Land use change prediction method and device based on map representation generation learning, equipment and medium

Abstract

This application discloses a land use change prediction method, apparatus, device, and medium based on map representation generation learning, relating to the fields of spatiotemporal big data mining and urban planning technology. The method includes: constructing multi-source map submaps and performing binarization and attention weighting to obtain binarized land use submaps and weighted geographic environment submap features; extracting deep aggregated spatial features from both to generate latent vector representations; flattening the representation into a one-dimensional feature sequence and querying the index of each element in a landscape codebook to form a vectorized index sequence; combining relative distance encoding, using a map sequence generation learning model to predict future vectorized index sequences, and querying the landscape codebook to obtain future vectorized tensors; iteratively decoding the tensor using a map decoder to generate land use change results. This application achieves more accurate land use change prediction by capturing mesoscale spatial local information and reducing unit-level error accumulation.

Description

Technical Field

[0001] This application relates to the fields of spatiotemporal big data mining and urban planning technology, and in particular to land use change prediction methods, devices, equipment and media based on map representation generation learning. Background Technology

[0002] With the rapid development of remote sensing, geographic information systems (GIS), and artificial intelligence technologies, land use change prediction has gradually shifted from traditional methods relying on static statistical data to intelligent modeling based on spatiotemporal big data. In applications such as land spatial planning, urban renewal, and ecological environmental protection, there is an urgent need for high-precision, dynamic, and fine-grained land use evolution simulation.

[0003] Currently, mainstream land use change prediction methods mainly include simulation models based on cellular automata (CA), Markov chain transition matrix methods, and deep learning models that have emerged in recent years. Deep learning methods often treat land use maps as two-dimensional raster images, employing architectures such as Convolutional Neural Networks (CNN), U-Net, and ConvLSTM to perform end-to-end training on historical land use maps from multiple periods to predict future land use distribution. Some studies further introduce Generative Adversarial Networks (GANs) or autoencoders to enhance the spatial continuity and structural rationality of the generated results.

[0004] However, existing practices generally suffer from insufficient representation of spatial scale. They typically perform independent predictions at the micro-unit level (such as individual pixels or regular grids), neglecting the spatial organization patterns, topological relationships, and semantic context among land use types at the meso-scale (i.e., the scale of city blocks or functional zones). This unit-level modeling approach not only struggles to capture the overall morphological characteristics of local areas but also easily leads to the accumulation of local prediction biases over time, resulting in structural distortions or logical contradictions in long-term predictions. Furthermore, traditional convolution operations are limited by a fixed receptive field, making it difficult to adaptively model the spatial heterogeneity of different regions, especially in complex scenarios such as urban-rural fringe areas and mixed-function zones, where generalization ability is limited. Therefore, how to capture local spatial information at the meso-scale and reduce the accumulation of unit-level errors to achieve more accurate land use change predictions has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a land use change prediction method, device, equipment and medium based on map representation generation learning, which aims to solve the technical problem of how to achieve more accurate land use change prediction by capturing mesoscale spatial local information and reducing the accumulation of unit-level errors.

[0006] To achieve the above objectives, this application proposes a land use change prediction method based on map representation generation learning, the method comprising: Construct a multi-source map sub-map, perform binarization and attention weighting on the multi-source map sub-map, and obtain the binarized land use sub-map and weighted geographic environment sub-map features; The map encoder in the map variational autoencoder performs deep aggregation spatial feature extraction on the features of the binarized land use submap and the weighted geographic environment submap to obtain latent vector representations. The map variational autoencoder also includes a vectorization module and a map decoder. The vectorization module contains a preset number of landscape codebooks. The latent vector representation is flattened by traversing the grid in a clockwise spiral path to obtain a one-dimensional feature sequence. The vectorization module queries the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain the vectorized index sequence. Based on the vectorized index sequence and relative distance encoding, a future vectorized index sequence is generated through a map sequence generation learning model, and the landscape codebook is queried according to the future vectorized index sequence to obtain the future vectorized tensor. The map decoder iteratively decodes the future vectorized tensor to generate land use change results.

[0007] Furthermore, to achieve the above objectives, this application also proposes a land use change prediction device based on map representation generation learning, the device comprising: The sub-map construction module is used to construct multi-source map sub-maps, perform binarization and attention weighting on the multi-source map sub-maps, and obtain binarized land use sub-maps and weighted geographic environment sub-map features; The feature extraction module is used to perform deep aggregation spatial feature extraction on the features of the binarized land use submap and the weighted geographic environment submap through the map encoder in the map variational autoencoder to obtain latent vector representations. The map variational autoencoder also includes a vectorization module and a map decoder. The vectorization module contains a preset number of landscape codebooks. The spiral flattening module is used to flatten the latent vector representation by traversing the grid in a clockwise spiral path to obtain a one-dimensional feature sequence. The vectorization module is used to query the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain a vectorized index sequence. The sequence generation module is used to generate a future vectorized index sequence based on the vectorized index sequence and the relative distance encoding through a map sequence generation learning model, and to query the landscape codebook based on the future vectorized index sequence to obtain a future vectorized tensor. The iterative decoding module is used to iteratively decode the future vectorized tensor through the map decoder to generate land use change results.

[0008] Furthermore, to achieve the above objectives, this application also proposes a land use change prediction device based on map representation generation learning. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the land use change prediction method based on map representation generation learning as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the land use change prediction method based on map representation generation learning as described above.

[0010] One or more technical solutions proposed in this application have at least the following technical effects: First, a multi-source map submap is constructed, and binarization and attention weighting are applied to obtain binarized land use submaps and weighted geographic environment submaps, effectively separating target land use types and highlighting the influence of key environmental factors. Second, the map encoder in the map variational autoencoder performs deep aggregation spatial feature extraction on the two types of submaps, generating compact latent vector representations, thereby integrating high-order semantic information of land use patterns and environmental driving factors. Next, this latent vector representation is flattened by traversing the raster along a clockwise spiral path, forming a one-dimensional feature sequence that preserves local spatial proximity relationships while adapting to the sequence modeling input format. Then, the vectorization module queries the landscape codebook for the index corresponding to each element of this sequence, obtaining a vectorized index sequence, realizing the conversion from continuous features to discrete structured representations. Subsequently, based on this index sequence and relative distance encoding, a map sequence generation learning model is used to generate future vectorized index sequences, and the landscape codebook is queried accordingly to obtain future vectorized tensors, enabling the model to learn the spatiotemporal dependence of land use evolution at the mesoscale. Finally, the future vectorized tensor is iteratively decoded using a map decoder to reconstruct land use change results with reasonable spatial structure and semantic consistency. This application achieves more accurate land use change prediction by capturing mesoscale spatial local information and reducing unit-level error accumulation. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an embodiment of the land use change prediction method based on map representation generation and learning in this application. Figure 2 This is a schematic diagram of the architecture of MapsVAE provided in Implementation Example 1 of the land use change prediction method based on map representation generation learning in this application; Figure 3 This is a schematic diagram of the architecture of EMTrans provided in Embodiment 1 of the land use change prediction method based on map representation generation learning in this application; Figure 4 This is a flowchart illustrating Embodiment 2 of the land use change prediction method based on map representation generation and learning in this application. Figure 5 A simplified flowchart illustrating the land use change prediction method based on map representation generation learning provided in Embodiment 2 of this application; Figure 6 This is a schematic diagram of the module structure of the land use change prediction device based on map representation generation learning, as described in an embodiment of this application. Figure 7 This is a schematic diagram of the hardware operating environment of the land use change prediction method based on map representation generation learning in the embodiments of this application.

[0014] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0016] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0017] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or land use change prediction system capable of the above functions. The following description uses a land use change prediction system as an example to illustrate this embodiment and the subsequent embodiments.

[0018] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0019] Based on this, embodiments of this application provide a land use change prediction method based on map representation generation learning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the land use change prediction method based on map representation generation learning in this application.

[0020] In this embodiment, the land use change prediction method based on map representation generation learning includes steps S10~S60: Step S10: Construct a multi-source map sub-map, perform binarization and attention weighting on the multi-source map sub-map, and obtain the binarized land use sub-map and weighted geographic environment sub-map features; It should be noted that multi-source map submaps refer to sub-regional maps covering the same local area extracted from geospatial data of different sources or types. These include, but are not limited to, land use maps, road network maps, POI (Point of Interest) distribution maps, elevation models, and nighttime light imagery, used to jointly represent the comprehensive geographic environmental characteristics of the area. Weighted geographic environmental submap features refer to the fused feature representation obtained by assigning different weights to each submap based on its importance to the current prediction task after processing the multi-source map submaps using an attention mechanism. This can highlight environmental factors that have a stronger explanatory or driving effect on land use change, thereby enhancing the model's ability to understand the local spatial context.

[0021] As an example, the steps of constructing multi-source map sub-maps, performing binarization and attention weighting on the multi-source map sub-maps to obtain binarized land use sub-maps and weighted geographic environment sub-map features include: constructing a multi-source map dataset containing land use and various geographic environments; performing block-based and normalization processing on the multi-source map dataset to obtain multi-source map sub-maps, wherein the multi-source map sub-maps include land use sub-maps and geographic environment sub-maps; converting the various land use semantic labels of the land use sub-maps into raster maps with multiple binarized channels to obtain binarized land use sub-maps; and performing attention weighting on the geographic environment sub-maps to obtain weighted geographic environment sub-map features.

