A method for identifying historical pattern of old city based on time section and historical layer accumulation

Abstract

The application provides an old city historical pattern recognition method based on time sections and historical layer accumulation, and the method comprises the following steps: collecting historical data, forming a holographic historical space database, and generating three-level dynamic time sections of a century, a decade and an interannual period; constructing an element gene map and a historical layer accumulation element map based on the time sections; recognizing pattern mutation points and quantifying evolution rules based on the maps; performing three-fold evidence verification and spatial analysis based on the rules to obtain verification results; and formulating a graded protection strategy based on the results. The application integrates multi-scale space-time correlation through three-level time sections, solves the problem of time dimension fragmentation, and accurately captures pattern mutations and periodic evolution. The three-fold evidence verification and spatial autocorrelation analysis improve the reliability and scientificity of the evaluation of historical authenticity and texture integrity. The global and local spatial analysis quantifies the texture integrity index, provides scientific support for the graded protection, and improves the systematicness of the old city historical pattern recognition and protection.

Description

Technical Field

[0001] This invention relates to the field of old town historical pattern identification technology, and in particular to an old town historical pattern identification method based on time cross-section and historical layering. Background Technology

[0002] Current methods for identifying the historical layout of old cities primarily rely on traditional GIS spatial analysis, remote sensing image interpretation, and archival documentation. In practice, scholars often overlay historical maps from a single time point with current remote sensing images, using visual interpretation or simple spatial overlay analysis to identify static spatial features such as building fabric and street layout. Some studies attempt to introduce a time dimension, but these are mostly limited to macro-narratives at a century-scale, lacking dynamic connections at meso- and micro-scales such as decades or years. In the verification phase, traditional methods mainly rely on expert judgment or single-document corroboration, such as verifying the location of a section of city wall through local gazetteers or determining the age based on architectural style, lacking a multi-source evidence cross-verification mechanism. Spatial continuity assessments often employ simple buffer analysis or kernel density estimation, making it difficult to quantify the dynamic evolution of the integrity of the urban fabric. Summary of the Invention

[0003] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it proposes an innovative method for identifying the historical pattern of old towns based on time cross-section and historical layering.

[0004] To achieve the above-mentioned objectives of this invention, this invention provides a method for identifying the historical pattern of old towns based on time cross-sections and historical layering, the method comprising: S1. Collect and preprocess historical maps, remote sensing images, document archives and oral history data, and form a holographic historical spatial database through georegistration and timestamp calibration; S2. Based on the holographic historical spatial database, a three-level dynamic time slice is generated by using the key event anchor point extraction algorithm and the wavelet transform periodic slicing algorithm, which includes macroscopic slices at the centennial scale, mesoscopic slices at the ten-year scale, and microscopic slices at the interannual scale. S3. Based on the three-level dynamic time section, use CNN feature extractor and graph neural network to construct element gene map, and generate historical layered element map through building texture segmentation, street network topology analysis and landscape pattern field simulation. S4. Based on the historical layered feature map, identify pattern abrupt change points and quantify evolution patterns using a spatiotemporal graph convolutional network, a long short-term memory network, and a mutation detection algorithm. S5. Based on the aforementioned evolutionary pattern, a multi-criteria verification system and spatial autocorrelation analysis are used to verify the historical authenticity through three pieces of evidence: literature, physical objects, and word-of-mouth. Spatial continuity analysis is used to assess the integrity of the texture and obtain the verification results. S6. Develop a graded protection strategy based on the verification results.

[0005] On the other hand, the present invention also provides a system for identifying the historical pattern of old towns based on time sections and historical layers, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the old town historical pattern recognition method based on time section and historical layer when executing the executable instructions.

[0006] The beneficial effects of this invention are as follows: By constructing a three-level dynamic time section, this invention achieves the organic integration of multi-scale spatiotemporal correlations, effectively solving the problem of fragmented time dimensions in traditional methods. It dynamically connects macro-narratives at the centennial scale, meso-level fluctuations at the decadal scale, and micro-level evolutions at the interannual scale, enabling precise capture of the abrupt changes in the historical pattern and the cyclical evolutionary patterns triggered by key historical events. Regarding the verification mechanism, it innovatively adopts a cross-verification system of three pieces of evidence—documents, physical objects, and word-of-mouth—combined with multi-dimensional quantitative assessment through spatial autocorrelation analysis. This breaks through the limitations of traditional single-evidence verification, significantly improving the credibility of historical authenticity assessment and the scientific rigor of the integrity assessment. At the spatial analysis level, through the combined application of the global Moran's I index and local LISA hotspot identification, it achieves a leap from static spatial feature identification to dynamic evolutionary pattern quantification. It can accurately quantify core indicators such as the diversity of types and chronological gradients of historical integrity, providing scientific and dynamic data support for the formulation of graded protection strategies. Ultimately, it forms a closed-loop solution of "identification-verification-assessment-protection," significantly improving the systematic nature of old city historical pattern identification and the scientific rigor of protection strategies.

[0007] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0008] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a method for identifying the historical pattern of an old city based on time cross-section and historical layering, according to the present invention. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] like Figure 1 As shown, a method for identifying the historical pattern of an old town based on time cross-sections and historical layering is presented. The method includes: S1. Collect and preprocess historical maps, remote sensing images, document archives and oral history data, and form a holographic historical spatial database through georegistration and timestamp calibration; Historical map preprocessing requires unifying the coordinate system across different eras of data, including Ming and Qing dynasty local gazetteer maps, Republic of China-era measured maps, and topographic maps from the post-1949 era. Permanent features such as city gates and ancient bridges are selected as control points for georegistration, with errors controlled within ±5 meters. For damaged or blurred maps, histogram equalization and edge detection techniques are used to restore details. OCR is used to extract street names, building types, and other information, which are then linked to spatial locations. Aerial imagery from years of remote sensing image processing undergoes radiometric correction to eliminate atmospheric scattering, geometric correction to correct topographic distortion, and multispectral and panchromatic image fusion to improve resolution to 0.5 meters. Key timestamps are added in conjunction with urban redevelopment and disaster event chronologies. Regarding archival documents, LDA thematic models are used to extract spatial event information such as street expansion and building repairs from local histories and urban construction archives, storing the time, location, and content in a structured manner. For spatial descriptions such as "Main Street Inside the East City Gate," place name matching algorithms are used to link geographic coordinates. After oral history data is transcribed from audio recordings, natural language processing is used to extract memory information such as "the courtyard house next to the old locust tree." Semantic similarity matching is used to associate it with historical map elements, while metadata such as the age of interviewees and length of residence is recorded to assess the weight of evidence. Finally, all preprocessed data is integrated into a spatiotemporal database to establish an index linking spatial elements (buildings, streets, landscapes) with temporal events, supporting multi-dimensional queries by time slice or spatial location.

[0011] S2. Based on the holographic historical spatial database, a three-level dynamic time slice is generated by using the key event anchor point extraction algorithm and the wavelet transform periodic slicing algorithm, which includes macroscopic slices at the centennial scale, mesoscopic slices at the ten-year scale, and microscopic slices at the interannual scale. S3. Based on the three-level dynamic time section, use CNN feature extractor and graph neural network to construct element gene map, and generate historical layered element map through building texture segmentation, street network topology analysis and landscape pattern field simulation. S4. Based on the historical layered feature map, identify pattern abrupt change points and quantify evolution patterns using a spatiotemporal graph convolutional network, a long short-term memory network, and a mutation detection algorithm. S5. Based on the aforementioned evolutionary pattern, a multi-criteria verification system and spatial autocorrelation analysis are used to verify the historical authenticity through three pieces of evidence: literature, physical objects, and word-of-mouth. Spatial continuity analysis is used to assess the integrity of the texture and obtain the verification results. S6. Develop a graded protection strategy based on the verification results.

