A geographic multi-factor classification and fusion modeling method based on state space model

Abstract

The present invention discloses a geographic multi-factor classification and fusion modeling method based on a state-space model, which relates to the fields of geographic information and environmental monitoring, including: sampling geographic data and dividing geographic element factors; optimizing temporal feature representation through multi-scale temporal information encoding, and mapping spatial coordinates to a continuous vector space; realizing selective sequence scanning and long-range dependency capture using a state-space model through selective encoder design and feature extraction; adopting a residual gated fusion module and multi-source factor selective fusion, realizing the gradual fusion of different types of geographic factors and spatiotemporal information through a cascade structure, and constructing a decoder to generate results suitable for specific earth science applications. The method of the present invention realizes differentiated processing and fusion of different types of geographic factors through a multi-factor partitioning strategy, a selection mechanism based on a state-space model, and a residual gated fusion mechanism, thereby improving modeling accuracy while maintaining linear computational complexity.

Description

Technical Field

[0001] The present invention relates to the fields of geographic information, remote sensing, environmental monitoring, and climate analysis, and in particular to a geographic multi-factor classification and fusion modeling method based on a state space model. Background Art

[0002] With the development of Earth observation technologies, sensor networks, and the Internet of Things, multi-source, multi-scale, and multi-resolution geographic spatiotemporal data continues to emerge. These include satellite remote sensing imagery, ground-based monitoring station data, drone data, and mobile device positioning data. These data provide a rich source of information for fields such as environmental monitoring, climate change research, urban planning, and disaster warning. However, the multi-source heterogeneity, complex spatiotemporal dependencies, and high-dimensional nonlinear relationships of these data pose significant challenges to data processing and modeling.

[0003] Existing geographic data modeling methods have the following limitations: statistical methods such as interpolation and weighted regression can take into account spatial autocorrelation, but have difficulty handling high-dimensional nonlinear relationships; machine learning methods such as random forests and support vector machines can handle complex nonlinear relationships, but have difficulty capturing the spatiotemporal dependencies between data and realizing multi-source data fusion; among deep learning methods, convolutional neural networks have difficulty effectively handling long-range dependencies, and recurrent neural networks have the gradient vanishing problem. Although Transformer-based models can capture long-range dependencies, their computational complexity increases quadratically with the sequence length, and they face serious problems of computational resource consumption and inefficiency in large-scale geographic data processing.

[0004] In addition, the geographic system simultaneously contains dynamic factors such as meteorological elements, relatively stable static factors such as terrain, and discrete category factors such as land cover types. These factors with different characteristics require differentiated processing strategies, but existing methods lack an effective multi-factor integration mechanism, making it difficult to simultaneously take into account both temporal periodicity and spatial location specificity, resulting in difficulty in achieving high-precision modeling while maintaining high computational efficiency. Summary of the Invention

[0005] The problem to be solved by the present invention is to provide a geographic multi-factor classification and fusion modeling method based on a state-space model. Through a multi-factor partitioning strategy, an adaptive spatiotemporal encoder and a selective fusion mechanism, the problems of multi-source heterogeneity, spatiotemporal dependence, multi-scale characteristics and high-dimensional complexity in geographic spatiotemporal data modeling are effectively solved, while maintaining high computational efficiency, the accuracy of geographic data modeling is significantly improved.

[0006] The present invention adopts the following technical solution: a geographic multi-factor classification and fusion modeling method based on a state space model, comprising the following steps:

[0007] Step 1: Geographic data sampling and factor division: Collect multi-source geographic spatiotemporal data and perform preprocessing, identify and divide geographic data into multiple factors based on temporal variation characteristics, construct a spatiotemporal reference grid, and perform gridding processing on each factor;

[0008] Step 2: Multi-scale temporal information encoding: Using optimized sine and cosine periodic encoding methods, the continuity of geographic data temporal information is preserved and periodic topological mapping is performed, and multi-scale periodic features are integrated through a hierarchical processing strategy.