[0022] As an example, the step of performing attention weighting on the geographic environment subgraph to obtain weighted geographic environment subgraph features includes: compressing the features of each environmental variable of the geographic environment subgraph through global average pooling to obtain the feature vectors corresponding to each environmental variable; processing the feature vectors through a one-dimensional convolutional layer with a preset number of channels to obtain the local interaction patterns between the feature vectors; applying Sigmoid activation to the local interaction patterns to generate attention weights corresponding to each environmental variable; performing pointwise multiplication of the attention weights with the original feature tensor of the geographic environment subgraph, and standardizing the resulting feature tensor to obtain the weighted geographic environment subgraph features.

[0023] A multi-source map dataset refers to a collection of geospatial data of various types, including land use maps and geographic environment layers reflecting different dimensions such as topography, transportation, population, economy, and ecology (e.g., road density maps, POI distribution maps, elevation maps, nighttime light index maps, etc.), used to comprehensively characterize the spatial features of the study area. A land use submap is a raster image of a local area extracted from a land use map, with each pixel carrying a corresponding land use category identifier, used to represent the spatial distribution of various land uses within that local area. A geographic environment submap is a collection of raster images of sub-regions extracted from various geographic environment layers corresponding to the same local area. Each layer represents an environmental variable (e.g., slope, distance to subway station, building density, etc.), collectively forming a multi-channel input describing the external driving factors of the area. Land use semantic labels refer to the category name or code corresponding to each pixel in the land use submap (e.g., cultivated land, forest land, residential land, industrial land, etc.), used to express the land function attributes of that location.

[0024] A binarized raster image refers to a single-channel raster image containing only 0s and 1s for a specific land use category, where 1 indicates belonging to that category and 0 indicates not belonging. If there are N land use categories, N such binarized channels can be generated, forming a multi-channel binarized representation. The feature vector corresponding to each environmental variable is a one-dimensional vector obtained by compressing the spatial dimension of each environmental variable channel (i.e., each type of environmental factor) in the geographic environment submap through a global average pooling operation. It is used to characterize the overall statistical characteristics of that environmental variable in the current submap region. The preset number of channels refers to the number of output channels manually set in the one-dimensional convolutional layer, used to control the representational power and model complexity when the convolutional operation transforms the feature vectors. The local interaction mode refers to the intermediate feature representation output after the one-dimensional convolutional layer models the dependencies between adjacent or related channels in the feature vectors of each environmental variable, reflecting the synergistic or competitive effects of different environmental factors in local combinations. Attention weights are scalars or vectors with values ​​between 0 and 1, generated after processing local interaction patterns using the Sigmoid activation function. They are used to quantify the importance of each environmental variable to the current prediction task. The original feature tensor refers to the multi-channel 3D data structure (height × width × number of channels) of the geographic submap before weighting, where each channel corresponds to raster data of one environmental variable, preserving complete spatial detail information.

[0025] (1) Construct a multi-source map dataset of land use and various geographic environments to adapt to the training and learning of geographic deep learning models. The multi-source map is divided into D×D grid sub-maps using a map partitioning strategy to avoid the inherent challenge of deep learning intelligent models learning large-size images. At the same time, data augmentation is used to increase the sample size and further ensure the model's learning effect. Then, the land use map and various geographic environment maps of the study area are partitioned to form the training and validation dataset of the map variational autoencoder. For the land use map prediction task of a large study area, sub-maps can be extracted sequentially based on the grid geospatial location and the size of the sub-map block D for prediction. Then, the predicted future geographic sub-maps are stitched together according to their spatial location to form a complete map of the study area. Before partitioning the geographic environment map into sub-maps, a global normalization calculation needs to be performed to ensure that all geographic environment attribute values ​​in the sub-map are in the range of [0, 1].

[0026] (2) The land use map containing M types of land use semantic labels is transformed into a raster map with M binary channels. Each channel uses a 0-1 matrix to specifically represent the membership of a certain land use semantic type. This spatial one-hot encoding transformation effectively solves the convolution calculation problem of discrete labels, while preserving the mutual exclusion or mixing between different land use semantic types.

[0027] (3) First, a geographic environment submap attention mechanism is designed. By simulating the decision-making preference process of human experts, higher weights are pre-assigned to geographic features with strong explanatory power of spatial structure from dozens or even hundreds of environmental variables. This mechanism includes two core links: geographic environment feature compression and feature interaction learning. The geographic environment feature compression layer extracts the essential features of environmental variables through multi-level features, and the feature interaction layer is used to capture the nonlinear correlation between multi-variable compressed features.

[0028] The geographic environment submap GM is a multi-channel cubic tensor representing the normalized values ​​of high-dimensional geographic features within the same region. First, global average pooling (GAP) is used to compress the features of each geographic environment variable, compressing the three-dimensional tensor GM (D×D×C) into a C-dimensional feature vector (1×1×C). Global average pooling compresses the global statistical features of each environment variable, avoiding spatial detail interference and significantly reducing computational complexity. Then, a one-dimensional convolutional layer (1×1×k) with k-channel filters is used to capture local interaction patterns between compressed features, learning nonlinear relationships between variables to measure the uniqueness of feature information. Based on this, the attention weights w of the geographic environment submap are generated using a sigmoid activation function: In the formula, This indicates a global average pooling operation. This represents a one-dimensional convolution operation with a kernel size of k. Apart from the original features of the input geographic submap, all variables in the formula are learnable parameters. Finally, the pre-computed weighted geographic submap features are obtained by pointwise multiplying the attention weight w with the original feature tensor GM. : Step S20: The map encoder in the map variational autoencoder performs deep aggregation spatial feature extraction on the features of the binarized land use submap and the weighted geographic environment submap to obtain latent vector representation. The map variational autoencoder also includes a vectorization module and a map decoder. The vectorization module contains a preset number of landscape codebooks. It's important to note that MapsVAE (Maps Variational Autoencoder) is a generative deep learning model specifically designed for map data. Based on the Variational Autoencoder (VAE) framework, it learns low-dimensional probability distribution representations from input map submaps and supports high-fidelity, structurally sound land use pattern reconstruction and generation. The map encoder, the feedforward neural network part of MapsVAE, receives binarized land use submaps and weighted geographic environment submap features as input. It extracts and aggregates deep spatial features step-by-step through multiple convolutional or attention mechanisms, ultimately outputting a compact latent representation describing the essential attributes of the map content. The vectorization module, located after the map encoder, transforms the high-dimensional feature map output by the encoder into latent spatial vectors that meet the requirements of the variational autoencoder. It typically includes flattening operations and fully connected layers, generating the mean and variance parameters of the latent distribution during this process to support subsequent random sampling.

[0029] The map decoder refers to the generative network part of MapsVAE. It receives latent vectors sampled from the latent space as input and gradually reconstructs a land use prediction map with the same size as the original input through deconvolution or upsampling operations, aiming to restore a reasonable spatial layout and semantic structure. The latent vector representation refers to the low-dimensional continuous vectors obtained after the map encoder is processed by the vectorization module. It implicitly encodes the core spatial patterns and evolutionary laws of the land use submap and its environmental context in the form of a probability distribution, serving as an intermediate semantic bridge connecting the encoding and decoding processes. The preset quantity refers to the number of discrete vectors included in the landscape codebook, manually specified during the model design phase, used to control the discretization granularity and representation capacity of the latent space. The landscape codebook is a set of learnable discrete representation vectors, each representing a typical land use pattern or spatial pattern (such as "high-density residential area," "industrial-green space mixed area," etc.), used to map continuous latent representations to semantically explicit discrete codes.

[0030] As an example, the construction steps of the map variational autoencoder include: constructing a map encoder based on a depthwise separable convolutional layer, a residual connection layer, a group normalization layer, and an ELU activation function; constructing a vectorization module based on a preset number of landscape codebooks and a vector mapping rule for nearest neighbor search, wherein the landscape codebook consists of multiple discrete representation vectors; constructing a map decoder based on the depthwise separable convolutional layer, the residual connection layer, the group normalization layer, the ELU activation function, a two-dimensional convolutional layer, and a softmax layer; and constructing a map variational autoencoder based on the map encoder, the vectorization module, and the map decoder.

[0031] Depthwise separable convolutional layers are a lightweight convolutional operation that decomposes standard convolution into channel-wise spatial convolution and cross-channel pointwise convolution, reducing the number of parameters and computational overhead while retaining strong spatial feature extraction capabilities. Residual connection layers introduce skip connections into the network, directly adding input features to the output of a layer to alleviate the vanishing gradient problem in deep network training and promote efficient information flow between different layers. Group normalization (GN) layers are a technique that normalizes feature channels within each group, independent of batch size, suitable for training small batches or high-resolution map data, and helps improve model stability and generalization ability. The ELU activation function, or Exponential Linear Unit activation function, has a non-zero output and is smoothly differentiable in the negative region, while remaining linear in the positive region, helping to accelerate convergence and improve model expressiveness.

[0032] The nearest neighbor search vector mapping rule refers to a mapping strategy that, during vectorization, calculates the distance (e.g., Euclidean distance) between the continuous latent vectors output by the encoder and all discrete representation vectors in the landscape codebook, selecting the closest codebook vector as its discrete counterpart. Discrete representation vectors refer to vectors of each fixed dimension in the landscape codebook, used to encode specific types of land use spatial structures in a discrete and interpretable manner. They are key elements for the model to achieve structured generation and semantic control. A two-dimensional convolutional layer is a standard convolutional layer that performs sliding window convolution operations on the input feature map in both height and width spatial dimensions. It is used to extract local spatial patterns and is often used in decoders to progressively recover spatial details of the map. A softmax layer is an activation layer located at the end of the network. It normalizes the class logits at each spatial location, outputting the probability distribution of each land use class, ensuring that the prediction results meet probability constraints and are convenient for subsequent classification or sampling.