[0012] Based on historical authenticity verification results (triple evidence support score, spatiotemporal alignment, and consistency coefficient) and texture integrity assessment results (spatial continuity index, type diversity index, and chronological gradient index), a three-tiered protection level classification model of "core protection - key control - general coordination" is constructed. The core protection level requires a triple evidence support score ≥ 0.85, a spatial continuity index ≥ 0.9, and a chronological gradient index ≤ 0.2; the key control level requires a support score of 0.7-0.85, a continuity index of 0.75-0.9, and a chronological gradient index of 0.2-0.4; and the general coordination level requires a support score < 0.7, a continuity index < 0.75, or a chronological gradient index > 0.4. For core protection level areas, authenticity protection measures are implemented: any alteration to the street network topology or architectural texture is prohibited; traditional materials and techniques are used to repair damaged elements; and millimeter-precision spatial monitoring sensors are installed to collect element status data monthly. For key control level areas, a restrictive renewal strategy is implemented: low-intensity functional replacement (such as converting traditional residences into cultural exhibition spaces) is permitted, but core parameters such as building height and street width must remain consistent with the historical stratification element map. Renewal plans must pass triple evidence cross-verification and spatial autocorrelation analysis before approval. For general coordination level areas, adaptive optimization measures are implemented: the spatial layout is reconstructed in accordance with contemporary urban functional needs, preserving key historical anchor elements such as ancient trees, ancient wells, and archways. The reconstruction plan must ensure a local spatial continuity index ≥ 0.6. A dynamic adjustment mechanism for protection strategies is established: every 3 years, the S1-S5 verification process is re-executed based on the updated holographic historical spatial database, and the protection level and corresponding measures are adjusted according to the new pattern evolution, forming a closed-loop management system of "assessment-adjustment-implementation".

[0013] As an optional embodiment of the present invention, optionally, in step S2, based on the holographic historical spatial database, a key event anchor point extraction algorithm and a wavelet transform periodic slicing algorithm are used to generate a three-level dynamic time slice containing a macroscopic slice at a hundred-year scale, a mesoscopic slice at a ten-year scale, and a microscopic slice at an interannual scale, including: S201. Based on the holographic historical spatial database, using the NLP hybrid model and entropy weight method, keywords are matched by regular expressions and combined with the contextual understanding of the BERT-wwm model to extract specific events. After calculating the influence weight of specific events, a key event anchor point library is formed. Regular expressions are used to construct keyword matching rules related to the evolution of the old city's layout, including event type keywords (such as "war destruction," "street and alley expansion," "policy promulgation," "disaster damage," and "building repair"), timestamp keywords (such as "Year XXXX" and "early days of the People's Republic of China"), and spatial location keywords (such as "XX city gate," "XX street and alley," and "XX ancient bridge"). Through preset regular expression patterns, text fragments containing potential key events are quickly filtered from archival documents and oral history transcripts, initially filtering out irrelevant information. The existing BERT-wwm model, based on a Chinese full-word mask pre-training mechanism, performs contextual semantic understanding on the text fragments filtered by the regular expressions, resolving ambiguity issues (such as distinguishing whether "demolition and construction" refers to partial repair or overall demolition). Simultaneously, it extracts the structured attributes of events, including the event subject (such as government departments), object (such as building clusters and street networks), action (such as new construction, demolition, and renovation), and descriptions of the scope of impact, forming a semantic feature vector for the event. Utilizing existing... The NLP hybrid model combines the efficient initial screening of regular expressions with the deep semantic analysis of BERT-wwm. First, regular expressions are used to quickly locate event-related text, then BERT-wwm is used to accurately extract event attributes, outputting structured event records (including fields such as event ID, occurrence time, spatial location, event type, and impact description). The entropy weighting method is used to calculate the impact weight of events. First, evaluation indicators are selected: the number of spatial units involved in the event (e.g., the number of building texture units covered, the number of street nodes), the duration of time (e.g., the project implementation period, the policy impact period), the frequency of mentions in multi-source documents (the number of times the event appears in local histories and urban construction archives), and the intensity of mentions in oral history (interviewees' ratings of their attention to the event, 1-5 points). After standardizing each indicator, information entropy is calculated, and the indicator weights are inferred from the entropy values ​​(the smaller the entropy value, the greater the weight). Finally, a weighted sum is used to obtain the comprehensive impact weight of each event, and events with a weight ≥ 0.7 are included in the key event anchor point library.

[0014] The expression for calculating the impact weight of a specific event is: in, Indicates an event Total influence weight, Represents the set of influencing factors. Indicator Factors The entropy weight is determined by the degree of variation; the greater the variation, the higher the weight. Indicator Factors In the event Standardized values ​​on the range (dimensions eliminated by range standardization, range [0,1]). Indicator Factors Information entropy Indicates an event In factors The original observations (e.g., the weight preset value of "city building event" type is 0.7). Represents a set of events. Indicates an event In factors Standardized values The proportion of the total standardized values ​​of all events for this factor.

[0015] S202. Based on the key event anchor point library, using the Kalman filter algorithm and tree-ring chronology data, through time ambiguity correction and absolute dating of building wooden components, an anchor point sequence with timestamp calibration is generated. The Kalman filter algorithm is used to address the ambiguity in historical event time records. For inaccurate time descriptions commonly found in archival documents and oral histories (such as "mid-to-late Guangxu period" or "late 1980s"), a Kalman filter model is constructed, incorporating a state equation and an observation equation. The state equation establishes time evolution rules based on the temporal logical relationships of historical events (such as the time interval between policy promulgation and implementation, and the seasonal characteristics of disaster events). The observation equation transforms the ambiguous time range into probability distribution observations. Through an iterative prediction-update process, existing precise time events in a key event anchor point library are integrated as prior constraints, outputting the optimal estimate and confidence interval of the ambiguous event's occurrence time (e.g., revising "street widening in the early Republic of China" to 1913±2 years, with a confidence level of 98%), thus achieving quantitative correction of time ambiguity.

[0016] Dendrochronological data and absolute dating of building timber components provide physical evidence to support the anchor point sequence: Timber component samples (such as main beams, eaves columns, and door and window frames) from existing traditional buildings in the core area of ​​the old city were collected. After sample preprocessing (removing decay layers and surface polishing), a high-resolution dendrochronology analyzer was used to obtain the tree ring width sequence. This sequence was cross-dated with a regional standardized dendrochronology (such as a centennial-scale chronology based on local ancient trees). The felling date of the timber components was determined through tree ring matching, thereby inferring the construction or major renovation time of the buildings (with an error controlled within ±1 year). These precise time points supported by physical evidence serve as supplementary anchor points, which, together with event anchor points extracted from literature, constitute a multi-source verification anchor point set.

[0017] Finally, the precise time anchors obtained from tree-ring dating were used as the observation input to the Kalman filter model to further optimize the time correction results for fuzzy events, forming a timestamp-calibrated anchor sequence. This sequence covers three time scales: century, decade, and interannual.

[0018] S203. Based on the anchor point sequence after the timestamp calibration, the Morlet wavelet function and cross-validation method are used to generate a three-level time slice containing a macro slice at the centennial scale, a meso slice at the decadal scale, and a micro slice at the interannual scale through time series decomposition. The time series decomposition uses the calibrated anchor point sequence as a benchmark to transform the state change indicators (such as the building fabric integrity index, street network connectivity, and landscape coverage) of various spatial elements (buildings, streets, and landscapes) in the holographic historical spatial database into multivariate time series. Each variable corresponds to the dynamic evolution trajectory of a type of element. The Morlet wavelet function, as the basis function of continuous wavelet transform, effectively captures the pattern evolution characteristics at different time scales due to its excellent temporal and frequency resolution. By calculating the inner product of the time series and the Morlet wavelet basis, the wavelet coefficients at different scale parameters at each time point are obtained. The larger the coefficient modulus, the more significant the periodic signal at that scale. To address the need for three-level time slices, the scale parameter ranges are defined as follows: The centennial-scale macro-slice corresponds to scale a∈[50,100] (unit: years), capturing long-term pattern changes driven by events such as dynastic changes and major wars; the decadal-scale meso-scale slice corresponds to a∈[5,10], reflecting medium-term changes such as urban planning adjustments and infrastructure construction; and the interannual-scale micro-slice corresponds to a∈[1,3], focusing on short-term fluctuations such as disaster damage and localized repairs. Cross-validation is used to optimize the scale division threshold: a 5-fold cross-validation method is employed, dividing the anchor point sequence into 5 subsets. The scale recognition model is trained using 4 subsets in turn, and the remaining subsets are used for validation. By calculating the time series reconstruction error RMSE under different scale thresholds, the parameter combination with the smallest reconstruction error is selected to ensure that the three-level slices not only match the temporal distribution of key events but also accurately reconstruct the evolution cycle of spatial patterns. The final generated three-level dynamic time slices contain spatial element status data and event-related information within their corresponding time range for each slice.