[0009] Step 3: Adaptive spatial information encoding: mapping geographic data spatial coordinates to continuous vector space;

[0010] Step 4: Selective encoder construction and feature extraction: The input factors are nonlinearly transformed and spatially expanded through the projection layer, and selective sequence scanning and long-range dependency capture are performed through the state space model;

[0011] Step 5: Residual gating and multi-source factor fusion: The residual gating fusion module is used to selectively fuse multi-source factors, and different types of geographic factors and spatiotemporal information are gradually integrated through a cascade structure to generate a unified geographic spatiotemporal fusion representation;

[0012] Step 6. Decoder module construction and spatiotemporal prediction output: Construct a decoding module, take the unified geographic spatiotemporal fusion representation as input, perform dimensionality reduction processing through the combined structure of the KAN layer and the selective encoder module, use the Mamba component stacking structure to model the sequence tensor, capture the dependencies between data points, and map the features to the target output space, perform result analysis and post-processing, and generate the final prediction value, category result, spatial distribution map or simulated data.

[0013] Preferably, the geographic data sampling and factor division in step 1 is achieved through the following sub-steps:

[0014] Step 1.1: Collection and preprocessing of original geographic spatiotemporal data: collecting original geographic data containing time, spatial location, and multidimensional attribute information, and cleaning, normalizing, and filling missing values ​​in the data;

[0015] Step 1.2, multi-factor identification and division. This sub-step divides the data into three types of factors according to the temporal and spatial variation characteristics of geographic elements: dynamic geographic factors , static geographical factors , Category Geographic Factor ;

[0016] Step 1.3: Spatiotemporal grid construction and sampling: Construct a unified spatiotemporal reference coordinate system, determine the appropriate spatial resolution and temporal granularity based on the study area and application requirements, generate a regular grid based on the determined spatiotemporal resolution, and grid and resample the three types of factor data to ensure that the data are aligned in the same spatiotemporal reference frame.

[0017] In particular, dynamic factors refer to factors that show obvious changes in time series, such as meteorological factors such as temperature, precipitation, and air humidity; static factors refer to factors that remain relatively stable within the research time scale, such as terrain features such as terrain elevation, surface slope, and slope aspect; categorical factors refer to categorical data with discrete characteristics, such as land use type, vegetation cover type, and other data with clear categorical attributes.

[0018] Preferably, the multi-scale time information encoding in step 2 is implemented by the following sub-steps:

[0019] Step 2.1: Multi-scale time coding uses sine and cosine periodic coding to optimize continuity preservation, periodic topological mapping, and multi-dimensional periodic feature fusion. , assuming the time period length is , then the obtained two-dimensional encoding vector ;

[0020] Step 2.2: For multi-scale periodic feature fusion, a hierarchical processing strategy is used to fuse multi-scale periodic features, and root mean square normalization and KAN mapping are applied to convert the time variable into a continuous vector representation. .

[0021] Preferably, in step 3, the adaptive spatial information encoding, the adaptive spatial information encoder connects the original coordinate vector , and then apply the sequential KAN layer to embed the spatial information into a continuous vector middle.

[0022] Preferably, the selective encoder construction and feature extraction described in step 4 are achieved through the following sub-steps:

[0023] Step 4.1: Input feature projection and spatial expansion aims to increase feature diversity and prepare for subsequent selective scanning: the input geographic factor or fusion feature tensor Through the KAN projection layer, nonlinear transformation is performed to expand the feature space dimension to , and get the tensor ;

[0024] Step 4.2, feature dimension reshaping and projection: tensor Reshape it and pass it through the KAN projection layer again to map it to the hidden dimension , get a tensor suitable for sequence scanning ;

[0025] Step 4.3, Selective sequence scanning based on the selection state space model Mamba: tensor Input into the stacked structure composed of Mamba, each Mamba component uses the state space model selection mechanism to select along the embedding space dimension Perform efficient one-way scanning processing to capture long-range dependencies.

[0026] In particular, residual connections are applied between stacked Mamba components to enhance the stability of gradient propagation and model training. The process is expressed as:

[0027] ;

[0028] in, represents the residual connection, stands for root mean square normalization, represents a feedforward network, express The number of layers, express The number of stacking times, It is a sequence processing component based on the selective state space model (SSM), which efficiently captures the feature extraction unit of long-range dependencies through structured state space representation and selective scanning mechanism.

[0029] Step 4.4: Based on the “scan-select-activate” mechanism, enhance feature generation and convert the tensor processed by the Mamba component into pass The normalization layer and KAN projection layer perform dimension compression and nonlinear activation to obtain the final enhanced feature representation .