[0033] (1) Map variational autoencoder construction: This embodiment proposes a "map-vectorization-map" geographic deep learning framework for geographic raster maps, enabling adaptive learning of deep aggregated spatial features of arbitrary geographic raster maps while ensuring that the raster maps are decodeable and mappable. The architecture comprises three core modules: a map encoder, a vectorization module, and a map decoder.

[0034] (2) Map Encoder (MapE): MapE integrates Depthwise Separable Convolutional Layers (DSConv) with residual connections. DSConv, the fundamental unit of MapE, consists of two key components: first, independent convolution operations are applied to each input channel to achieve channel-wise depthwise convolution; then, pointwise convolution operations are used to achieve cross-channel feature interaction learning. Building upon DSConv, a Depthwise Separable Residual (DSRes) module is designed to further enhance feature learning capabilities by introducing residual connections after DSConv. Each DSRes module contains two cascaded DSConv layers, each followed by GroupNorm and ELU activation functions.

[0035] MapE first extracts shallow features through an initial DSConv layer, then gradually enhances the feature aggregation range and depth through two deep DSRes modules, finally compressing the input pre-computed multi-channel raster map into a compact latent vector representation. Taking the land use submap LM as an example, the latent vector representation is as follows: Similarly, the encoding process for the geographic environment submap GM can be represented as follows: In the formula, The tensor representing the geographic subgraph in its embedding space (the latent vector representation of the geographic subgraph).

[0036] (3) Vectorization module: Vectorization methods transform the embedding space of a continuous surface into a representative vector representation of a finite number of points. Taking the land use submap LM as an example, the land use landscape codebook is first defined as a set of K discrete representation vectors, used to represent the deep embedding space distribution of the land use submap extracted by MapE. Here, K represents the number of representative vectors in the land use landscape representation learning, i.e., the potential maximum number of land use landscapes; n is the length of the deep encoding vectors learned by MapE.

[0037] Given the encoding tensor of any land use subgraph its i-th element The vectorization process follows the nearest neighbor search principle: In the formula, Landscape codebook representing land use Zhongyu The representative embedding vector with the closest Euclidean distance; This indicates that for the i-th potential vector The result of the vectorization operation is the discrete representation vector corresponding to the index obtained from the landscape codebook corresponding to land use. This indicates the search for index j that minimizes the Euclidean distance; Let k represent the discrete representation vectors of all candidates in the landscape codebook, and k traverses the index range of the entire codebook. Based on this, the entire coding tensor of any land use submap... The vectorized result can be represented as a d×d dimensional tensor. : The same vectorization calculation process is applied to the geographic environment submap GM: In the formula, This represents the vectorized tensor corresponding to the geographic environment submap. For geographical environment landscape codebook middle The nearest neighbor vector; This indicates that for the i-th potential vector The result of the vectorization operation is the discrete representation vector corresponding to the index obtained from the landscape codebook corresponding to the geographic environment. This represents the j-th discrete representation vector in the geographical environment landscape codebook; This represents the discrete representation vector of all candidates in the geographic environment codebook.

[0038] (4) Map Decoder (MapD): When the vectorization module embeds the representation tensor of the continuous embedding space Transformed into a discrete space landscape tensor V( After that, MapD's task is to use a network structure similar to MapE, based on V( MapD reconstructs the original land use submap LM. The difference is that MapD's final layer uses standard 2D convolution and softmax layers instead of MapE's group normalization layers. For the land use submap LM, MapD outputs a multi-channel logit raster map, where each channel value represents the probability of different land use semantic types. In the formula, and MapD represents the vectorized tensor V based on the land use subgraph. The characteristics of stepwise extraction and These represent a two-dimensional convolutional layer and a reconstructed probability map of land use semantic types, respectively.

[0039] Please refer to Figure 2 , Figure 2 This is a schematic diagram of the MapsVAE architecture provided in Embodiment 1 of the land use change prediction method based on map representation generation learning in this application. The left side represents the input, showing the land use sub-map and geographic environment sub-map extracted from multi-source map data. The land use sub-map contains multiple semantic labels, while the geographic environment sub-map consists of multiple environmental variables. These two types of sub-maps together constitute the input to the multi-sub-map construction module. Subsequently, the system preprocesses these sub-maps: the land use sub-map is binarized to generate a multi-channel binarized raster map, while the geographic environment sub-map is enhanced with an attention weighting mechanism to improve the expressive power of key environmental factors, forming weighted geographic environment sub-map features. The preprocessed sub-maps are then fed into two structurally identical but parameter-independent sub-variable autoencoder branches—namely, the "land use sub-map encoder" and the "geographic environment sub-map encoder." Each encoder contains a depthwise separable convolutional layer, a residual connection layer, a group normalization layer, and an ELU activation function to extract deep spatial features and output latent vector representations. The latent vector then enters the vectorization module, where nearest neighbor search is performed using the landscape codebook to map the continuous latent representation into a discrete index sequence. Finally, the decoder receives this discrete index sequence and reconstructs the original map structure step by step through reverse operations, achieving high-fidelity reconstruction of the historical map and thus completing the entire training and representation learning process of MapsVAE.

[0040] As an example, the map variational autoencoder includes two sub-variable autoencoders constructed from the land use submap and the geographic environment submap, respectively. The two sub-variable autoencoders have the same structure and their parameters are independent. The training steps of the map variational autoencoder include: constructing a total loss function based on reconstruction loss, vectorization loss, and commitment loss; constructing a training dataset based on multiple land use submap training samples and a corresponding number of geographic environment submap training samples, wherein the land use submap training samples and the geographic environment submap training samples are raster submaps with a preset tile size; using a preset optimizer and a pass-through gradient estimator, iteratively training the map variational autoencoder based on the training dataset; in each iteration, calculating the loss function value based on the total loss function, and adjusting the model parameters inversely based on the loss function value; stopping training when the number of iterations reaches a preset iteration threshold or the loss function value converges to less than or equal to the preset loss threshold, thus obtaining the trained map variational autoencoder.

[0041] Sub-variable autoencoders (SAEs) refer to two structurally identical but parameter-distributed SAE branches within a map SAE, independently constructed for the land use submap and the geographic environment submap, respectively. These are denoted as LMapVAE (for land use) and GMapVAE (for geographic environment), each performing map feature extraction and reconstruction for its specific input type. Reconstruction loss measures the difference between the model output (i.e., the map reconstructed by the decoder) and the original input map, typically calculated using pixel-level cross-entropy or mean squared error (MSE), and is used to drive the model's ability to accurately reconstruct the input land use pattern. Vectorization loss, introduced in the vectorization module, aims to bring the encoder's output latent vectors closer to the discrete representation vectors in the landscape codebook. It is usually defined as the mean squared error between the codebook vectors and the encoder output, ensuring the effectiveness of the discretization process. Commitment loss is a regularization term set to prevent the encoder output from deviating excessively from the codebook space. Its function is to constrain the encoder to "commit" to using vectors from the codebook for representation, typically achieved by freezing the codebook gradient during backpropagation and applying this loss only to the encoder. Land use submap training samples refer to local raster images cropped from historical land use maps at fixed regions, used for model training. Each pixel carries a clear land use category label and serves as input to LMapVAE. Geographic environment submap training samples refer to local raster images cropped from various geographic environment layers (such as road density, POI, elevation, etc.) that are spatially aligned with the land use submap. These serve as input to GMapVAE to provide external driving factor information.

[0042] Preset chunk size refers to the submap cropping size manually set during the data preprocessing stage. This standardizes the spatial scale of the input data, facilitating batch training and model generalization. A raster submap is a local two-dimensional array of map fragments extracted from continuous geospatial data after dividing it into regular grids. Each element corresponds to a spatial location attribute value, serving as the basic unit for deep learning models to process map data. A pre-selected optimizer is an algorithm chosen beforehand during training to update model parameters. Here, it specifically refers to the Adam optimizer (Adaptive Moment Estimation), which combines momentum and adaptive learning rate mechanisms to accelerate convergence. A pass-through gradient estimator is a gradient approximation strategy used in the vectorization module to handle non-differentiable discrete mappings. It performs nearest neighbor replacement in forward propagation and directly "passes through" the gradient to the encoder output in backpropagation, achieving end-to-end training. The loss function value is the specific numerical value of the total loss function (a weighted sum of reconstruction loss, vectorization loss, and commitment loss) calculated based on the current model parameters in each training iteration. It is used to evaluate model performance and guide parameter updates. The number of iterations refers to the number of parameter update rounds the model has completed during the entire training process. Each round uses a batch of training data for forward computation and backpropagation. The preset iteration threshold is the maximum number of training rounds set before training begins. When the actual number of iterations reaches this value, training will terminate regardless of whether the loss has converged. The preset loss threshold is a manually set criterion for determining loss convergence. If the loss function value is lower than or stabilizes at this threshold for several consecutive rounds, the model is considered to have converged sufficiently, and training can be stopped early.