[0019] S204. Calibrate the three-level time slices; The calibration of the third-level time slice is divided into two dimensions: spatial calibration and temporal calibration. The slice accuracy is optimized through cross-validation using multi-source reference data. Spatial calibration dimension: Based on historical measured maps (such as the "Detailed Map of XX Old City" from the 1930s and the urban construction survey map from the 1950s), the affine transformation algorithm is used to geometrically correct the coordinates of spatial elements in the slice. First, common feature points in the slice and the base map (such as the center points of city gates, ancient bridges, and landmark buildings) are selected to establish a coordinate transformation model, unifying the spatial data of the slice into the coordinate system of the base map (such as the Beijing 54 coordinate system). The spatial overlap of the corrected slice is calculated. If the IoU (Intersection over Union) with the base map is less than 0.95, the transformation parameters are repeatedly adjusted until the standard is met, ensuring that the spatial position error of the building texture and street network is ≤5 meters.

[0020] Time calibration dimension: Integrating archaeological stratigraphic chronology data with remote sensing image time series analysis results, the accuracy of the time boundaries of the slices is verified. For example, for macro-scale slices at the centennial scale, the start and end times of the slices are adjusted by comparing the chronology of the cultural layers excavated within the slice's coverage period (such as the carbon-14 dating results of Ming and Qing dynasty strata) to ensure that the slice's time range matches the chronological intervals of the cultural layers by ≥90%. For micro-scale slices at the interannual scale, aerial remote sensing image sequences from 1960 to 2020 are used to analyze whether the changes in architectural texture in the slices are consistent with the actual renovation time in the images. If the time deviation exceeds 2 years, the time nodes of the slices are corrected based on the image timestamps.

[0021] Error assessment and iterative optimization: The deviation between the calibrated slice and the benchmark data is quantified using the root mean square error (RMSE). A spatial RMSE ≤ 3 meters and a temporal RMSE ≤ 1 year are considered acceptable. If these standards are not met, the process returns to step S203 to adjust the scale parameter of the Morlet wavelet function or the number of folds in the cross-validation, regenerates the slice, and repeats the calibration process until the accuracy requirements are met. The final output, a calibrated three-level dynamic time slice, accurately reflects the true evolution of the old city's historical layout at different time scales.

[0022] S205. Based on the calibrated three-level time slices, the event pattern transmission path is analyzed using a spatiotemporal graph convolutional network. Through spatiotemporal correlation identification, a three-level dynamic time slice is formed, which includes macroscopic slices at the centennial scale, mesoscopic slices at the decadal scale, and microscopic slices at the interannual scale.

[0023] The Spatiotemporal Graph Convolutional Network (ST-GCN) takes calibrated three-level time slices as input to construct a dynamic graph structure that integrates spatial topology and time series: First, the old city space is divided into several basic units (such as architectural texture units, street and alley nodes, and historical anchor points) as nodes of the graph. The node features include the spatial attributes (coordinates, area, type) and temporal attributes (state change values ​​in each slice) of the corresponding unit; the adjacency relationships of the base space (such as physical connectivity between units and functional dependence) are used to construct the edges of the graph, and the edge weights are dynamically adjusted according to the distance and functional dependence between units. The network structure adopts a combination of "spatial convolutional layer + temporal convolutional layer + attention mechanism": the spatial convolutional layer uses graph convolution operators (such as ChebNet convolutional kernels) to extract spatial correlation features between nodes and capture the mutual influence of different spatial units within the same time slice; the temporal convolutional layer uses 1D convolution or gated recurrent units (GRU) to process the node feature sequences of continuous time slices and explore the temporal dependencies of pattern evolution; the attention mechanism module sets weights for slices at different time scales, assigning higher long-term trend attention weights to slices at the centennial scale and strengthening local attention to short-term events to slices at the interannual scale, ensuring that the network focuses on key spatiotemporal nodes.

[0024] Spatiotemporal correlation identification, based on the output features of ST-GCN, is achieved through the following steps: 1) Event impact range extraction: Cluster analysis (such as K-means algorithm) is performed on the node features output by the network to identify spatial unit clusters affected by the same key event and determine the direct impact range of the event; 2) Transmission path construction: The feature similarity and time delay between clusters are calculated to construct the transmission path of the event from the source node to the target node (such as the chain path of "policy promulgation → street and alley widening → building cluster adjustment"), and the transmission strength is quantified (cosine similarity is used to represent the correlation degree of node features on the path); 3) Multi-scale correlation fusion: For three-level slices The analysis extracts macroscopic transmission at a century-scale (e.g., the impact of dynastic changes on the overall urban landscape), mesoscopic transmission at a decade-scale (e.g., the radiation path of infrastructure construction on surrounding blocks), and microscopic transmission at an interannual scale (e.g., the impact path of local renovations on adjacent buildings), and integrates these paths into corresponding time slices. 4) Correlation verification: By comparing documentary records in the holographic historical spatial database with field survey data (e.g., the chronological order of building renovations and the spatial diffusion trajectory of street and alley expansion), the accuracy of the transmission paths is verified. If the verification accuracy is below 90%, the ST-GCN's attention weights or node feature dimensions are adjusted, and the analysis is repeated. Finally, the three-level dynamic time slices of the spatiotemporal transmission paths are integrated to fully present the evolution process and internal correlation mechanism of the old city's historical landscape at different time scales.

[0025] As an optional embodiment of the present invention, optionally, in step S3, based on the three-level dynamic time section, a feature extraction CNN and graph neural network are used to construct an element gene map, and a historical layered element map is generated through building texture segmentation, street network topology analysis, and landscape pattern field simulation, including: S301. Based on the three-level dynamic time section, using a CNN feature extractor, and through transfer learning, the data overlaid with remote sensing images and historical maps is processed to generate three classification results: Category A: typical traditional buildings; Category B: suspected traditional buildings; and Category C: modern buildings. The CNN feature extractor adopts an improved architecture based on ResNet50. The weights pre-trained on the ImageNet dataset provide a rich foundation of general visual features for feature extraction. The transfer learning process is divided into two stages: The first stage is feature adaptation, which overlays remote sensing images (RGB three channels) with rasterized single-channel data after vectorization of historical maps to form a four-channel input tensor (uniform size of 256×256 pixels). The first 17 convolutional layers of ResNet50 are frozen, and only the three newly added fully connected layers (including dropout layers to prevent overfitting) are trained, outputting a three-class probability distribution. The second stage is fine-tuning, which unfreezes the last three convolutional layers of ResNet50 and trains them together with the top fully connected layer to learn the unique textures (such as the eaves, blue bricks and gray tiles of traditional buildings), morphology (such as courtyard layout), and annotation features in historical maps (such as the words "temple" and "residential building"). The cross-entropy loss function was used during training, and the Adam optimizer was selected (with an initial learning rate of 1e-4, decaying by 10% every 5 epochs). The training and validation sets were split in a 7:3 ratio, and supervised training was conducted using labeled data from a field survey of the old city (including 1200 typical traditional buildings (Category A), 800 suspected traditional buildings (Category B), and 1500 modern buildings (Category C). The classification results were output as pixel-level masks, with Category A regions marked in red, Category B in yellow, and Category C in blue. Classification performance was evaluated using a confusion matrix, requiring a recall rate of ≥95% for Category A and a precision rate of ≥85% for Category B. If these targets were not met, the Dropout rate or learning rate decay strategy was adjusted, and retraining was performed until the accuracy requirements were met. These classification results will serve as the foundational data for subsequent building texture segmentation.