[0030] Preferably, the residual gating and multi-source factor fusion in step 5 are achieved through the following sub-steps:

[0031] Step 5.1: Residual gated fusion is based on a gate-like mechanism that uses a feedforward network FFN to accurately control the feature information flow. Given the input feature and , the output features are generated by the residual gated fusion module ;

[0032] Step 5.2: Select and fuse the dynamic factors and categorical factors, and combine the dynamic geographic factors obtained in step 1 Feature representation and categorical geographical factors The feature representation of is taken as input and input into the cascade structure of the first-level residual gate fusion and selective encoder modules to achieve interactive fusion of the two factor features and selective feature selection and enhancement, generating a fusion representation of dynamic and categorical geographic factors. ;

[0033] Step 5.3: Fusion of geographical factors and time coding: Fusion of dynamic and categorical geographical factors generated in step 5.2 The continuous vector representation generated by the time scale encoder obtained in step 2 As input, it is input into the cascade structure of the second-level residual gated fusion and selective encoder modules, and residual gated fusion and selective processing are performed to generate a representation that integrates time, dynamics and category geographical factors. ;

[0034] Step 5.4: Fusion of the fusion representation and the static factor. The fusion representation generated in step 5.3 is The static geographic factor obtained in step 1 The feature representation is taken as input and input into the cascade structure of the third-level residual gate fusion and selective encoder modules for fusion processing to generate a representation that integrates time and dynamic, classification, and static geographic factors. ;

[0035] Step 5.5: Generate the final geographic spatiotemporal fusion representation: Generate the fusion representation generated in step 5.4 The position encoding vector generated by the adaptive spatial information encoding in step 3 As input, it is input into the fourth-level cascade structure for final residual fusion and selective processing to generate a unified geographic spatiotemporal fusion representation containing all geographic factors and spatiotemporal information. , as the input of the decoder described in step 6.

[0036] Preferably, the decoder module in step 6 is constructed and outputs the spatiotemporal prediction, and receives the unified geographic spatiotemporal fusion representation generated after processing in step 5.5. It takes as input and generates final output suitable for specific downstream geoscience applications based on this input, which includes the following sub-steps:

[0037] Step 6.1: Reduce the dimension of the encoder output and represent the unified geographic spatiotemporal fusion obtained in step 5.5 As input, it undergoes nonlinear transformation and feature selection processing through a combination of multiple Kolmogorov-Arnold network layers and selective encoder modules to reduce the feature dimension and generate a sequence tensor suitable for downstream tasks. ;

[0038] Step 6.2: Sequence modeling based on state space Mamba, convert the sequence tensor generated in step 6.1 The input is a stacked structure composed of multiple Mamba components. Each Mamba component uses a state space model and a content-related selection mechanism to select along the sequence dimension. Model the proximity and dependencies between input data points to further extract and refine features ;

[0039] Step 6.3, the final output projection, the sequence tensor processed by step 6.2 The final nonlinear projection and dimension transformation is performed through one or more KAN layers. , mapping the feature space to the dimension of the target output space;

[0040] Step 6.4: Output generation: Project the output of step 6.3 based on the needs of specific downstream earth science tasks. After analysis and post-processing, it relies on various loss functions and end-to-end back-propagation training to generate output results in the form of final prediction values, classification results, spatial distribution maps or simulated data.

[0041] Preferably, end-to-end back-propagation training and parameter optimization includes the following sub-steps:

[0042] Step 7.1: Build an end-to-end differentiable computational graph from data input to final output.

[0043] Step 7.2: Select a loss function suitable for the task type and apply the optimizer to update the model parameters.

[0044] Step 7.3. Calculate the gradient of the loss function with respect to the model parameters through automatic differentiation.

[0045] Step 7.4: Iterate the training process on the computing platform until the model converges or reaches the preset accuracy requirement.

[0046] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0047] 1. Efficient processing of heterogeneous geographic data: The method of the present invention can adaptively process geographic data of different types and sources through the division strategy of dynamic, static and classified geographic factors, thereby improving the model's adaptability to heterogeneous data.

[0048] 2. Accurately capture spatiotemporal dependencies: By leveraging the efficient sequence processing capabilities of the state-space model and the precise numerical fitting capabilities of KAN, the present invention can simultaneously capture the temporal dynamics and spatial correlations of geographic data.