[0043] MapVAE learning objective constraint construction: For the land use submap and the geographic environment submap, separate MapsVAE models with identical architectures are constructed, named LMapVAE and GMapVAE respectively. The learnable parameters of each model are trained independently. This "isomorphic but different parameter" model design ensures both the consistency of the landscape learning hierarchy and the specific learning of land use and integrated geographic environment landscape.

[0044] For the land use submap variational autoencoder LMapVAE, its training process is subject to three constraints: reconstruction loss Vectorization loss and promised losses Among them, reconstruction losses Ensure the accuracy of land use submap reconstruction; vectorization loss Promote the embedding spatial distribution of landscape codebooks to approximate real subgraph samples; commitment loss Ensure the stability of the MapE learning process. Total loss function. for: In the formula, sg[ ] represents the gradient truncation operation, β is the weight hyperparameter of the committed loss, and || || represents the norm, the subscript 2 indicates the L2 norm, and the superscript 2 indicates the square. To address the non-differentiable nature of the vectorization process, a pass-through gradient estimator is used to directly copy the decoder gradient to the encoder, approximating gradient backpropagation to complete end-to-end model training.

[0045] In the training of the geographic subgraph variational autoencoder GMapVAE, only reconstruction loss is used. different: During training, land use sub-map training samples of preset tile sizes from the training dataset are first input into LMapVAE, and geographic environment sub-map training samples are input into GMapVAE. Based on their respective input samples, the two sub-variable autoencoders perform feature extraction, vectorization transformation, and reconstruction operations through their own map encoders, vectorization modules, and map decoders. Then, their respective loss function values ​​are calculated according to the total loss function (including reconstruction loss, vectorization loss, and commitment loss). To address the non-differentiable nature of the vectorization process, a pass-through gradient estimator is used to approximate gradient backpropagation. A preset optimizer is then applied to the two sub-variable autoencoders. The gradient of the loss function value is calculated, and then the convolutional kernel weights, residual connection parameters, group normalization parameters, landscape codebook vectors in the vectorization module, and 2D convolutional weights in the map decoder are adjusted in reverse. The parameter adjustments of the two sub-variable autoencoders are independent of each other. Then, the above process of "sample input - feature processing - loss calculation - gradient calculation - parameter adjustment" is repeated for iterative training until the number of iterations reaches the preset iteration threshold, or the loss function values ​​of the two sub-variable autoencoders converge to less than or equal to the preset loss threshold. Training is stopped and the trained map variational autoencoder is obtained.

[0046] Step S30: Flatten the latent vector representation by traversing the grid in a clockwise spiral path to obtain a one-dimensional feature sequence; It should be noted that a one-dimensional feature sequence refers to a linear sequence formed by traversing the spatial positions of a latent vector representation (usually a two-dimensional or three-dimensional tensor) point by point from the center or corner along a clockwise spiral path, and arranging the feature values ​​of the positions traversed in sequence. This sequence preserves the relative order information of the local neighborhood in the original latent space, while converting the multi-dimensional structure into a one-dimensional input format suitable for sequence modeling (such as Transformer or RNN).

[0047] Understandably, the MapsVAE module transforms the original semantically sparse land use maps and geographic environment maps into compact latent representations (latent vector representations) through multi-layer spatial convolution computation. It aggregates landscape information based on local micro-scale individual features, and the vectorization operation provides a spatiotemporal sequence for learning and generating future maps.

[0048] This embodiment uses a clockwise spiral flattening (CSF) strategy to flatten the raster spatial data by taking into account the neighborhood structure. Starting from the top left raster, all cells are traversed in a clockwise spiral path until the middle raster, flattening the two-dimensional raster into a one-dimensional sequence.

[0049] First, the land use change prediction system treats the latent vector representation output by the map variational autoencoder as a regular two-dimensional raster matrix and determines the starting point of the spiral traversal (usually the top left corner). Then, the system moves clockwise sequentially: first traversing to the right from the current row to the rightmost column, then traversing down from the current column to the bottom row, then traversing left from the bottom row to the leftmost column, then traversing up from the leftmost column to the previous row, then shrinking the boundary and repeating the above process until all raster positions have been visited once. During the traversal, the system extracts the feature vector corresponding to each position in the order of visit and concatenates them into a continuous sequence. Finally, the system outputs this concatenated result as a one-dimensional feature sequence. The purpose of this is to preserve local spatial proximity and structural context information as much as possible during the flattening process, providing spatially order-aware input for subsequent sequence modeling modules (such as the Transformer).

[0050] Step S40: Using the vectorization module, query the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain the vectorized index sequence; It should be noted that the vectorized index sequence refers to the sequence of index numbers of the discrete representation vectors matched by each element in the one-dimensional feature sequence in the corresponding landscape codebook, where each index value identifies the unique identity of the potential land use or geographic environment pattern at that location in the codebook.

[0051] It is understandable that for the land use sub - map LM of any area in the real world and its corresponding geographical environment sub - map GM given in this embodiment, after being encoded by MapsVAE, the vectorized tensors V( ) and V( ) are obtained respectively. By querying the index corresponding to V( ) in the landscape codebook , it can be converted into a one - dimensional index sequence (i.e., the vectorized index sequence): where, represents that the i - th latent vector is mapped to the discrete representation vector with index in the landscape codebook.

[0052] Step S50: Based on the vectorized index sequence and relative distance encoding, generate a future vectorized index sequence through a map sequence generation learning model, and query the landscape codebook according to the future vectorized index sequence to obtain a future vectorized tensor; It should be noted that relative distance encoding refers to the positional perception encoding constructed by introducing the relative geometric relationship (such as row offset and column offset) in the original two - dimensional grid space for any two positions in a one - dimensional feature sequence during the sequence modeling process, which is used to compensate for the lost spatial structure information after the sequence is flattened, enabling the model to perceive the actual spatial proximity and directionality between elements. The map sequence generation learning model (EMTrans) refers to a generative Transformer architecture designed specifically for map sequences. By fusing the vectorized index sequence and relative distance encoding, it learns the spatio - temporal evolution law of land use patterns in an autoregressive or non - autoregressive manner and outputs the discrete latent representation of the future state. The future vectorized tensor refers to the discrete latent representation of the land use pattern in the future period predicted by the map sequence generation learning model according to the historical vectorized index sequence. Its structure is a two - dimensional tensor with the same size as the original latent vector representation, and the value at each position corresponds to an index in the landscape codebook, which can be restored to continuous features through the codebook look - up table for decoding.

[0053] It is understandable that based on the vectorized index sequences of the land use sub - map and the geographical environment sub - map, the future land use prediction task is essentially an autoregressive generation problem of the vectorized index sequence: Given the vectorized index sequence S( , ; t + 1) before the historical moment t + 1, construct a sequence learning model (i.e., EMTrans) to predict the vectorized index S( ; t + 1) at the future moment t + 1, that is where, Indicates a step in future time Land use prediction probability distribution The maximum value of the class probability at each position is selected element by element to obtain the final class prediction result.

[0054] This embodiment utilizes a sliding window to continuously predict future map patches, ultimately stitching them together to form a complete future land use map. A relative distance coding (RDC) method based on target graticets is designed. By normalizing the relative distance as a learning bias term for the Transformer's global attention mechanism, the spatial neighborhood understanding capability of the Transformer model is enhanced. This embodiment is based on the future vectorized index sequence predicted by EMTrans. Constructing future vectorized tensors by searching the land use landscape codebook .

[0055] First, the land use change prediction system concatenates or fuses the vectorized index sequence with relative distance encoding and inputs it into the map sequence to generate a learning model. Then, internally, this model uses a multi-layered attention mechanism to predict the index value of the corresponding location at the next time step, based on the index of each location and its relative spatial offset (i.e., relative distance encoding) to other locations, ultimately outputting a complete-length future vectorized index sequence. Next, the system iterates through each index in this future vectorized index sequence, searching for its corresponding discrete representation vector in the landscape codebook. Finally, following the original two-dimensional raster arrangement of the latent vector representations (e.g., the reverse order of a spiral path), the retrieved discrete representation vectors are reorganized into a two-dimensional tensor, resulting in the future vectorized tensor. This is done to restore the discrete sequence prediction results to a spatially continuous and decodeable map latent representation while preserving semantic structure, providing a foundation for the subsequent generation of high-fidelity land use maps.

[0056] As an example, the training steps of the map sequence generation learning model include: obtaining the multi-time-period vectorized tensors output by the trained map variational autoencoder, and arranging the multi-time-period vectorized tensors in chronological order to form a vectorized index sequence dataset; dividing the vectorized index sequence dataset into a training set and a validation set; constructing a map sequence generation learning model based on the input layer, sliding window prediction module, global attention module, and output layer; training the map sequence generation learning model based on a preset number of iterations, a preset batch size, a preset learning rate, and the training set, and monitoring the model's prediction accuracy through the validation set; stopping training when the model's prediction accuracy converges, thus obtaining the trained map sequence generation learning model.