[0026] S302. Based on the three-classification results, the betweenness centrality and proximity centrality of street and alley nodes are calculated using graph neural networks. The historical accessibility is verified by combining spatial syntax theory, and high-integration paths such as the north-south city axis are generated. Optionally, the expression for calculating the betweenness centrality of street nodes is: in, Represents a node betweenness centrality, Indicates street and alley nodes, and Indicates other street and alley nodes, express arrive The shortest path passes through Quantity, Represented as arrive The total number of shortest paths; Optionally, the expression for calculating proximity centrality is: in, Represents a node The proximity centrality of a node reflects its average reachability to other nodes. The variable representing the total number of nodes. Represents the path length variable, node arrive The shortest path distance; The graph neural network employs a Graph Attention Network (GAT) architecture, constructing a dynamic graph based on a street and alley network within a three-level time slice. Street and alley intersections, endpoints, and historical landmarks (such as city gates and post stations) are used as nodes in the graph. Node features include the physical attributes of the street and alley (width, material, length), historical functional labels (commercial, government office, residential), and state changes in the corresponding time slice (such as expansion or abandonment). Street and alley segments are used as edges in the graph, with edge weights dynamically assigned based on the street and alley's accessibility (width × road surface smoothness index). The GAT model learns the dependencies between nodes through a multi-head attention mechanism. When calculating betweenness centrality, it utilizes the statistical results of the shortest path weighted by attention weights to enhance the identification of key nodes. The calculation of proximity centrality combines attention weights to correct the path distance between nodes, making the results more consistent with the actual accessibility in historical scenarios (e.g., considering the pedestrian priority characteristics of traditional streets and alleys, assigning lower weights to non-pedestrian streets and alleys).

[0027] The application of spatial syntax theory focuses on the calculation and verification of integration index: First, based on the node centrality results output by the graph neural network, candidate high integration paths are extracted; then, the axis analysis method of spatial syntax is used to transform the street network into an axis model, and the global integration (reflecting accessibility in the entire urban network) and local integration (reflecting accessibility in the surrounding area) of each axis are calculated; by weighting and fusing the centrality score of the graph neural network with the integration value of spatial syntax (the weight is determined by the importance of street functions in historical documents, such as setting the weight of official post streets and alleys to 1.2 and ordinary residential streets and alleys to 0.8), the final set of high integration paths is generated. For macro-level slices on a 100-year scale, the focus is on verifying the stability of the global integration degree of historical axes such as the North-South City Axis during dynastic changes (requiring the integration degree ranking to remain in the top 5% for three consecutive dynasties); for meso-level slices on a 10-year scale, the impact of urban planning adjustments on local integration degree is analyzed (e.g., a local integration degree increases by ≥20% after the widening of a certain street or alley); for micro-level slices on an interannual scale, the short-term fluctuations of node centrality caused by disasters or repairs are captured (e.g., a node's betweenness centrality decreases by ≥15% due to damage).

[0028] The evaluation process involves two dimensions: first, documentary verification, comparing the spatial overlap between the high-integration paths generated by the model and the paths recorded in the documentary records with historical records in the holographic historical spatial database, requiring an overlap of ≥88%; second, field verification, matching the street and alley remains (such as pavement layers and drainage systems) from archaeological excavations with the path features output by the model (such as width and orientation), requiring a feature matching rate of ≥85%. If the requirements are not met, the number of attention heads in the GAT model (currently set to 4, but can be adjusted to 6) or the weighting coefficient of the spatial syntactic integration is adjusted, and the model is recalculated and verified again until the accuracy requirements are met. The final generated high-integration paths, such as the North-South City Axis, clearly present the core framework of the old city's historical accessibility and its evolutionary patterns.

[0029] S303. Based on the three-level dynamic time section, a three-dimensional terrain model is constructed using DEM elevation data and water system vector data. The old city structure is simulated by the Viewshed field of view analysis algorithm, and the viewpoint position is dynamically adjusted to generate a field of view coverage heat map. The expression for the Viewshed view analysis algorithm is: in, This indicates the view coverage heatmap in coordinates The value at that location, Represents the set of all viewpoints. This indicates the result of the binary visibility determination. Indicate viewpoint The weighting coefficients, Indicates from the viewpoint To the target point The discrete set of points along the line of sight. Representing path points DEM elevation values ​​at the location, Indicates the line of sight is at the point. Vertical height increment at that location Indicate viewpoint The observation height (such as the height of a watchtower). Indicates the calculation of Euclidean distance; The construction of a 3D terrain model requires the integration of multi-source spatiotemporal data: First, the DEM elevation data is processed in layers. The 100-year scale slices use stratigraphic elevation reconstruction data obtained from archaeological drilling (combined with carbon-14 dating to determine time nodes). The 10-year scale slices are overlaid with terrain modification records from urban construction archives (such as elevation changes from dike construction and lake filling). The interannual scale slices are supplemented with elevation adjustment data from temporary projects (such as the temporary height of flood control dikes). The river system vector data is based on records of river course direction, width, and seasonal water level changes in historical documents. After vectorization, it is spatially registered with the DEM data of the corresponding time slices to generate a 3D mesh model that includes terrain undulations, river system distribution, and historical geomorphological features (such as city walls, hills, and ponds). For different time scales, the model dynamically updates key elements: the 100-year scale focuses on long-term terrain evolution such as river course changes and mountain erosion; the 10-year scale highlights the artificial modification of the terrain by infrastructure construction (such as road widening and bridge construction); and the interannual scale captures short-term terrain changes caused by disasters (such as floods and earthquakes).

[0030] The application of the Viewshed view analysis algorithm needs to be optimized in conjunction with historical scene characteristics: viewpoint selection is set differently according to time scales. On a century-scale, the core high points of the city (such as city gate towers, pagodas, and mountaintop observation decks) are selected; on a decade-scale, regional landmarks (such as the main gate of government offices and the main hall of temples) are selected; and on an interannual scale, temporary observation points (such as the roof of a building under renovation and construction scaffolding) are selected. The viewpoint weight coefficient is assigned according to its historical functional importance (such as setting the weight of city gate towers to 1.5 and the weight of ordinary residential roofs to 0.6). When calculating the sight path, historical vegetation occlusion correction is introduced: through the vegetation types (such as ancient locust trees and pines) and height data recorded in historical documents, the elevation increment Δh of the path points is dynamically adjusted (such as increasing Δh by 2-3m at tall trees) to ensure that the visibility determination is more in line with historical reality. When generating the view coverage heatmap, the number of visible viewpoints at each coordinate point is normalized to the range of 0-100. Red indicates high coverage area (≥80, i.e. visible to most key viewpoints), yellow indicates medium coverage area (50-79), and blue indicates low coverage area (<50), which intuitively presents the visual relationship pattern of the old city space under different time slices.

[0031] For literature verification, the spatial matching degree between the viewpoint coverage area generated by the model and the literature description is compared with the records of landscape sightlines in the holographic historical spatial database (such as "climbing the Drum Tower can overlook the entire city streets and alleys" and "the south city pond can be seen from the West Gate Tower"). The requirement is ≥90%. For field verification, the model's viewpoint settings and viewpoint results are matched with the archaeologically excavated view-related relics (such as the orientation of the lookout holes and the foundation elevation of the viewing platform). The requirement is ≥85% match rate. If the standard is not met, the interpolation accuracy of the DEM data is adjusted (such as increasing the resolution from 5m to 2m) or the viewpoint weight coefficient is adjusted, the 3D terrain model is reconstructed, and viewpoint analysis is performed until the accuracy requirements are met. The final generated 3D terrain model and viewpoint coverage heat map will be integrated with the building classification results and street integration paths to form a multi-dimensional historical stratification element map.

[0032] S304. Based on the three-classification results, the high-integration path and the view coverage heatmap, the elements are integrated into an element relationship graph model. The graph neural network is used to aggregate node features and identify functional groups to generate an element gene map containing core and edge layered structures. The construction of the element relationship graph model requires the integration of multi-source historical layered elements: Building classification areas (A / B / C categories), highly integrated street and alley nodes, and key areas of the view coverage heatmap are used as the core nodes of the graph. Node features integrate three aspects of information: building nodes carry classification labels, texture morphology features, and historical functional attributes; street and alley nodes embed betweenness / proximity centrality scores, spatial syntactic integration values, and physical attributes; view nodes include heat coverage values, associated viewpoint weights, and visibility range. Edge construction is based on a dual logic of spatial association and historical functional association: spatial association edges reflect the physical adjacency between elements (e.g., if the distance between a building and an adjacent street or alley is ≤5m, an edge is established); historical functional association edges are based on records in the holographic historical spatial database (e.g., if "temples and surrounding residences have a sacrificial movement connection," an edge is established). Edge weights are assigned by the association strength (spatial adjacency weight is set to 0.7, and historical functional association weight is set to 1.3).