[0049] 3. Significantly improved computational efficiency: The method of the present invention adopts the Mamba selection mechanism with linear complexity, which can process long-sequence geographic spatiotemporal data more efficiently than the quadratic complexity of the traditional Transformer model.

[0050] 4. Significantly improved model accuracy: This invention effectively integrates multi-source geographic factor information through an adaptive spatiotemporal encoder and a selective fusion mechanism, thereby improving the prediction accuracy and generalization capability of the geographic spatiotemporal model.

[0051] 5. Strong adaptability: The method of the present invention adopts the cascade structure of residual gated fusion and selective encoder modules, and the modular design of the decoder enables it to flexibly adapt to different geographic spatiotemporal modeling tasks, including but not limited to air quality prediction, climate change analysis, land use change monitoring and other application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flowchart of the geographic multi-factor classification and fusion modeling method based on the state space model of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the application are further elaborated in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in this field on this embodiment fall within the scope of protection of the present invention. At the same time, the step numbers in the embodiments of the present invention are only set for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0054] In one embodiment of the present invention, a geographic multi-factor classification and fusion modeling method based on a state space model is provided. Figure 1 As shown, the following steps are included:

[0055] (1) Geographic data sampling and factor division.

[0056] (1.1) Collection and preprocessing of original geographic spatiotemporal data:

[0057] This example collects raw geographic data from a variety of sources, including satellite remote sensing imagery, ground monitoring stations, drones, and mobile devices. This data contains temporal information, spatial location information, and multidimensional attribute information. To ensure data quality and consistency, the collected data undergoes necessary preprocessing, including but not limited to: data cleaning to remove or correct redundant and malformed records; data normalization; and missing value filling.

[0058] (1.2) Multi-factor identification and classification:

[0059] After data preprocessing is completed, the multidimensional attribute factors in the data are divided into three categories according to the changing characteristics of geographic elements in time series or their inherent properties:

[0060] Dynamic factors: refers to geographical elements whose attribute values ​​change significantly over time, such as temperature, precipitation, air humidity, surface temperature, wind speed, etc. These factors usually have obvious time dependence and volatility.

[0061] Static factors refer to geographical elements whose attribute values ​​are relatively stable, change slowly or do not change within the research time scale, such as terrain elevation, surface slope, slope aspect, soil type, average annual precipitation, etc. These factors mainly reflect spatial characteristics.

[0062] Categorical factors: refer to geographic features whose attribute values ​​are discrete categories, such as land use types (such as forests, cultivated land, urban areas), vegetation cover types, administrative divisions, etc. These factors describe the categorical attributes of geographic entities.

[0063] (1.3) Construction and sampling of space-time grids: Construct a regular space-time reference grid covering the study area:

[0064] First, according to the specific application requirements and the scope of the study area, the appropriate spatial resolution and temporal granularity (e.g., daily, monthly) are determined. Based on the determined spatiotemporal resolution, a two-dimensional or three-dimensional spatiotemporal grid is generated, which contains a spatial location index and a time step index.

[0065] Then, the dynamic, static, and categorical factor data divided in (1.2) are gridded. For data located in the same grid cell, aggregation or resampling is performed to ensure that all factor data are aligned in the same spatiotemporal reference frame to form a structured input tensor.

[0066] (2) Multi-scale temporal information encoding.

[0067] (2.1) Multi-scale time coding uses sine and cosine periodic coding methods, which are optimized for continuity preservation, periodic topological mapping, and multi-dimensional periodic feature fusion.

[0068] Specifically, for a single input time tensor , assuming the time period length is , then the obtained two-dimensional encoding vector for:

[0069] ;

[0070] in, 、 They are time tensors Cosine period encoding and sine period encoding.

[0071] (2.2) For multi-scale periodic feature fusion, a hierarchical processing strategy is adopted:

[0072] ;

[0073] in, is the maximum index value of the year embedding, is the actual year value, is the element-wise multiplication operator, It is an embedding layer in a neural network that maps discrete indices to dense vectors; represents the cosine-sine encoding applied to the contained multiscale time variables, They are the time granularity units such as month, day, hour, minute, etc. Refers to the tensor connection operation, They are respectively the embedded representation of the year and the periodic encoding of other time units; The dimension of the final dimension of the tensor, represents the RMS normalization layer, Represents the tensor Mapping to dimensions using the Kolmogorov Arnold Network (KAN) A function of a continuous vector space.