[0057] Multi-time-period vectorized tensors refer to a set of two-dimensional discrete latent tensors with identical structures, arranged chronologically, obtained by encoding and vectorizing land use submaps and geographic environment submaps from multiple consecutive historical periods using a trained map variational autoencoder. Each element in the tensor is the index value of the corresponding location in the landscape codebook. The vectorized index sequence dataset refers to the dataset where the aforementioned multi-time-period vectorized tensors are expanded step-by-step according to the same spatial location and organized into several input-target pair sequences based on a preset time-sliding window strategy. Each sequence consists of vectorized indices from consecutive time steps and is used to train the map sequence generation learning model. The input layer is the starting module of the map sequence generation learning model, responsible for receiving the input vectorized index sequence and converting it into a dense vector representation that can participate in neural network calculations through the embedding layer. It also incorporates relative distance encoding to inject spatial location relationship information. The sliding window prediction module is a structure in the model used to limit the temporal context. It slides a fixed-length time window from front to back, extracting several consecutive periods of historical vectorized indices as input each time to predict the index sequence of the next period, thereby achieving temporal dependency modeling and rolling prediction.

[0058] The global attention module refers to a multi-head self-attention mechanism component based on the Transformer architecture. It can globally interact and model features across all time steps and spatial locations within the sliding window, capturing long-range spatiotemporal dependencies without being limited to local neighborhoods or fixed receptive fields. The output layer is the linear projection layer at the end of the model, mapping the high-dimensional features output by the global attention module to a classification logits space the size of the landscape codebook, and then generating the predicted vectorized index for the next time step through softmax or argmax operations. The preset number of iterations refers to the maximum number of training rounds set before training begins for the map sequence generation learning model, used to control the total training time. The preset batch size refers to the number of samples used for each parameter update, i.e., the number of sliding window sequence samples processed in parallel in each iteration, affecting training stability and memory usage. The preset learning rate refers to the initial step size used by the optimizer when updating model parameters, used to control the magnitude of parameter adjustments, and is usually used in conjunction with a learning rate scheduling strategy. Model prediction accuracy refers to the degree of consistency between the future vectorized index sequence output by the model and the true vectorized index sequence on the validation set. It is usually measured by metrics such as accuracy or F1 score to determine whether the model has converged.

[0059] First, the land use change prediction system sequentially applies a pre-trained map variational autoencoder to land use submaps and geographic environment submaps from multiple consecutive historical periods, outputting vectorized indices for each period and restoring them to multi-time-period vectorized tensors according to the original two-dimensional spatial layout. Then, the system stacks these tensors chronologically and uses a fixed-length time sliding window (e.g., using the previous k periods to predict the k+1 period) to segment input-target pairs from each time series, forming a vectorized index sequence dataset consisting of historical vectorized index sequences and their corresponding future indices. This dataset is then divided into training and validation sets in a ratio (e.g., 8:2). Next, the system constructs a map sequence generation learning model, sequentially connecting an input layer (converting indices into embedding vectors and adding relative distance encoding), a sliding window prediction module (limiting the input time range), a global attention module (calculating the spatiotemporal dependencies between all locations), and an output layer (outputting the index predictions for each location at the next time step). Subsequently, the system samples data from the training set at a preset batch size, updates parameters using an optimizer with a preset learning rate, and repeats this process until a preset number of iterations is reached. After each training iteration, the system calculates the matching accuracy between the predicted index and the true index on the validation set as the model's prediction precision. Finally, when this precision no longer significantly improves within several consecutive rounds (i.e., convergence, or the change value is repeatedly less than a preset value), the system stops training, saves the model parameters, and obtains the trained map sequence generation learning model. The purpose of this is to enable the model to learn the spatiotemporal patterns of land use evolution from historical discrete index sequences and to possess the ability to make structurally consistent and semantically reasonable predictions of future patterns.

[0060] Step S60: The future vectorized tensor is iteratively decoded using the map decoder to generate land use change results.

[0061] It should be noted that the land use change result refers to the raster image output by the map decoder after decoding the future vectorized tensor, which represents the spatial distribution of land use types in the prediction period. The value of each pixel corresponds to a specific land use category (such as residential land, cultivated land, forest land, etc.), which is used to reflect the evolution of land use patterns from the current or historical state to the future period.

[0062] Understandably, the land use change prediction system first looks up the index value of each location in the future vectorized tensor in the corresponding landscape codebook, replacing it with the corresponding discrete representation vector to reconstruct a continuous latent feature tensor. Then, the system inputs this tensor into the map decoder, sequentially passing it through multiple upsampling modules consisting of depthwise separable convolutional layers, residual connection layers, group normalization layers, and ELU activation functions, progressively restoring spatial resolution and fusing detailed information. Next, the map decoder converts the feature map into a class probability distribution map with the same size as the original input sub-map through a final two-dimensional convolutional layer and a softmax layer. Finally, the system selects the class with the highest probability for each pixel as the final predicted value, outputting a complete raster image, i.e., the land use change result. This is done to reconstruct, with high fidelity, a discrete, compact future latent representation into a land use pattern with clear semantics and a reasonable spatial structure, to support subsequent analysis and decision-making.

[0063] Please refer to Figure 3 , Figure 3 This is a schematic diagram of the architecture of EMTrans, provided in Embodiment 1 of the land use change prediction method based on map representation generation learning in this application. The left side shows the historical global map input. First, the MapsVAE model encodes the multi-source map data, transforming it into a discrete vectorized index sequence. Specifically, MapsVAE extracts latent representations from the historical land use map and geographic environment map, and maps them to index values ​​in the corresponding landscape codebook through a vectorization module, forming a series of vectorized indices at each time step. Subsequently, these vectorized indices are flattened into a one-dimensional sequence in chronological order and combined with relative distance encoding to form the input sequence used to construct the vectorized index sequence. The middle part is the Transformer model, which receives this sequence as input and uses a sliding window mechanism to generate the vectorized index prediction results for future time steps. This prediction result is expanded and fed into the trained decoder to complete the reconstruction from discrete indices to continuous features. The right side shows the future land use generation process. The feature tensor output by the decoder is further decoded into a future global land use map, realizing the prediction of the spatiotemporal evolution of land use patterns. The entire process achieves end-to-end generative modeling from historical maps to future maps, integrating spatial structure information and temporal evolution patterns.

[0064] This embodiment provides a land use change prediction method based on map representation generation learning. First, multi-source map sub-maps are constructed, and these sub-maps are binarized and attention-weighted to obtain binarized land use sub-maps and weighted geographic environment sub-maps, effectively separating target land use types and highlighting the influence of key environmental factors. Second, the map encoder in a map variational autoencoder performs deep aggregation spatial feature extraction on the two types of sub-maps, generating compact latent vector representations, thereby fusing high-order semantic information of land use patterns and environmental driving factors. Next, this latent vector representation is flattened by traversing the raster grid along a clockwise spiral path, forming a one-dimensional feature sequence that preserves local spatial proximity relationships while adapting to the sequence modeling input format. Then, the vectorization module queries the index corresponding to each element of the sequence in the landscape codebook to obtain a vectorized index sequence, realizing the transformation from continuous features to discrete structured representation. Subsequently, based on this index sequence and relative distance encoding, a map sequence generation learning model is used to generate a future vectorized index sequence, and the landscape codebook is queried accordingly to obtain a future vectorized tensor, enabling the model to learn the spatiotemporal dependency of land use evolution at the mesoscale. Finally, the future vectorized tensor is iteratively decoded by a map decoder to reconstruct land use change results with reasonable spatial structure and semantic consistency. This embodiment achieves more accurate land use change prediction by capturing mesoscale spatial local information and reducing unit-level error accumulation. This embodiment's landscape representation method based on map variational autoencoder learning learns rich landscape codes from multi-view maps, which can fully capture complex spatial local information and overcome the shortcomings of existing methods that ignore the mesoscale land use landscape connotation. This embodiment can automatically learn the spatiotemporal local dependency of map sequence data, avoiding the dilemma of high prior information requirements and complex hyperparameter calibration faced by the quantitative process of unit transformation rules in traditional methods.

[0065] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 , Figure 4 This is a flowchart illustrating the second embodiment of the land use change prediction method based on map representation generation learning in this application. Steps Sx0 of the land use change prediction method based on map representation generation learning include steps S61 to S64: Step S61: The future vectorized tensor is decoded by the map decoder to obtain a future land use map; Step S62: Predict the first quantity of various land uses in the next year using a Markov model, and calculate the quantity difference between the first quantity and the second quantity of various land uses in the future land use map; Step S63: When the quantity difference is greater than the preset difference threshold, the parameters of the map sequence generation learning model are adjusted, and the steps of generating future vectorized index sequences based on the vectorized index sequence and relative distance encoding through the map sequence generation learning model, and querying the landscape codebook according to the future vectorized index sequence to obtain the future vectorized tensor are returned. Step S64: When the quantity difference is less than or equal to the preset difference threshold, generate land use change results based on the future land use map.

[0066] It should be noted that the future land use map refers to a raster image generated by a map decoder after decoding the future vectorized tensor, representing the spatial distribution of various land use types in the prediction period (next year), where each pixel corresponds to a specific land use category. A Markov model is a land use quantity prediction model based on state transition probabilities. It assumes that the area of ​​a certain type of land use in the future depends only on the current state and estimates the total amount of various land uses in the next year through a historical transition matrix. The first quantity of various land uses in the next year refers to the total area or number of pixels of each land use type predicted by the Markov model based on historical land use statistics and transition probabilities. The second quantity refers to the area or quantity of each type of land use obtained by counting the actual number of pixels appearing in each land use category in the future land use map. The preset difference threshold is an upper limit of allowable error set manually during the model calibration stage, used to determine whether the deviation between the macroscopic quantity predicted by the Markov model and the map decoding result is within an acceptable range.