[0033] The graph neural network employs a hybrid architecture combining graph convolutional networks (GCN) and graph attack vectors (GAT): First, GCN performs preliminary aggregation of the basic features of nodes to capture global spatial dependencies; then, the multi-head attention mechanism of GAT (set to 6 heads) strengthens the local correlations between key elements (e.g., the attention weight of core buildings and highly integrated streets and alleys is increased to 1.5). Functional cluster identification is achieved through clustering algorithms: based on the aggregated node features, spectral clustering is used to divide nodes into different functional clusters. Combined with historical records of functional areas (e.g., "the two sides of the north-south market axis during the Ming and Qing dynasties were commercial functional areas"), the clusters are labeled to generate functional clusters such as commercial, government offices, residences, and sacrificial areas.

[0034] In terms of literature verification, the spatial overlap between the functional clusters identified by the computational model and the functional areas recorded in the holographic historical spatial database must be ≥90%. For field verification, functional remains discovered through archaeological excavation (such as the patterns of floor tiles in shop ruins and the shape of gate piers in government offices) are matched with cluster features, requiring a feature matching rate of ≥85%. If the standard is not met, the edge weight assignment rules or the number of attention heads are adjusted, and the graph model is retrained until the accuracy requirements are met. The final generated element gene map is presented in a layered visualization: the core layer consists of long-term stable historical functional clusters (such as the core area of ​​government offices and the sacrificial center), while the peripheral layer consists of auxiliary functional areas that evolve over time (such as temporary markets and residential expansion areas), clearly revealing the layered structure and evolutionary trajectory of the old city's historical pattern.

[0035] S305. Based on the element gene map, use a spatiotemporal graph convolutional network to analyze the spatiotemporal relationship between events and elements, identify the layered transmission path, and generate a historical layered element map.

[0036] The construction of the Spatiotemporal Graph Convolutional Network (ST-GCN) requires the integration of static features of the element gene map and dynamic spatiotemporal information of historical events. It employs a spatiotemporal architecture combining Graph Convolutional Networks (GCN) and Gated Recurrent Units (GRUs). The node features of the element gene map are used as the spatial dimension input, while historical event sequences (such as dynastic changes, war destruction, commercial policy adjustments, and disasters) are used as the temporal dimension input. Historical event data is extracted from a holographic historical spatial database, including the event's timestamp, spatial impact range (e.g., administrative region, street segment), event type (political, economic, social, natural), and intensity level (1-5), which is then encoded into a temporal feature vector. ST-GCN captures the relationships between element nodes through spatial convolution and learns the temporal dependencies of events through GRU units, thus achieving spatiotemporal relationship modeling between events and elements.

[0037] The identification of stratified transmission paths first involves filtering out strong correlation pairs with a weight ≥ 0.6 using the event-feature association weight matrix output by ST-GCN; then, based on chronological order and spatial proximity, transmission chains from events to features and from features to features are constructed (e.g., "Ming Dynasty capital relocation event → relocation of the core area of ​​government offices → increased integration of the north-south city axis → enhanced functionality of surrounding commercial buildings"); finally, the transmission chains are sorted using the existing PageRank algorithm, and the top 10 core transmission paths are extracted as key evolutionary clues for historical stratification.

[0038] In an optional embodiment of the present invention, step S4 may involve identifying pattern abrupt change points and quantifying evolution patterns based on the historical layered feature map using a spatiotemporal graph convolutional network, a long short-term memory network, and a mutation detection algorithm, including: S401. Based on the historical layered element map, the spatiotemporal correlation features of building texture, street network and landscape pattern are extracted using spatiotemporal graph convolutional network. The information of neighboring nodes is aggregated by spatial graph convolution and the receptive field is expanded along the time dimension to generate a spatiotemporal feature tensor containing macro trends at the centennial scale, meso fluctuations at the ten-year scale and micro changes at the interannual scale. In this embodiment, the spatiotemporal graph convolutional network adopts a two-layer architecture of "spatial graph convolutional layer + multi-scale temporal recurrent layer": the spatial graph convolutional layer designs differentiated neighborhood aggregation rules for three types of nodes: architectural texture, street network, and landscape pattern. The neighborhood of architectural texture nodes includes adjacent buildings within a radius of 10m and directly connected street nodes. When aggregating, architectural morphological features (such as courtyard layout and roof type) are given higher weight (coefficient 1.2). The neighborhood of street network nodes is the street segments connected upstream and downstream and building nodes within 5m on both sides. The focus is on aggregating spatial syntax integration degree and centrality score (weight 1.1). The neighborhood of landscape pattern nodes (rivers, hills, ponds) covers the terrain nodes within a 20m radius and related bridge and dam nodes. The DEM elevation difference and water system connectivity features are preferentially integrated (weight 1.0). The spatial convolution kernel adopts an adaptive size design. The building node uses a 3×3 convolution kernel to capture local texture, the street and alley node uses a 1×5 convolution kernel to adapt to linear extension features, and the landscape node uses a 5×5 convolution kernel to cover wide-area terrain associations.

[0039] The multi-scale temporal recurrent layer is implemented based on gated recurrent units (GRUs), and differentiated temporal sampling strategies are set for three-level time sections: macro trends at the centennial scale use a time step of 50 years, inputting long-term evolution characteristics of nodes (such as changes in the proportion of building types and adjustments in the topology of street networks); meso-level fluctuations at the decadal scale use a time step of 10 years, inputting changes in elements brought about by urban planning or policy adjustments (such as fluctuations in the integration degree after street widening); and micro-level changes at the interannual scale use a time step of one year, inputting the impact data of short-term events such as disasters and repairs (such as the building damage rate after floods). A multi-scale attention mechanism is used to weight and fuse features from different time steps, with a weight of 0.6 for macro trends, 0.3 for meso-level fluctuations, and 0.1 for micro-level changes, ensuring that the spatiotemporal feature tensor simultaneously captures long-term patterns and short-term details.

[0040] The spatiotemporal feature tensor is designed with the dimensions of "total number of nodes × feature dimension × time steps". The total number of nodes covers all building, street, and landscape nodes (approximately 12,000). The feature dimension integrates the spatially aggregated morphological, topological, and terrain features (64 dimensions in total). The time steps include a fusion sequence of 10 steps at a 100-year scale, 50 steps at a 10-year scale, and 200 steps at an interannual scale (a total of 260 steps). The resulting spatiotemporal feature tensor will serve as the core input for subsequent identification of abrupt change points and quantification of evolution patterns.

[0041] S402. Based on the spatiotemporal feature tensor, a long short-term memory network is used to capture the long-term dependencies of historical patterns. Key evolution information is filtered and noise interference is suppressed through a gating mechanism to generate a hidden state sequence that reflects the dynamic evolution of the pattern. This embodiment employs a two-layer stacked Long Short-Term Memory (LSTM) network architecture, with 128 hidden units in each layer to enhance the ability to capture long-term dependencies in historical patterns. The gating mechanism is designed with differentiated optimization for features at different time scales: the forget gate calculates the weight for retaining historical information using a sigmoid activation function, assigning a high weight of 0.8 to the topological changes in building fabric at the centennial scale to retain core pattern features, and a low weight of 0.3 to filter noise from the impact of temporary disasters at the interannual scale; the input gate controls the inclusion of new features, setting the activation threshold to 0.5 for features corresponding to key events such as policy adjustments and major construction projects; if the threshold is exceeded, the new feature is integrated into the current hidden state; the output gate combines the spatiotemporal features of the current time step with the historical hidden state, generating the hidden state output at the current moment using a tanh activation function, while introducing layer normalization to stabilize the training process.

[0042] The generation process of the hidden state sequence needs to be aligned with the time dimension of the spatiotemporal feature tensor: the spatiotemporal feature tensor is input into the LSTM network in time step order, and the hidden state of each step integrates the architectural, street, and landscape features of the current time point as well as the cumulative evolution information of the previous time steps. For example, when processing the time step corresponding to the adjustment of the Qing Dynasty's canal transport policy, the hidden state will integrate the water system connectivity features before the policy and the changes in street integration after the policy, forming a continuous pattern evolution trajectory. To avoid overfitting, dropout layers (dropout rate set to 0.2) are added between LSTM layers, and the Adam optimizer (learning rate 0.001) is used to minimize the mean squared error (MSE) loss function. Iterative training stops when the validation set loss does not decrease for 5 consecutive rounds. The final generated hidden state sequence has a dimension of 260×128 (number of time steps × number of hidden units), which fully preserves the dynamic evolution information of the pattern from a century-long to an annual scale.