[0074] (3) Adaptive spatial information coding, which converts the two-dimensional spatial location information of geographic data into a continuous vector representation that can be processed by the model, and enables it to adaptively generate codes according to the specific spatial location.

[0075] Specifically, the adaptive position encoder will connect the original coordinate vector , and then apply the sequential KAN layer to embed the spatial information into a continuous vector middle:

[0076] ;

[0077] in, It is a multi-layer nonlinear mapping function based on the Kolmogorov-Arnold network, used to encode spatial information Mapping to dimensions A continuous vector space.

[0078] (4) Selective coding module design and feature extraction, which is used to extract geographical factor features efficiently and selectively, especially to capture long-range dependencies.

[0079] (4.1) Input feature projection and spatial expansion:

[0080] Aims to increase the diversity of features and prepare for subsequent selective scanning, by inputting geographic factors or fusion feature tensors Through the nonlinear transformation of the KAN projection layer, the feature space dimension is expanded to , and get the tensor ,in, They are batch size, spatial sequence length, and input feature dimension.

[0081] (4.2) Feature dimension reshaping and projection:

[0082] The tensor Reshape it and pass it through the KAN projection layer again to map it to the hidden dimension , get a tensor suitable for sequence scanning .

[0083] (4.3) Selective sequence scanning based on the selection state space model Mamba:

[0084] The tensor The input is fed into a stacked structure composed of Mamba components. Mamba components select the model along the embedding space dimension based on the state space model selection mechanism. Perform efficient one-way scanning processing to capture long-range dependencies.

[0085] In this embodiment, residual connections are applied between Mamba components to enhance the stability of gradient propagation and model training. The process is expressed as follows:

[0086] ;

[0087] in, represents the residual connection, stands for root mean square normalization, represents a feedforward network, express The number of layers, express The number of stacking times, It is a sequence processing component based on the selective state space model (SSM), a feature extraction unit that efficiently captures long-range dependencies through structured state space representation and selective scanning mechanism.

[0088] (4.4) Enhanced feature generation based on the “scan-select-activate” mechanism:

[0089] The tensor processed by the Mamb component pass The normalization layer and KAN projection layer perform dimension compression and nonlinear activation to obtain the final enhanced feature representation .

[0090] (5) The residual gated fusion mechanism is integrated with multi-source factors to achieve the selective fusion of different types of geographic factors and spatiotemporal coding.

[0091] (5.1) Residual gated fusion is based on a gate-like mechanism:

[0092] Using a feedforward network Precisely control the flow of feature information, given the input features and , the output features are generated by the residual gate fusion module for:

[0093] ;

[0094] in, is the learnable weight parameter matrix of the i-th KAN linear layer, is the Gaussian error linear unit activation function, which provides smooth nonlinear transformation.

[0095] (5.2) Dynamic factor and classification factor selection fusion:

[0096] The dynamic geographic factor obtained in step (1) Feature representation and categorical geographical factors The feature representation of is taken as input and input into the cascade structure of the first-level residual gated fusion and selective encoder modules to realize the interaction of the two factor features and residual fusion and selective feature selection and enhancement based on the state space model to generate a fusion representation of dynamic and categorical geographic factors. .

[0097] (5.3) Geographic factor fusion representation and time coding fusion:

[0098] The fusion representation of the dynamic and categorical geographical factors generated in step (5.2) The time-coded tensor generated by the time scale encoder obtained in step (2) As input, it is input into the cascade structure of the second-level residual gated fusion and selective encoder modules, and residual gated fusion and selective processing are performed to generate a representation that integrates time, dynamics, and category geographical factors. .

[0099] (5.4) Fusion representation and static factor fusion:

[0100] The fusion representation generated in step (5.3) and the static geographic factor obtained in step (1) The feature representation is taken as input and input into the cascade structure of the third-level residual gate fusion and selective encoder modules for fusion processing to generate a representation that integrates time and dynamic, category, and static geographic factors. .