[0067] Understandably, firstly, the MapD module of the decoder in the trained land use map variational autoencoder model LMapVAE can... ( The land use data (t+1) is reconstructed into a human-visualized land use map. Then, Markov modeling is used to model the changes in the quantity of various land uses over consecutive years to predict the total quantity of each land use type in the next year. Based on this, the EMTrans model is iteratively run to output a future vectorized index. Then, the landscape codebook is queried based on the future vectorized index to obtain the future vectorized tensor. ( The LMapVAE's MapD module continuously decodes the data until the difference between the decoded future land use map and the Markov's predicted land use quantities is minimized. Then, the current land use map is used as the prediction result for the corresponding future year.

[0068] As an example, the training steps of the Markov model include: acquiring land use type quantity data for several consecutive years to form a historical quantity dataset; calculating the annual transfer probability between various land use types based on the historical quantity dataset and constructing a transfer probability matrix; dividing the historical quantity dataset into a model training set and a model test set; training the Markov model based on the model training set and the transfer probability matrix, and verifying the model prediction error based on the test set; stopping training when the model prediction error is less than a preset error threshold, and obtaining the trained Markov model.

[0069] Land use type quantitative data refers to the total number of pixels or area statistics corresponding to each land use type within a study area in a specific year, used to quantify the scale of various land uses. Historical quantitative datasets are collections of land use type quantitative data organized chronologically over several consecutive years, used to model the temporal evolution of land use quantity. Land use types refer to land use categories classified according to national or industry standards, such as cultivated land, forest land, grassland, water area, construction land, and unused land, each with a clear semantic definition and functional attributes. Annual transition probability refers to the probability that a certain land use type will transform into another between two adjacent years, calculated by the ratio of historical changes to the initial state quantity. The transition probability matrix is ​​a square matrix indexed by land use type, where each element represents the probability of transforming from one land use type to another within a year, comprehensively describing the mutual transformation relationships between various land use types. Model prediction error refers to the deviation between the predicted and actual observed values ​​of land use quantity in future years of the test set after the Markov model predicts the land use quantity based on the transition probability matrix obtained during training; it is usually measured using mean squared error or mean absolute error. The preset error threshold is an acceptable upper limit of error set before model training. It is used to determine whether the prediction accuracy of the Markov model meets the requirements and serves as the basis for terminating training.

[0070] First, the land use change prediction system statistically analyzes the total number of pixels for each land use type year by year from historical land use maps, obtaining continuous years of land use type quantity data, which is then organized into a historical quantity dataset in chronological order. Next, the system iterates through all adjacent year pairs (e.g., year t and year t+1) in this dataset. For each year t, the system counts the number of pixels for each land use type that will change to other types in year t+1, divides this by the original total number of that type in year t, calculates the annual transition probability between types, and fills these probabilities into a square matrix indexed by land use type, forming a transition probability matrix. Then, the system divides the historical quantity dataset into a model training set (e.g., the first N-k years) and a model test set (e.g., the next k years), and multiplies the initial year quantity vector of the training set with the transition probability matrix to predict the land use quantity for subsequent years year by year. Finally, the system compares the prediction results with the actual quantities for the corresponding years in the model test set, calculating the mean absolute error as the model prediction error. Finally, if the error is less than a preset error threshold, the transition probability matrix is ​​considered to accurately reflect the macroscopic evolution of land use in the region. Training is then stopped, and the transition probability matrix is ​​output as the trained Markov model. This is done to ensure that the generated map is not only spatially reasonable but also conforms to long-term statistical patterns in terms of total trends, thereby improving the reliability and practicality of the prediction results.

[0071] As an example, the step of adjusting the parameters of the map sequence generation learning model includes: calculating the proportion of the quantity difference to the first quantity; if the proportion is greater than a first preset proportion, increasing the sliding window prediction step size of the map sequence generation learning model by a preset step size increment, and increasing the autoregressive generation probability threshold of the map sequence generation learning model by a preset threshold increment; if the proportion is less than or equal to the first preset proportion, decreasing the sliding window prediction step size by the preset step size increment, and decreasing the autoregressive generation probability threshold by the preset threshold increment; obtaining the initial weight allocation of the elements corresponding to various land uses in the vectorized index sequence; for land use types with positive quantity differences, increasing the weight of their corresponding index elements by half based on the initial weight; for land use types with negative quantity differences, decreasing the weight of their corresponding index elements by half based on the initial weight, thus completing the parameter adjustment.

[0072] The first preset ratio refers to a pre-set threshold ratio in the parameter adjustment strategy, used to determine whether the difference in quantity is significant relative to the first quantity predicted by the Markov model, thereby determining the adjustment direction of the sliding window prediction step size and the autoregressive generation probability threshold. The sliding window prediction step size refers to the length of the time context used by the map sequence generation learning model during training or inference, i.e., how many consecutive time steps of vectorized index sequences from the past are used to predict the output of the next time step. The preset step size increment refers to a manually set adjustment unit for the sliding window prediction step size, used to increase or decrease the step size with each parameter adjustment. The autoregressive generation probability threshold refers to the confidence threshold for the prediction result of each location during the autoregressive generation of future vectorized index sequences. Only when the generation probability of a certain type of index exceeds this threshold is it adopted; otherwise, a resampling or correction mechanism is triggered. The preset threshold increment refers to a manually set adjustment unit for the autoregressive generation probability threshold, used to fine-tune this threshold during parameter adjustments. Initial weight allocation refers to the pre-defined category weight vectors for index elements corresponding to different land use types before the vectorized index sequence is input into the map sequence to generate the learning model. This is used to balance the contribution of each category in loss calculation or attention mechanism. By default, it is usually initialized according to category frequency or uniform distribution.

[0073] First, the system calculates the quantity difference for each land use category and divides it by its corresponding first quantity to obtain a ratio. Then, it determines whether this ratio is greater than a first preset ratio: if so, the sliding window prediction step size of the map sequence generation learning model is increased by a fixed preset step size increment, and the autoregressive generation probability threshold is simultaneously increased by a preset threshold increment; if not, the sliding window prediction step size is decreased by the preset step size increment, and the autoregressive generation probability threshold is decreased by the preset threshold increment. Next, the system reads the initial weight assignments of each land use index element in the vectorized index sequence. For categories with positive quantity differences (i.e., fewer areas in the future land use map), the weights of their corresponding index elements are increased by multiplying the initial value by (1 + ratio ÷ 2); for categories with negative quantity differences (i.e., more areas), their weights are decreased by multiplying the initial value by (1 - ratio ÷ 2). Finally, the adjusted sliding window prediction step size, autoregressive generation probability threshold, and category weights are updated in the map sequence generation learning model. This is done to guide the model to gradually correct the total bias in subsequent generation by dynamically adjusting the length of the time context, the generation confidence threshold, and the class importance, so that the prediction results maintain both the rationality of the spatial structure and conform to the macro statistical trend.

[0074] This embodiment first decodes the future vectorized tensor using a map decoder to obtain a future land use map with spatial details. Second, it uses a Markov model to predict the first quantity of each type of land use for the next year, and calculates the actual quantity of each type of land use in the map as the second quantity, then calculates the difference between the two. When this difference is greater than a preset difference threshold, it indicates that the predicted result deviates from the historical trend in terms of total quantity. The system then adjusts the parameters of the map sequence generation learning model and returns to regenerate the future vectorized tensor and subsequent decoding steps. When the difference is less than or equal to the preset difference threshold, the current future land use map is used as the final output to generate the land use change result. This embodiment calibrates the generated result by introducing macro-level quantity constraints, ensuring that the total quantity of each category conforms to long-term evolutionary patterns while maintaining the rationality of the spatial structure.

[0075] As an example, land use data from a certain urban cluster from 2000 to 2022 will be used to illustrate the specific implementation of the present invention. The specific implementation steps of the present invention will be described in detail below with reference to this example: (1) Data Preparation. The land use data used in this study came from the 30-meter resolution annual land cover dataset of China. This study used land use product data from 2000 to 2022 to verify the effectiveness of the proposed model. The mode strategy was used to resample the original data to a 300-meter resolution. Geographical environment driving factors are divided into three categories: natural environment, socio-economic conditions, and transportation accessibility, including DEM, slope, nearest distance to water system, population density, GDP output, kernel density of administrative sites, nearest distance to station, nearest distance to airport, nearest distance to railway, and nearest distance to highway, etc.

[0076] Table 1 (Geographical Environmental Characteristics Driving Land Use Change) (2) Model training. First, land use maps and multi-dimensional geographic environment maps from 2000 to 2022 were prepared. To evaluate the model's predictive ability, the data maps from 2000 to 2015 were used for model training, while the data from 2016 to 2022 were used as an independent validation set. The validation data was never used during the training process.

[0077] The model training employs a two-stage strategy: the first stage trains the MapsVAE model, and the second stage trains the EMTrans model based on the vectorized index output by MapsVAE. Specifically, in the first stage, 10,000 map tiles are randomly selected from the 2015 map, and 64×64 land use sub-maps and geographic environment sub-maps are extracted for training the LMapVAE and GMapVAE modules. In the second stage, the land use map and geographic environment map of the study area from 2000 to 2015 are segmented into non-overlapping, interconnected sub-maps. The trained MapsVAE model is then used to encode all land use sub-maps and geographic environment sub-maps from 2000 to 2015 into V( ) and V( And arranged in chronological order as S( , ; t=2000), S( , ; t=2001), ..., S( , The sequence is t=2015. Based on this, the EMTrans model can be based on S( , ;<2016) Iterative prediction S( (2016) This study coupled the MapD and Markov models trained in LMapVAE to determine a land use prediction map that meets the 2016 land use quantity constraints. The future land use maps for 2017-2022 are predicted iteratively using the same method: the EMTrans model is responsible for iteratively generating a vectorized tensor of future land use, the MapD module in LMapVAE is responsible for decoding the vectorized tensor into a land use distribution map, and the Markov model predicts the quantities of various land uses in the next year to determine the best-matching land use distribution map.