[0043] S403. Based on the hidden state sequence, a combination strategy of CUSUM mean detection, Pelt multi-mutation point algorithm and Chow structural fracture test is used to identify the set of pattern mutation points of macro-mutation on a century-scale, meso-level turning point on a decade-scale and micro-level disturbance on an interannual scale through cumulative sum threshold judgment, penalty term optimization segmentation and segmented regression significance test. In step S403, the CUSUM mean detection is based on the principle of cumulative sum control charts and is used to capture significant abrupt changes in the mean of the latent state sequence. Differentiated parameters are set for different time scales: in the 100-year macro-trend analysis, a 10-step sliding window is used to calculate the cumulative sum of the latent state sequence mean, with a threshold set to 3 times the sequence standard deviation (σ); in the 10-year meso-level fluctuation analysis, the sliding window is set to 5 steps, with a threshold of 2.5σ; and in the interannual micro-level change analysis, the sliding window is set to 2 steps, with a threshold of 2σ. The specific operation is as follows: for each time step of the latent state sequence, the cumulative sum of the deviations between the current time and the mean of the preceding window is calculated. When the absolute value of the cumulative sum exceeds the threshold of the corresponding scale, that time step is marked as a potential abrupt change point. For example, in the 100-year scale analysis, the cumulative sum of the deviations of the latent state mean at the time step corresponding to the Ming-Qing dynastic transition exceeds 3σ, thus it is marked as a potential point of macro-level pattern change.

[0044] The Pelt multi-mutation point algorithm employs a dynamic programming framework to efficiently identify multiple mutation points in a sequence. Penalty terms are set for different time scales: 50 for a century-scale, 30 for a decade-scale, and 10 for an interannual scale (the penalty term value is negatively correlated with the intensity of sequence volatility). The algorithm iteratively calculates the cost function (including the intra-sequence variance and the penalty term) for each possible split point, selecting the splitting method with the minimum cost to obtain a candidate set of mutation points. For example, in a decade-scale analysis, time steps such as the Qing Dynasty's canal transport policy adjustments and the expansion of streets and alleys during the Republic of China era are used as candidate splitting points. Through cost optimization, three key meso-level turning point mutations are ultimately retained.

[0045] The Chow structural break test is used to verify whether candidate mutation points are significant structural breakpoints. The hidden state sequence is segmented according to candidate mutation points, and a piecewise linear regression model is constructed: with the time step as the independent variable and the hidden state value as the dependent variable, regression equations are fitted before and after the mutation point. The F-statistic is calculated (to test whether there is a significant difference between the regression coefficients of the two segments). When the F-value is greater than the critical value at the α=0.05 level, the point is confirmed as a structural breakpoint. For example, when testing the time step corresponding to the Opium War, the F-value is significantly greater than the critical value, verifying it as a macro-level structural breakpoint on a century-long timescale.

[0046] The combined strategy employs a three-step process of "screening-optimization-verification": First, all potential mutation points are quickly screened using CUSUM mean detection; second, the Pelt algorithm is used to optimize and segment the potential points, removing redundant candidates; finally, the significance of the candidate points is verified using the Chow test, resulting in the final set of pattern mutation points. For example, 15 potential points are obtained through CUSUM, 8 candidates are retained after Pelt optimization, and then 5 effective mutation points are confirmed by the Chow test (including 2 centennial macro-mutations, 2 ten-year meso-level turning points, and 1 interannual micro-level disturbance).

[0047] The cumulative sum threshold is dynamically adjusted to assess the volatility of latent state sequences at different time scales: for centennial sequences with moderate volatility, the threshold is set to 3σ to reduce false positives; for decadal sequences with moderate volatility, the threshold is set to 2.5σ; and for interannual sequences with dramatic volatility, the threshold is set to 2σ to capture short-term disturbances. The determination rule is: when the absolute value of the cumulative sum exceeds the threshold for three consecutive time steps, the starting time step of that interval is marked as a breakout point.

[0048] In optimizing segmentation using penalty terms, the penalty term setting needs to balance the number of mutation points and segmentation accuracy: for centennial scales, due to the shorter sequence length (10 steps), the penalty term is set to 50 to avoid over-segmentation; for decadal scales, with a medium sequence length (50 steps), the penalty term is set to 30; and for interannual scales, with a longer sequence length (200 steps), the penalty term is set to 10 to allow for more micro-perturbation points. During the segmentation optimization process, priority is given to retaining candidate points corresponding to key events recorded in historical documents, such as time steps of disaster occurrence and policy promulgation.

[0049] For each candidate mutation point, after constructing a piecewise regression model, the t-test value of the regression coefficient and the R² value of the model are calculated. If the absolute value of the t-test is greater than 2 (α=0.05) and the R² increases by more than 15%, the mutation point is considered significant. For example, for the time step corresponding to the relocation of the capital in the Ming Dynasty, the R² increased from 0.6 to 0.8 after piecewise regression, and the t-test value was 2.3, confirming it as a significant mutation point.

[0050] Through the above combination strategy, a complete set of pattern change points covering a century-long timescale (such as dynastic changes and major wars), a decade-long timescale (such as policy adjustments and urban planning), and an interannual timescale (such as disasters and major renovations) was finally identified.

[0051] S404. Based on the set of abrupt change points in the pattern, the evolution rate is quantified by trend slope analysis, periodic fluctuations are identified by Morlet wavelet transform, and the abrupt change intensity is calculated by Euclidean distance. A quantitative evolution law is generated that includes the spatiotemporal location of the abrupt change points, the direction of the evolution trend, periodic characteristic parameters, and the abrupt change intensity index.

[0052] Trend slope analysis segments the latent state sequences before and after abrupt change points in the pattern, using the least squares method to calculate the linear trend slope to quantify the rate of evolution. For centennial-scale abrupt changes (such as the Ming-Qing transition), five time steps of the latent state sequence are taken before and after each point. An absolute slope value ≥0.8 indicates a fast rate of evolution (e.g., a rapid decline in street network integration), 0.4-0.8 indicates a moderate rate, and <0.4 indicates a slow rate. For decadal-scale abrupt changes (such as adjustments to the Qing Dynasty's Grand Canal policy), three time steps are taken before and after each point, with slope thresholds set to 1.0, 0.5, and 0.2, respectively. For interannual-scale abrupt changes (such as floods), two time steps are taken before and after each point, with slope thresholds set to 1.2, 0.6, and 0.3. A positive slope indicates an expanding trend in pattern elements (such as commercial building density), while a negative slope indicates a contracting trend.

[0053] Morlet wavelet transform captures periodic fluctuations at different time scales by adjusting the wavelet scale. For time series of spatiotemporal feature tensors, the wavelet scale range is set to 1-50 (corresponding to 1-50 year cycles): the centennial scale focuses on identifying long cycles of 50-100 years (such as dynastic change cycles), the decadal scale identifies medium cycles of 10-20 years (such as policy adjustment cycles), and the interannual scale identifies short cycles of 1-5 years (such as disaster occurrence cycles). By calculating the wavelet energy spectrum, the periodic values ​​corresponding to the peaks are extracted as the main periodic feature parameters. For example, performing a Morlet transform on the street network evolution sequence of an old city reveals a 20-year medium-term periodic fluctuation, corresponding to street widening policies that occur every 20 years.

[0054] When calculating the intensity of a catastrophe using Euclidean distance, the hidden state vectors at one time step before and after the catastrophe point are used as input to calculate the Euclidean distance between the vectors. The intensity levels of catastrophes on a centennial scale are categorized as follows: distance > 5.0 indicates a strong catastrophe (such as a reshaping of the political landscape due to war), 3.0-5.0 indicates a medium catastrophe (such as the relocation of the core area due to the relocation of the capital), and < 3.0 indicates a weak catastrophe. On a decadal scale, distance > 4.0 indicates a strong catastrophe, 2.0-4.0 indicates a medium catastrophe, and < 2.0 indicates a weak catastrophe. On an interannual scale, distance > 3.0 indicates a strong catastrophe, 1.0-3.0 indicates a medium catastrophe, and < 1.0 indicates a weak catastrophe. For example, the Euclidean distance between the hidden state vectors before and after the catastrophe point corresponding to the Opium War is 6.2, and it is classified as a strong catastrophe on a centennial scale.