[0101] (5.5) Final geographic spatiotemporal fusion representation generation:

[0102] The fusion representation generated in step (5.4) The position encoding vector generated by the adaptive spatial information encoding in step (3) As input, it is input into the fourth-level cascade structure for final residual fusion and selective processing to generate a unified geographic spatiotemporal fusion representation containing all geographic factors and spatiotemporal information. .

[0103] (6) Decoding module design and spatiotemporal prediction output.

[0104] Specifically, receive the unified geographic spatiotemporal fusion representation generated after step (5.5) It takes as input and generates final output results suitable for specific downstream geoscience applications based on this input.

[0105] (6.1) Encoder output dimensionality reduction:

[0106] The unified geographic spatiotemporal fusion representation obtained in step (5.5) As input, it undergoes nonlinear transformation and feature selection processing through a combination of multiple Kolmogorov-Arnold network layers and selective encoder modules to reduce the feature dimension and generate a sequence tensor suitable for downstream tasks. .

[0107] (6.2) Sequence modeling based on state space Mamba:

[0108] The sequence tensor generated in step (6.1) Input is a stack of multiple Mamba components, each of which has a state space model and a content-dependent selection mechanism along the sequence dimension. Model the proximity and dependencies between input data points to further extract and refine features .

[0109] (6.3) Final output projection:

[0110] The sequence tensor processed by step (6.2) The final nonlinear projection and dimension transformation is performed through one or more KAN layers. , which maps the feature space to the dimension of the target output space.

[0111] (6.4) Output generation:

[0112] The results of step (6.3) are projected according to the needs of specific downstream earth science tasks. Perform parsing and post-processing, relying on various loss functions and step (7) end-to-end back-propagation training, and finally generate output results in the form of final prediction values, category results, spatial distribution maps or simulated data.

[0113] In particular, the geographic multi-factor classification and fusion modeling method based on the state space model in this embodiment also includes: step (7) end-to-end training and parameter optimization.

[0114] Specifically, the process from the input data of step (1) to the final output of step (6) is constructed as an end-to-end differentiable computation graph, and the model parameters (including but not limited to KAN layer parameters, Embedding layer parameters, Mamba component parameters, and RMSNorm parameters, etc.) are optimized through the standard deep learning training process.

[0115] First, choose a suitable loss function, which measures the difference between the model output and the true target value; for example, the mean squared error (MSE) can be used for regression tasks, and the cross-entropy loss (Cross-Entropy Loss) can be used for classification tasks.

[0116] Then, use an optimizer (such as Adam, AdamW, etc.) to update the model parameters.

[0117] During training, batches of labeled geographic spatiotemporal data are fed into the model. The loss between the model output and the true target is calculated, and the gradient of the loss function with respect to all model parameters is calculated using automatic differentiation techniques. Finally, the optimizer updates the model parameters based on the gradient to minimize the loss function. This process is iteratively executed on a high-performance computing platform (for example, leveraging the parallel computing power of GPUs or TPUs) until the model converges on the validation set or achieves high-precision modeling.

[0118] In summary, the present invention provides a method for spatiotemporal modeling of geographic big data based on multi-factor partitioning and fusion of state-space models. It effectively partitions and fuses dynamic, static, and categorized geographic factor data, while taking into account feature extraction in both temporal and spatial dimensions. This method achieves differentiated processing and effective fusion of different types of geographic factors, effectively resolving the multi-source heterogeneity, spatiotemporal dependency, multi-scale characteristics, and high-dimensional complexity issues in geographic spatiotemporal data modeling. Furthermore, it can effectively capture the spatiotemporal characteristics of various geographic elements, significantly improving the accuracy of geographic data modeling while maintaining high computational efficiency. It is universally applicable and adaptive to the comprehensive multi-factor modeling of geographic data.