[0078] (3) Parameter settings. The initial learning rate was set to 5×10⁻⁵. The landscape codebook size for both LMapVAE and GMapVAE was set to 128, representing that the potential landscape type can be represented by a maximum of 128 vectors, ensuring the richness of the potential landscapes learned by the MapsVAE model. Based on the number of input feature channels and the number of convolutional layers, the landscape codebook vector dimensions for LMapVAE and GMapVAE were calculated to be 32 and 128, respectively. The input block size ri and output block size ro were set to 10 and 2, respectively, that is, using a historical grid area of ​​size 10×10 to predict a future grid area of ​​size 2×2. The Epoch and Batch Size for LMapVAE and GMapVAE were set to 500 and 128, respectively, and the Epoch and Batch Size for EMTrans were set to 100 and 600, respectively. The model training used the Adam optimizer and the pass-through gradient estimator. The MapsGT model was deployed on a Windows computer equipped with a single NVIDIA GTX 4090D GPU, running PyTorch 2.4.1.

[0079] (4) Prediction Validation. Comparative experiments were conducted with the FLUS model and the CNN-LSTM-CA model. Quantitative evaluation was performed using FoM (Figure of Merit), Producer Precision (PA), and User Precision (UA). PA is positively correlated with the model's ability to hit real-world changing grids. UA is the proportion of grids predicted by the model that have actually changed, and it has a negative relationship with the model's false alarm rate. The larger the UA, the lower the false alarm rate; conversely, the smaller the UA, the higher the model's false alarm rate. FoM represents the model's overall performance in terms of hitting actual changes and the effectiveness of predicting changes. The higher the FoM, the stronger the model's ability to predict changes correctly, and the relatively lower the false alarm rate.

[0080] In the formula, A represents the number or area of ​​rasters whose actual land use semantics have changed but whose model predictions have not changed; B represents the number or area of ​​rasters whose land use semantics have changed in both the real world and the prediction results, and whose predicted land use semantics are also correct; C represents the number or area of ​​rasters whose land use semantics have changed in both the real world and the prediction results, but whose predicted land use semantics are incorrect; and D represents the number or area of ​​rasters whose actual land use semantics have not changed but whose model predictions have changed.

[0081] Experimental results show that our method achieves the highest FoM and UA indices in short-term predictions, but exhibits lower PA and the highest UA in the Pearl River Delta region, indicating that our model cautiously generates high-confidence future land use maps. ① Regarding the FoM index: The MapsGT model has the highest FoM in short-term predictions, but its FoM advantage gradually decreases and becomes comparable to the baseline model level as the number of predicted future years increases. The CA model (FLUS) and the deep learning-enhanced CA model (CNN-LSTM-CA) show a relative advantage in long-term simulations, which may be due to the top-level control advantage of the cellular simulation system feedback. This indicates that if deep learning models only use historical data to learn land use evolution patterns, it is difficult to infer long-term evolutionary patterns or evolutionary processes that have occurred less frequently in history. ② Regarding the PA index: The PA index reflects the model's recall ability of actual change grids, which is basically consistent with the FoM trend, indicating that our model is not confident in predicting long-term land use changes. Re-examining the MapsGT model reveals that when the generative prediction module EMTrans generates future land use maps based on flattened raster long sequences, the exponential parameters of the Transformer model limit the length of the input year sequence. The CNN-LSTM-CA model, by learning long-term evolutionary patterns through the LSTM module, performs better on the PA index in long-term predictions. ③ On the UA index, the CNN-LSTM-CA model achieves the highest PA index in long-term predictions at the cost of the highest false alarm rate, reflected in the lowest UA index. The UA index represents the realism of the model's predicted changes in the raster: an excessively low UA usually indicates that the prediction model is too bold and aggressive, easily predicting semantic changes in land use.

[0082] Table 2 (Quantitative evaluation results of MapsGT, FLUS and CNN-LSTM-CA models) (5) Regional Consistency Assessment. This study designed the Regional Consistency Index (RCI) based on a moving window to measure the local regional accuracy of the prediction results. The RCI does not focus on the absolute correctness of a single raster, but rather on the overall reliability of the simulation results at the regional scale, thus serving as a mesoscale supplementary measure of the tolerance for prediction errors of a single raster.

[0083] In the formula, loc and These represent the selected center grid and its Moore neighborhood, respectively. Representative at Within the region, the absolute difference in quantity or area between the actual map and the predicted map of land use type k. M and N represent the total number and area of ​​land use types within the study area, respectively. The total number of grid cells or total area within the region. In this experiment, M is set to 4, representing cultivated land, vegetation, water area, and built-up land, respectively. The RCI value ranges from [0, 1]. The larger the RCI value, the higher the consistency of the simulation at that scale. RCI can not only reflect the regional error at a specific location, but also assess the overall regional consistency level of the entire study area by calculating the average RCI of all locations.

[0084] Experimental results show that even when faced with complex land use systems, this method can still effectively capture and reproduce real land use landscape patterns, demonstrating a strong ability to learn, understand, and dynamically reconstruct complex landscapes. Figure 7 The spatial distribution map of the RCI index of the 2022 land use prediction results of the MapsGT, CNN-LSTM-CA and FLUS models is shown, and the simulation accuracy is further tested at the regional scale. Figure 7 middle The size is set to 11×11. In practical applications, decision-makers can flexibly specify the size based on the scale of interest or specific task requirements. Different sizes were used to analyze the reliability of land use prediction models under different planning decision scales. Figure 7 Medium blue indicates that the RCI is close to 1, meaning that the predicted land use composition in the region is closer to the actual situation; red indicates that there is a significant difference between the model's predicted land use patterns and the actual land use distribution. Figure 7 The results indicate that regional inconsistencies in the MapsGT model are primarily concentrated in the Xijiang (Pearl River main stream) basin, and CA-based models also encounter similar problems. Satellite field surveys reveal that the Xijiang basin has a dense network of waterways, with paddy fields dominating arable land. As is well known, paddy fields possess characteristics of both water and arable land. In recent years, with rapid urbanization and flood control projects, the water characteristics of the Xijiang basin have weakened, while the arable land characteristics have become more prominent. The sensitivity of land use product interpretation may create the illusion of rapid land use change, and the band similarity between paddy fields, arable land, and water areas increases the difficulty of predicting multiple land use types. This phenomenon reveals a common challenge faced by current raster-label-based land use prediction models: when the actual landscape undergoes gradual changes, prediction systems based on finite discrete labels struggle to accurately capture the transitional state of landscape content, leading to misjudgment or neglect of land use changes.

[0085] For example, to help understand the implementation process of the land use change prediction method based on map representation generation learning obtained by combining this embodiment with the above embodiment one, please refer to... Figure 5 , Figure 5A simplified flowchart of a land use change prediction method based on map representation generation learning is provided, specifically: On the left is the Multi-Source Map Variational Autoencoder (MapsVAE), whose input is multi-source map data, including land use maps and various geographic environment layers. This data is fed into two structurally identical but parameter-independent sub-networks—one for the land use sub-map and the other for the geographic environment sub-map. The encoder extracts features and generates latent vector representations. These latent vectors are then mapped to discrete indices in the landscape codebook via a vectorization module, forming a vectorized representation. The decoder receives this vectorized representation and reconstructs a map output similar to the original input, completing the encoding-decoding closed-loop training of the historical map. On the right is the Embedded Map Transformer (EMTrans), whose input is the historical vectorized sequence extracted from MapsVAE. The Transformers module, combined with relative distance encoding, models the sequence to generate a future vectorized tensor. This future vectorized tensor is then decoded by the trained decoder, ultimately outputting a future land use map, enabling prediction of land use change. The entire process achieves end-to-end generative modeling from multi-source map data to future spatial patterns.

[0086] This application also provides a land use change prediction device based on map representation generation learning. Please refer to [link / reference]. Figure 6 The land use change prediction device based on map representation generation learning includes: Sub-map construction module 10 is used to construct multi-source map sub-maps, perform binarization processing and attention weighting on the multi-source map sub-maps, and obtain binarized land use sub-maps and weighted geographic environment sub-map features; Feature extraction module 20 is used to perform deep aggregation spatial feature extraction on the features of the binarized land use submap and the weighted geographic environment submap through the map encoder in the map variational autoencoder to obtain latent vector representation. The map variational autoencoder also includes a vectorization module and a map decoder. The vectorization module contains a preset number of landscape codebooks. The spiral flattening module 30 is used to flatten the latent vector representation by traversing the grid in a clockwise spiral path to obtain a one-dimensional feature sequence. The vectorization module 40 is used to query the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain a vectorized index sequence. The sequence generation module 50 is used to generate a future vectorized index sequence based on the vectorized index sequence and the relative distance encoding through a map sequence generation learning model, and to query the landscape codebook according to the future vectorized index sequence to obtain a future vectorized tensor. The iterative decoding module 60 is used to iteratively decode the future vectorized tensor through the map decoder to generate land use change results.

