product of the absolute value of the amplitude coefficient and the absolute value of the frequency coefficient, and add a preset smoothing constant to the product result to obtain the local functional complexity index. When updating the weights of the backbone prediction network through backpropagation, a bounded spatial enhancement coefficient is generated based on the local functional complexity index, and the bounded spatial enhancement coefficient is multiplied into the feature backpropagation gradient of the corresponding spatial location, so that the grid region with higher local functional complexity contributes more to the update of the shared weights within a preset upper limit.

7. The method according to claim 1, characterized in that, The incorporation of a time-smoothing regularization term into the loss function, which penalizes changes in shape control parameters generated at consecutive time steps within the same raster, includes: Obtain the shape control parameter vector output by the same geographic location raster at the current time step, and the historical shape control parameter vector output at the previous consecutive time step; Calculate the difference vector between the current time step and the shape control parameter vector of the previous time step; Solve for the square of the L2 norm of the difference vector, and use the square of the L2 norm of the difference vector as a penalty term for a single time step; The penalty terms for all monitored grids and all time steps are averaged and multiplied by a preset regularization hyperparameter to obtain the time smoothing regularization term. The time smoothing regularization term is added to the meteorological prediction error loss to form the total loss function used for joint optimization of model parameters.

8. A geographic information-driven meteorological monitoring system, characterized in that, include: The module is used to acquire time-series meteorological data and multi-source geographic information data including elevation, slope and underlying surface type of the area to be monitored. It extracts multi-source geographic information data and encodes it for each geographic location raster, generates a static geographic feature vector corresponding to the raster, and constructs a model including a parameter generation module and a backbone prediction network. The calculation module receives static geographic feature vectors and local meteorological state features from the previous time step from the parameter generation module, and outputs the activation function morphological control parameters for the current time step. The activation function is a composite function determined by the morphological control parameters, implemented by combining aperiodic and periodic basis functions. The morphological control parameters determine the amplitude and frequency of the periodic basis functions and the scaling weights of the aperiodic basis functions. The time-series meteorological data is input into the backbone prediction network for forward computation. In the time series processing unit, the gating activation mechanism of the Long Short-Term Memory network remains unchanged, and the nonlinear activation function of the hidden state output stage is instantiated and configured by the morphological control parameters for the current time step. The spatial convolution and time series processing unit are used to extract the local meteorological state features of each grid at the current time step. The output module is used to calculate the local functional complexity index for each grid based on the morphological control parameters during the model training phase, adjust the feature backpropagation gradient of the corresponding spatial location using the index, incorporate a time smoothing regularization term into the loss function, and penalize the changes in morphological control parameters generated in consecutive time steps for the same grid; during the inference phase, it outputs meteorological monitoring results through the backbone prediction network.

9. The system according to claim 8, characterized in that, The step of extracting and encoding multi-source geographic information data for each geographic location raster to generate a static geographic feature vector corresponding to the raster includes: Continuous features are obtained by standardizing the range of the raster elevation and slope data; The underlying surface type data is converted into discrete feature vectors using one-hot encoding; The continuous features and discrete feature vectors are concatenated along the feature dimension to form an initial feature concatenation vector; The initial feature concatenation vector is input into a multilayer perceptron containing a nonlinear activation function. The nonlinear mapping transformation is performed through a fully connected layer, and the hidden layer representation with fixed dimensions is output as the static geographic feature vector corresponding to the grid.

10. The system according to claim 8, characterized in that, The extraction of local meteorological state features for each grid cell at the current moment using spatial convolution and time series processing units includes: Using a two-dimensional convolutional layer with a convolutional kernel of a preset size, a sliding window operation is performed on the grid map of the area to be monitored to extract the spatial proximity meteorological feature matrix of the central grid and its neighborhood. Extract the channel features corresponding to the target grid in the spatial proximity meteorological feature matrix, flatten the channel features into a one-dimensional spatial feature vector, combine it with the hidden state vector of the long short-term memory network at the previous time, and input it into the long short-term memory network at the current time. By using the forget gate, input gate, and output gate of a long short-term memory network to perform nonlinear fusion of historical meteorological memory features and current spatial features, the output hidden state vector at the current moment is the local meteorological state feature of the grid.

Images