[0119] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A geographic multi-factor classification and fusion modeling method based on a state space model, characterized by: The steps include: Step 1: Geographic data sampling and factor division: Collect multi-source geographic spatiotemporal data and perform preprocessing, identify and divide geographic data into multiple factors based on temporal variation characteristics, construct a spatiotemporal reference grid, and perform gridding processing on each factor; Step 2: Multi-scale temporal information encoding: Using optimized sine and cosine periodic encoding methods, the continuity of geographic data temporal information is preserved and periodic topological mapping is performed, and multi-scale periodic features are integrated through a hierarchical processing strategy. Step 3: Adaptive spatial information encoding: mapping geographic data spatial coordinates to continuous vector space; Step 4: Selective encoder construction and feature extraction: The input factors are nonlinearly transformed and spatially expanded through the projection layer, and selective sequence scanning and long-range dependency capture are performed through the state space model; Step 5: Residual gating and multi-source factor fusion: The residual gating fusion module is used to selectively fuse multi-source factors, and different types of geographic factors and spatiotemporal information are gradually integrated through a cascade structure to generate a unified geographic spatiotemporal fusion representation; Step 6. Decoder module construction and spatiotemporal prediction output: Construct a decoding module, take the unified geographic spatiotemporal fusion representation as input, perform dimensionality reduction processing through the combined structure of the KAN layer and the selective encoder module, use the Mamba component stacking structure to model the sequence tensor, capture the dependencies between data points, and map the features to the target output space, perform result analysis and post-processing, and generate the final prediction value, category result, spatial distribution map or simulated data.

2. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 1 is characterized in that: The geographic data sampling and factor division described in step 1 includes the following sub-steps: Step 1.

1. Collection and preprocessing of original geographic spatiotemporal data: Collect multi-source original geographic data containing time, spatial location, and multidimensional attribute information, and perform data cleaning, normalization, and missing value filling. Step 1.2, multi-factor identification and classification: According to the spatiotemporal variation characteristics of geographic elements, the multi-dimensional attribute factors in the original geographic data are divided into three categories: dynamic geographic factors , static geographical factors , Category Geographic Factor ; Step 1.3, space-time grid construction and sampling: Construct a unified space-time reference coordinate system, determine the space-time resolution according to the study area and application requirements, including spatial resolution and temporal granularity, and generate a regular grid. Grid the three types of factor data and resample them to align the data in the same space-time reference frame to form a structured input tensor.

3. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 2 is characterized in that: In step 1.2, the dynamic geographical factors refer to meteorological elements that show obvious changes in time series, including temperature, precipitation, and air humidity; The static geographical factors refer to the terrain elements that remain relatively stable within the research time scale, including: terrain elevation, surface slope, and slope direction; The categorical geographic factors refer to categorical data with discrete characteristics, including land use types and vegetation cover types.

4. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 2 is characterized in that: The multi-scale temporal information encoding in step 2 includes the following sub-steps: Step 2.1, using sine and cosine periodic encoding to optimize continuity preservation, periodic topological mapping and multi-dimensional periodic feature fusion, for a single input time tensor , the time period length is , get the two-dimensional encoding vector for: ； in, 、 They are time tensors Cosine period coding and sine period coding; Step 2.2: Fusion of multi-scale periodic features through a hierarchical processing strategy, application of RMS normalization and KAN mapping, and conversion of time variables into continuous vector representations : ； in, is the maximum index value of the year embedding, is the actual year value, is the element-wise multiplication operator, It is an embedding layer in a neural network that maps discrete indices to dense vectors; represents the cosine-sine encoding applied to the contained multiscale time variables, They are month, day, hour, and minute time granularity units respectively; Refers to the tensor connection operation, They are respectively the embedded representation of the year and the periodic encoding of other time units; The dimension of the final dimension of the tensor, represents the RMS normalization layer, Represents the tensor Mapping to dimensions using KAN A function of a continuous vector space.

5. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 4 is characterized in that: Step 3: Adaptive spatial information encoding, connecting the coordinate vector of the original geographic data through the adaptive spatial information encoder , applying a sequential KAN layer to embed spatial information into continuous vectors middle: ； in, It is a multi-layer nonlinear mapping function based on the Kolmogorov-Arnold network, used to encode spatial information Mapping to dimensions A continuous vector space.

6. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 5 is characterized in that: The selective encoder construction and feature extraction described in step 4 includes the following sub-steps: Step 4.1: Input feature projection and spatial expansion: Input geographic factors or fusion feature tensors Through the nonlinear transformation of the KAN projection layer, the feature space dimension is expanded to , and get the tensor , increase feature diversity; among them, They are batch size, spatial sequence length, and input feature dimension; Step 4.2: Reshape and Project the Feature Dimension: Reshape the Tensor , and then passes through the KAN projection layer and is mapped to the hidden dimension , get a tensor suitable for sequence scanning ; Step 4.3, Selective sequence scanning based on the selection state space model Mamba: tensor Input into the stacked structure composed of Mamba, each Mamba component uses the state space model selection mechanism to select along the embedding space dimension Perform one-way scanning processing to capture long-range dependencies; Step 4.4: Enhance feature generation based on the scan-select-activate mechanism: The tensor processed by the Mamba component pass The normalization layer and KAN projection layer perform dimension compression and nonlinear activation to obtain the final enhanced feature representation .

7. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 6 is characterized in that: In step 4.3, in the stacked structure composed of Mamba components, residual connections are applied between the stacked Mamba components to enhance gradient propagation and model training stability, which is expressed as: ； in, represents the residual connection, stands for root mean square normalization, represents a feedforward network, express The number of layers, is a sequence processing component based on a selective state-space model, express Number of stacking times.

8. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 7 is characterized in that: The residual gating and multi-source factor fusion described in step 5 include the following sub-steps: Step 5.1: Residual gating fusion based on class gating mechanism: using feedforward network Control feature information flow, given input features and , the output features are generated by the residual gate fusion module for: ； in, is the learnable weight parameter matrix of the i-th KAN linear layer, It is the Gaussian error linear unit activation function, which provides smooth nonlinear transformation; Step 5.2: Fusion of dynamic factors and categorical factors: Combine the dynamic geographic factors obtained in step 1 Feature representation and categorical geographical factors The feature representation is input into the cascade structure of the first-level residual gate fusion and selective encoder modules to perform interactive fusion of the two factor features and selective feature selection and enhancement to generate a fusion representation of dynamic and categorical geographic factors. ; Step 5.3: Fusion of geographical factors and time coding: Fusion representation The continuous vector representation obtained in step 2 , input to the cascade structure of the second-level residual gated fusion and selective encoder modules, perform residual gated fusion and selective processing, and generate a representation that integrates time, dynamics and category geographical factors ; Step 5.4: Fusion of fusion representation and static factor: Fusion representation The static geographic factor obtained in step 1 The feature representation is input into the cascade structure of the third-level residual gate fusion and selective encoder modules for fusion processing to generate a representation that integrates time and dynamic, classification, and static geographic factors. ; Step 5.5: Generate the final geographic spatiotemporal fusion representation: The position encoding vector generated by the adaptive spatial information encoding in step 3 , input to the fourth level R cascade structure, perform final residual fusion and selective processing, and generate a unified geographic spatiotemporal fusion representation containing all geographic factors and spatiotemporal information , as the input to the decoder module described in step 6.

9. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 8 is characterized in that: Step 6: Decoder module construction and spatiotemporal prediction output, including the following sub-steps: Step 6.1, encoder output dimensionality reduction: the unified geographic spatiotemporal fusion representation obtained in step 5.5 is converted into As input, it undergoes nonlinear transformation and feature selection processing through a combination of multiple Kolmogorov-Arnold network layers and selective encoder modules to reduce the feature dimension and generate a sequence tensor suitable for downstream tasks. ; Step 6.2, Sequence Modeling Based on State Space Mamba: Convert the sequence tensor The input is a stacked structure composed of multiple Mamba components. Each Mamba component uses a state space model and a content-related selection mechanism to select along the sequence dimension. Model the proximity and dependencies between input data points, extract and refine features, and obtain sequence tensors ; Step 6.3, output projection: pass through the sequence tensor Perform nonlinear projection and dimension transformation through one or more KAN layers to obtain the projected result , mapping the feature space to the dimension of the target output space; Step 6.4: Output generation: The projected results are generated based on the specific downstream earth science task requirements. Perform parsing and post-processing, relying on loss functions and end-to-end back-propagation training and parameter optimization to generate the final prediction value, classification result, spatial distribution map or simulation data output.

10. The geographic multi-factor classification and fusion modeling method based on the state space model according to claim 8 is characterized in that: The end-to-end back-propagation training and parameter optimization includes the following sub-steps: Step 7.1: Build an end-to-end differentiable computational graph from data input to final output. Step 7.2: Select a loss function suitable for the task type and apply the optimizer to update the model parameters. Step 7.

3. Calculate the gradient of the loss function with respect to the model parameters through automatic differentiation. Step 7.4: Iterate the training process on the computing platform until the model converges or reaches the preset accuracy requirement.

Images