[0055] When generating quantitative evolution patterns, it is necessary to integrate various analytical results: Spatiotemporal positioning should clearly define the specific time stamp of the abrupt change point (e.g., the Ming-Qing transition in 1644) and its spatial impact range (e.g., the entire core area of ​​the old city); the trend direction should be determined by the slope (e.g., a slope of -0.9 indicates a decrease in the integration of the core area); periodic characteristic parameters should record the main fluctuation cycles (e.g., a 20-year cycle); and abrupt change intensity indicators should be used to label the level (e.g., strong abrupt changes). The final output of quantitative patterns can provide data support for old city protection planning, for example, by formulating targeted pattern protection strategies for historical events corresponding to strong abrupt change points.

[0056] As an optional embodiment of the present invention, optionally, in step S5, based on the evolutionary law, a multi-criteria verification system and spatial autocorrelation analysis are used to verify the historical authenticity through three pieces of evidence: literature, physical objects, and word-of-mouth; and spatial continuity analysis is used to assess the integrity of the texture. The verification results obtained include: S501. Based on the spatiotemporal location of the mutation point in the evolution law, a multi-criteria verification system is used to generate a historical authenticity verification result containing triple evidence support score, spatiotemporal alignment degree and consistency coefficient by calculating the TF-IDF support of literature evidence, verifying the consistency of physical evidence through chi-square test and aligning the Kappa coefficient of reputation evidence. The multi-criteria verification system is centered on "triple evidence verification," ensuring consistency between evolutionary patterns and historical facts through cross-verification of documentary, physical, and oral evidence. The steps for calculating the TF-IDF support of documentary evidence are as follows: First, a keyword database for the target mutation point is constructed (e.g., for the mutation point "damage in the 1937 war," keywords include "war damage," "building collapse," and "street blockage"); second, relevant documents such as local chronicles, historical archives, and academic works are collected, each document is segmented, and the product of the term frequency (TF) and inverse document frequency (IDF) of the keywords is calculated to obtain the support of a single document; finally, the average support of all relevant documents is taken as the documentary evidence support for that mutation point (≥0.7 is high support, 0.4-0.7 is medium support, and <0.4 is low support). Consistency Verification of Physical Evidence via Chi-square Test: For spatial pattern changes corresponding to abrupt changes (such as the proportion of building types and adjustments in street width), physical evidence such as archaeological reports and architectural remains survey data is collected. A contingency table of pattern characteristics before and after the abrupt change is constructed (e.g., "the proportion of government buildings was 15% before the abrupt change, and 5% after the abrupt change"). The chi-square test is used to calculate the χ² value. If χ² > 3.841 (α = 0.05, 1 degree of freedom), the physical evidence is considered consistent with the pattern changes in the evolutionary pattern, and the consistency coefficient is assigned a value of 1; otherwise, it is 0. Kappa Coefficient Alignment of Oral Evidence: Oral history and folk records of long-time residents are encoded, and event descriptions related to the abrupt change are extracted (e.g., "the flood in 1954 destroyed the South Gate"). The encoded results are matched with the abrupt events in the evolutionary pattern, and the Kappa coefficient is calculated to measure the consistency between the two (Kappa ≥ 0.8 indicates high alignment, 0.6-0.8 indicates medium alignment, and < 0.6 indicates low alignment). Ultimately, the historical authenticity verification results integrate the support scores of the three types of evidence (a weighted average of scores from documents, physical objects, and word of mouth, with weights of 0.4, 0.3, and 0.3, respectively), spatiotemporal alignment (a time difference of ≤5 years between the timestamp of the mutation point and the document / physical record is considered high alignment, 5-10 years is considered medium alignment, and >10 years is considered low alignment) and consistency coefficient (a comprehensive score of the consistency results of the three).

[0057] S502. Based on the historical authenticity verification results, the spatial autocorrelation analysis model is used to calculate spatial continuity through the global Moran's I index, identify local LISA hotspots, and aggregate the comprehensive scores using the entropy weight method to generate a texture integrity assessment result that includes spatial continuity index, type diversity index, and age gradient index. The expression for calculating spatial continuity using the global Moran's I exponent is: in, Represents the global Moran's I index, quantifying the spatial continuity of textural features. This indicates the total number of spatial units (such as architectural texture units and street / alley node units). Represents the elements in the spatial weight matrix With unit The weight value (based on adjacency or distance decay). and It represents the textural characteristics (such as building age, building type, street width, etc.). It represents the mean of the texture characteristics, reflecting the overall average level.

[0058] Existing spatial autocorrelation analysis models focus on combining global and local features, quantifying the spatial organization patterns of old city fabric through multi-dimensional indicators. The global Moran's I index ranges from -1 to 1: when the index approaches 1, it indicates that the fabric features (such as building age and street width) exhibit a significant spatial clustering pattern, with fabric units of the same type or age tending to be distributed adjacently; when the index approaches -1, it exhibits a spatial discrete pattern, with fabric units distributed randomly and exhibiting strong heterogeneity; when the index approaches 0, it is spatially random. For example, the global Moran's I index for the building age fabric of a certain old city core area is 0.72, indicating that buildings of the same age in this area show obvious clustering, reflecting the characteristics of historical large-scale construction.

[0059] Local LISA hotspot identification, based on local Moran's I statistics, can accurately locate spatial clusters of textural features: High-High hotspots are characterized by high textural feature values ​​for themselves and surrounding units (e.g., areas with concentrated wide streets and alleys); Low-Low coldspots are characterized by low textural feature values ​​for themselves and surrounding units (e.g., traditional residential areas with narrow streets and alleys); High-Low anomalies are characterized by high textural feature values ​​for themselves but low values ​​for surrounding units (e.g., modern buildings embedded in traditional areas); and Low-High anomalies are the opposite. For example, local LISA analysis revealed that the Qing Dynasty commercial district in the old city is a High-High hotspot (wide streets and alleys surrounded by wide streets and alleys), while the scattered residential areas from the Republican era are High-Low anomalies (single modern buildings embedded in traditional narrow streets and alleys). These identification results can directly guide the delineation of key areas in conservation planning, such as prioritizing the preservation of clustered features in hotspot areas.

[0060] The entropy weighting method is used to aggregate comprehensive scores to integrate multi-dimensional texture indicators: First, the spatial continuity index (global Moran's I), type diversity index (Shannon entropy, which calculates the richness of building types), and chronological gradient index (the average difference in age between adjacent building units, reflecting the smoothness of chronological transition) are standardized (the index values ​​are mapped to the [0,1] interval); second, the information entropy of each indicator is calculated. The smaller the entropy value, the higher the indicator's distinguishability and the greater its weight (e.g., if the entropy value of the type diversity index is 0.3, the weight is set to 0.4; if the entropy value of the spatial continuity index is 0.4, the weight is set to 0.3; if the entropy value of the chronological gradient index is 0.5, the weight is set to 0.3); finally, the standardized indicators are multiplied by their corresponding weights and summed to obtain the comprehensive score for texture integrity. A score ≥ 0.8 is judged as a "complete texture area" that requires strict protection; 0.6-0.8 is a "relatively complete texture area" that can be moderately updated; < 0.6 is a "fragmented texture area" that requires restorative treatment. For example, a historical and cultural district in an old city scored 0.85 overall and was classified as a complete urban fabric area. The planning prohibits large-scale demolition and construction, and only allows localized micro-renewal.

[0061] Through the above spatial autocorrelation analysis and comprehensive score aggregation, the final texture integrity assessment results can be combined with the historical authenticity verification results. Authenticity protection can be implemented for areas that are historically authentic and have intact textures, restoration can be carried out for areas that are historically authentic but have fragmented textures, and historical scene restoration can be performed for areas with intact textures but low historical authenticity.

[0062] S503. Based on the texture integrity assessment results, a spatiotemporal graph convolutional network is used to perform a correlation analysis of the verification results. Anomalies are identified through spatiotemporal feature aggregation. The historical layered feature map is corrected by combining a feedback mechanism to generate a verification result that includes verification correction parameters, anomaly location, and correction strategy.