[0087] The land use change prediction device based on map representation generation learning provided in this application, employing the land use change prediction method based on map representation generation learning in the above embodiments, can solve the technical problem of how to achieve more accurate land use change prediction by capturing mesoscale spatial local information and reducing unit-level error accumulation. Compared with the prior art, the beneficial effects of the land use change prediction device based on map representation generation learning provided in this application are the same as those of the land use change prediction method based on map representation generation learning provided in the above embodiments, and other technical features in the land use change prediction device based on map representation generation learning are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0088] This application provides a land use change prediction device based on map representation generation learning. The land use change prediction device based on map representation generation learning includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the land use change prediction method based on map representation generation learning in the above embodiment 1.

[0089] The following is for reference. Figure 7 This document illustrates a structural schematic diagram of a land use change prediction device based on map representation generation learning, suitable for implementing embodiments of this application. The land use change prediction device based on map representation generation learning in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The land use change prediction device based on map representation generation learning shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0090] like Figure 7As shown, the land use change prediction device based on map representation generation learning may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the land use change prediction device based on map representation generation learning. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the land use change prediction device based on map representation generation learning to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a land use change prediction device based on map representation generation learning with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0091] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0092] The land use change prediction device based on map representation generation learning provided in this application, employing the land use change prediction method based on map representation generation learning in the above embodiments, can solve the technical problem of how to achieve more accurate land use change prediction by capturing mesoscale spatial local information and reducing unit-level error accumulation. Compared with the prior art, the beneficial effects of the land use change prediction device based on map representation generation learning provided in this application are the same as those of the land use change prediction method based on map representation generation learning provided in the above embodiments, and other technical features in this land use change prediction device based on map representation generation learning are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0093] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0094] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0095] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the land use change prediction method based on map representation generation learning in the above embodiments.

[0096] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0097] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a land use change prediction device based on map representation generation and learning, cause the land use change prediction device based on map representation generation and learning to: construct multi-source map sub-maps; perform binarization and attention weighting on the multi-source map sub-maps to obtain binarized land use sub-maps and weighted geographic environment sub-map features; and perform deep aggregation spatial feature extraction on the binarized land use sub-maps and the weighted geographic environment sub-map features through a map encoder in a map variational autoencoder to obtain latent vector representations. The map variational autoencoder further includes a vectorization module and... The map decoder includes a vectorization module containing a preset number of landscape codebooks. The latent vector representations are flattened by traversing the grid along a clockwise spiral path to obtain a one-dimensional feature sequence. The vectorization module queries the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain a vectorized index sequence. Based on the vectorized index sequence and relative distance encoding, a future vectorized index sequence is generated using a map sequence generation learning model. The future vectorized index sequence is then used to query the landscape codebook to obtain a future vectorized tensor. The map decoder iteratively decodes the future vectorized tensor to generate land use change results.

[0098] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0099] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described land use change prediction method based on map representation generation learning. This method addresses the technical problem of how to achieve more accurate land use change prediction by capturing mesoscale spatial local information and reducing unit-level error accumulation. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the land use change prediction method based on map representation generation learning provided in the above embodiments, and will not be repeated here.

[0100] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A land use change prediction method based on map representation generation learning, characterized in that, The method includes: Construct a multi-source map sub-map, perform binarization and attention weighting on the multi-source map sub-map, and obtain the binarized land use sub-map and weighted geographic environment sub-map features; The map encoder in the map variational autoencoder performs deep aggregation spatial feature extraction on the features of the binarized land use submap and the weighted geographic environment submap to obtain latent vector representations. The map variational autoencoder also includes a vectorization module and a map decoder. The vectorization module contains a preset number of landscape codebooks. The latent vector representation is flattened by traversing the grid in a clockwise spiral path to obtain a one-dimensional feature sequence. The vectorization module queries the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain the vectorized index sequence. Based on the vectorized index sequence and relative distance encoding, a future vectorized index sequence is generated through a map sequence generation learning model, and the landscape codebook is queried according to the future vectorized index sequence to obtain the future vectorized tensor. The map decoder iteratively decodes the future vectorized tensor to generate land use change results.

2. The method as described in claim 1, characterized in that, The step of iteratively decoding the future vectorized tensor using the map decoder to generate land use change results includes: The future vectorized tensor is decoded using the map decoder to obtain a future land use map; The first quantity of various land uses in the next year is predicted using a Markov model, and the difference between the first quantity and the second quantity of various land uses in the future land use map is calculated. When the quantity difference is greater than the preset difference threshold, the parameters of the map sequence generation learning model are adjusted, and the steps of generating future vectorized index sequences based on the vectorized index sequence and relative distance encoding through the map sequence generation learning model, and querying the landscape codebook according to the future vectorized index sequence to obtain the future vectorized tensor are returned. When the quantity difference is less than or equal to the preset difference threshold, land use change results are generated based on the future land use map.

3. The method as described in claim 1, characterized in that, The latent vector representation is as follows: in, This refers to the aforementioned land use sub-map. This refers to depth-separable convolutional layers. This refers to the shallow features of the land use submap. This refers to depth-separable residual module groups. This refers to the deeper features of the land use submap. This refers to the group normalization layer. This refers to the potential vector representation corresponding to the land use submap.

4. The method as described in claim 1, characterized in that, The steps of constructing a multi-source map sub-map, performing binarization and attention weighting on the multi-source map sub-map, and obtaining the binarized land use sub-map and weighted geographic environment sub-map features include: A multi-source map dataset containing land use and various geographic environments is constructed. The multi-source map dataset is then divided into blocks and normalized to obtain multi-source map sub-maps, which include land use sub-maps and geographic environment sub-maps. The various land use semantic tags of the land use submap are converted into a raster map with multiple binarized channels to obtain a binarized land use submap; Attention weighting is applied to the geographic environment submap to obtain weighted geographic environment submap features.

5. The method as described in claim 4, characterized in that, The step of performing attention weighting on the geographic environment sub-map to obtain weighted geographic environment sub-map features includes: The feature vectors corresponding to each environmental variable in the geographic environment sub-map are obtained by performing feature compression on each environmental variable through global average pooling. The feature vectors are processed by a one-dimensional convolutional layer with a preset number of channels to obtain the local interaction patterns between the feature vectors. Sigmoid activation is applied to the local interaction pattern to generate attention weights corresponding to each of the environmental variables; The attention weights are multiplied point-by-point with the original feature tensor of the geographic environment subgraph, and the resulting feature tensor is standardized to obtain the weighted geographic environment subgraph features.

6. The method as described in claim 1, characterized in that, The construction steps of the map variational autoencoder include: A map encoder is constructed based on depthwise separable convolutional layers, residual connection layers, group normalization layers, and ELU activation functions. A vectorization module is constructed based on a preset number of landscape codebooks and the vector mapping rules of nearest neighbor search. The landscape codebooks consist of multiple discrete representation vectors. A map decoder is constructed based on the depth-separable convolutional layer, the residual connection layer, the group normalization layer, the ELU activation function, the two-dimensional convolutional layer, and the softmax layer. A map variational autoencoder is constructed based on the map encoder, the vectorization module, and the map decoder.

7. The method according to any one of claims 1 to 6, characterized in that, The map variational autoencoder includes two sub-variational autoencoders constructed from the land use submap and the geographic environment submap, respectively. The two sub-variational autoencoders have the same structure and their parameters are independent of each other. The training steps of the map variational autoencoder include: Construct the total loss function based on the reconstruction loss, vectorization loss, and commitment loss; A training dataset is constructed based on multiple land use submap training samples and a corresponding number of geographic environment submap training samples. Both the land use submap training samples and the geographic environment submap training samples are raster submaps with a preset block size. The map variational autoencoder is iteratively trained based on the training dataset using a preset optimizer and a pass-through gradient estimator. In each iteration, the loss function value is calculated based on the total loss function, and the model parameters are adjusted in reverse based on the loss function value. Training stops when the number of iterations reaches a preset iteration threshold or the loss function value converges to less than or equal to the preset loss threshold, and the trained map variational autoencoder is obtained.

8. A land use change prediction device based on map representation generation learning, characterized in that, The device includes: The sub-map construction module is used to construct multi-source map sub-maps, perform binarization and attention weighting on the multi-source map sub-maps, and obtain binarized land use sub-maps and weighted geographic environment sub-map features; The feature extraction module is used to perform deep aggregation spatial feature extraction on the features of the binarized land use submap and the weighted geographic environment submap through the map encoder in the map variational autoencoder to obtain latent vector representations. The map variational autoencoder also includes a vectorization module and a map decoder. The vectorization module contains a preset number of landscape codebooks. The spiral flattening module is used to flatten the latent vector representation by traversing the grid in a clockwise spiral path to obtain a one-dimensional feature sequence. The vectorization module is used to query the index value of each element in the one-dimensional feature sequence from the corresponding landscape codebook to obtain a vectorized index sequence. The sequence generation module is used to generate a future vectorized index sequence based on the vectorized index sequence and the relative distance encoding through a map sequence generation learning model, and to query the landscape codebook based on the future vectorized index sequence to obtain a future vectorized tensor. The iterative decoding module is used to iteratively decode the future vectorized tensor through the map decoder to generate land use change results.

9. A land use change prediction device based on map representation generation learning, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the land use change prediction method based on map representation generation learning as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the land use change prediction method based on map representation generation learning as described in any one of claims 1 to 7.

Images