[0063] The Spatiotemporal Graph Convolutional Network (ST-GCN) constructs a spatial graph structure using old city spatial units as nodes and spatial adjacency relationships as edges. It simultaneously inputs historical layered sequences as the temporal dimension, capturing spatial dependencies of textural features through spatial convolution and extracting temporal dynamic features of evolutionary patterns through temporal convolution. In the spatiotemporal feature aggregation stage, the historical authenticity verification score of S501 (e.g., triple evidence support) and the textural integrity indicators of S502 (e.g., spatial continuity index, type diversity index) are embedded as node features, fusing the spatiotemporal correlation information through multi-layer convolution operations. Verification anomaly identification is based on the prediction residuals of the network output: when the absolute value of the residual between the predicted verification result (e.g., the matching degree between historical authenticity and textural integrity) and the actual calculation result of a spatial unit exceeds 0.2, or when the local LISA hotspot type of the unit does not match the functional area recorded in historical documents (e.g., predicted as a High-High wide street hotspot but recorded as a traditional narrow street area in the literature), it is marked as a verification anomaly. When the feedback mechanism corrects the historical stratified feature map, it traces back to the quantitative evolution of outliers in S404 as follows: If the outlier originates from a mutation point identification bias, the t-test threshold or penalty term parameters of the piecewise regression are readjusted; if it originates from a texture feature extraction error, supplementary architectural remains survey data or oral history data for the area are collected, and the texture attribute values ​​of spatial units (such as building age and street width) are updated. The final validation results include: validation correction parameters (such as adjusted mutation point time steps and texture feature weights), spatial coordinates and timestamps of outliers, classification of deviation causes (such as missing data or unreasonable parameter settings), and correction strategies (such as conducting supplementary field surveys for outliers with missing data and re-optimizing model parameters for outliers with parameter deviations). These corrected results will be synchronously updated to the historical stratified feature map.

[0064] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for identifying the historical pattern of old towns based on time sections and historical layers, characterized in that, The method includes: Historical maps, remote sensing images, documentary archives and oral history data are collected and preprocessed, and a holographic historical spatial database is formed through georegistration and timestamp calibration. Based on the holographic historical spatial database, a three-level dynamic time slice is generated by using the key event anchor point extraction algorithm and the wavelet transform periodic slicing algorithm, which includes macro slices at the centennial scale, meso slices at the decadal scale, and micro slices at the interannual scale. Based on the three-level dynamic time section, a feature gene map of elements is constructed using a CNN feature extractor and a graph neural network. A historical layered feature map is generated through building texture segmentation, street and alley network topology analysis and landscape pattern field simulation. Based on the historical layered feature map, spatiotemporal graph convolutional networks, long short-term memory networks, and mutation detection algorithms are used to identify pattern mutation points and quantify evolution patterns. Based on the aforementioned evolutionary patterns, a multi-criteria verification system and spatial autocorrelation analysis are used to verify the historical authenticity through three types of evidence: literature, physical objects, and word-of-mouth. Spatial continuity analysis is used to assess the integrity of the texture and obtain verification results. A graded protection strategy is formulated based on the verification results; Generating a three-level dynamic time section includes: Based on the holographic historical spatial database, using the NLP hybrid model and entropy weight method, keywords are matched by regular expressions and combined with the contextual understanding of the BERT-wwm model to extract specific events. After calculating the influence weight of specific events, a key event anchor point library is formed. Based on the key event anchor point library, using the Kalman filter algorithm and tree-ring chronology data, and through time ambiguity correction and absolute dating of building wooden components, a timestamp-calibrated anchor point sequence is generated. Based on the anchor sequence calibrated by the timestamp, the Morlet wavelet function and cross-validation method are used to generate a three-level time slice containing a macro slice at a century scale, a meso slice at a decade scale, and a micro slice at an interannual scale through time series decomposition. The three-level time slices are calibrated; Based on the calibrated three-level time slices, the event pattern transmission path is analyzed using a spatiotemporal graph convolutional network. Through spatiotemporal correlation identification, a three-level dynamic time slice is formed, which includes macro slices at the centennial scale, meso slices at the decadal scale, and micro slices at the interannual scale. The proposed historical layered feature map includes: Based on the three-level dynamic time section, a CNN feature extractor is used to process the overlay data of remote sensing images and historical maps through transfer learning, generating three classification results: Category A: typical traditional buildings, Category B: suspected traditional buildings, and Category C: modern buildings. Based on the three classification results, the betweenness centrality and proximity centrality of street and alley nodes are calculated using graph neural networks. The historical accessibility is verified by combining spatial syntax theory, and highly integrated paths such as the north-south city axis are generated. Based on the three-level dynamic time section, a three-dimensional terrain model is constructed using DEM elevation data and water system vector data. The old city structure is simulated by the Viewshed field of view analysis algorithm, and the viewpoint position is dynamically adjusted to generate a field of view coverage heat map. Based on the three-classification results, the high-integration path and the view coverage heatmap, an element relationship graph model is integrated. The graph neural network is used to aggregate node features and identify functional groups to generate an element gene map containing core and edge layered structures. Based on the aforementioned element gene map, spatiotemporal graph convolutional networks are used to analyze the spatiotemporal correlation between events and elements, identify the layered transmission path, and generate a historical layered element map.

2. The method for identifying the historical pattern of an old town based on time cross-section and historical layering as described in claim 1, characterized in that, The expression for calculating the impact weight of a specific event is: in, Indicates an event Total influence weight, Represents the set of influencing factors. Indicator Factors Entropy weight, Indicator Factors In the event Standardized values ​​on Indicator Factors Information entropy Indicates an event In factors The original observations on Represents a set of events. Indicates an event In factors Standardized values The proportion of the total standardized values ​​of all events for this factor.

3. The method for identifying the historical pattern of an old town based on time cross-section and historical layering as described in claim 1, characterized in that, The quantitative evolution patterns include: Based on the historical layered element map, the spatiotemporal correlation features of building texture, street network and landscape pattern are extracted using spatiotemporal graph convolutional network. The information of neighboring nodes is aggregated by spatial graph convolution and the receptive field is expanded along the time dimension to generate a spatiotemporal feature tensor containing macro trends at the centennial scale, meso fluctuations at the ten-year scale and micro changes at the interannual scale. Based on the spatiotemporal feature tensor, a long short-term memory network is used to capture the long-term dependencies of historical patterns. Key evolution information is filtered and noise interference is suppressed through a gating mechanism to generate a hidden state sequence that reflects the dynamic evolution of the pattern. Based on the hidden state sequence, a combination strategy of CUSUM mean detection, Pelt multi-mutation point algorithm and Chow structural fracture test is used to identify the set of pattern mutation points of macro-mutation on a century-scale, meso-level turning point on a decade-scale and micro-level disturbance on an interannual scale through cumulative sum threshold judgment, penalty term optimization segmentation and segmented regression significance test. Based on the set of abrupt change points in the pattern, the evolution rate is quantified by trend slope analysis, periodic fluctuations are identified by Morlet wavelet transform, and the intensity of abrupt change is calculated by Euclidean distance. A comprehensive quantitative evolution law is generated, which includes the spatiotemporal location of abrupt change points, the direction of evolution trend, periodic characteristic parameters, and the intensity index of abrupt change.

4. The method for identifying the historical pattern of an old town based on time cross-section and historical layering as described in claim 1, characterized in that, The verification results include: Based on the spatiotemporal location of the mutation points in the aforementioned evolutionary pattern, a multi-criteria verification system is used to generate historical authenticity verification results, including triple evidence support scores, spatiotemporal alignment, and consistency coefficients, through TF-IDF support calculation of literature evidence, chi-square test consistency verification of physical evidence, and Kappa coefficient alignment of reputation evidence. Based on the historical authenticity verification results, a spatial autocorrelation analysis model is used to calculate spatial continuity through the global Moran's I index, identify local LISA hotspots, and aggregate the comprehensive scores using the entropy weight method, generating a texture integrity assessment result that includes spatial continuity index, type diversity index, and age gradient index. Based on the texture integrity assessment results, a spatiotemporal graph convolutional network is used to perform correlation analysis of the verification results. Verification anomalies are identified through spatiotemporal feature aggregation, and historical layered feature maps are corrected by combining a feedback mechanism to generate verification results that include verification correction parameters, anomaly location, and correction strategies.

5. The method for identifying the historical pattern of an old town based on time cross-section and historical layering as described in claim 4, characterized in that, The expression for calculating spatial continuity using the global Moran's I exponent is: in, Represents the global Moran's I index. Indicates the total number of spatial units. Represents the elements in the spatial weight matrix With unit The weight value, and Indicates the textural characteristic value, This represents the mean of the texture characteristic values.

Images