Intelligent deduction method and system for building time sequence succession in urban renewal

Abstract

The application relates to the field of urban renewal information technology, and discloses an intelligent deduction method and system for building time sequence succession in urban renewal, wherein the intelligent deduction method for building time sequence succession in urban renewal comprises the following steps: gathering building contour vectors, multi-period images and account to construct a building succession object graph; observing the visibility of each succession state by building and year; constructing a succession label set containing confirmed samples, interval missing samples and weak negative example samples; training a neighborhood message passing network to generate positive and negative modulation coefficients from neighbor succession stages to obtain a non-homogeneous state transition kernel; under the account conservation conditions of funds, construction operation faces and resettlement housing sources, a multi-year building-level succession path is deduced by rolling and sorting according to the urgency, and a scenario comparison evaluation and transition contribution attribution are made. The application can deduce the executable update time sequence of a building unit year by year and trace back to the corresponding evidence and constraints under the observation visibility difference and the preposition dependence of the chain.

Description

Technical Field

[0001] This invention relates to the field of information technology for urban renewal, and more specifically, to an intelligent deduction method and system for the temporal evolution of buildings in urban renewal. Background Technology

[0002] As urban development enters the stage of improving the quality of existing buildings, the buildings within the renewal unit are not renovated synchronously over time. Some buildings maintain their current state or undergo repair and reinforcement, some complete the conversion of their uses, and others successively undergo vacating, demolition, and reconstruction on the original site. The district-level urban renewal center needs to prepare an implementation plan on an annual basis, specifying in which year each building will enter which stage of evolution, and provide opinions on the order of adjustment after the funding amount is adjusted.

[0003] Current support methods primarily rely on spatiotemporal simulation at the land use scale and evolution event prediction at the grid scale, supplemented by building-scale renewal potential prediction and building change detection for historical processes. The outputs at the land use and grid scales are raster types or regional event labels, unable to reflect the sequence of building vacating, demolition, and reconstruction; building-scale potential predictions provide static trend values, failing to answer the question of the annual update sequence; and change detection focuses on past processes and does not undertake forward extrapolation.

[0004] The aforementioned methods also share two common technical problems. First, building-level succession labels rely on the reverse interpretation of remote sensing and street view imagery. Demolition and reconstruction leave traces on orthophotos, while repairs and equipment updates often leave no trace or are limited by facade orientation. The meaning of not detecting different states is not the same, and directly using them as negative samples will suppress the probability of relocation in the remediation category. Moreover, when two images span across years, it is difficult to pinpoint the specific year of relocation. Second, the suppression of construction during the construction period of adjacent buildings and the driving force after completion cannot be expressed in the unidirectional neighborhood weights. The pre-requisite dependence of the relocation chain of on-site rolling resettlement is also difficult to express in a self-consistent manner in calculable deductions. Summary of the Invention

[0005] This invention provides an intelligent method and system for extrapolating the temporal evolution of buildings in urban renewal, which solves the technical problems in related technologies where building-level succession labels are systematically lacking due to differences in observation visibility, and where the pre-requisite dependence of the relocation chain under resource constraints is difficult to express self-consistently in multi-year extrapolation.

[0006] This invention provides an intelligent method for extrapolating the temporal succession of buildings in urban renewal, comprising the following steps:

[0007] S1 collects building outline vectors, multi-period images, safety assessments and project ledgers. After standardization and unique cross-year coding of entities, it constructs a building succession object map with individual buildings as nodes and adjacent and access edges as edges, and outputs image slice indexes.

[0008] S2, on the object map and image tile index, through facade line intersection, roof occlusion calculation and state mode detection rate statistics, the visibility of each successive state is quantified to obtain a set of visibility coefficients coupled with the state category;

[0009] S3, based on the image slice index, makes a change determination, combines the visibility coefficient set, divides the intervals and weights the visibility to make a confidence assignment, and forms a succession label set containing three types of samples;

[0010] S4, using the succession label set and object graph attributes as training basis, constructs a neighborhood message passing network that generates positive and negative modulation coefficients from the neighbor succession stage, and trains to obtain a building-by-building and year-by-year non-homogeneous state transition kernel.

[0011] S5. Based on the candidate transfer probabilities given by the transfer core, under the condition of conservation of the three types of accounts (funds, construction work area, and resettlement housing), the candidates for default are included in descending order by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, and the multi-year building-level succession path, the annual trajectory of the three types of accounts, and the extension record are obtained.

[0012] S6 performs scenario-based assessments of multi-year building-level succession paths, annual trajectories of three types of accounts, and continuation records. It calls the transfer kernel to attribute transfer contributions and outputs a building renewal implementation timeline, account constraint trigger records, and a list of key driving factors.

[0013] In a preferred embodiment, the unique cross-year code for entities in S1 is determined by the ratio of the overlapping area of ​​the building outlines across years to their respective areas reaching an overlap area ratio threshold, which indicates that they are the same building entity and share the same code.

[0014] Adjacent edge is established and recorded as normalized spacing when the minimum distance between two building outlines is within the adjacent distance threshold or when they share a gable wall; Passage edge is the outer width of the construction influence range determined according to the building height and structural type. The influence range surface is obtained by buffering the building outline according to the outer width. The shortest path to the unit entrance is solved along the center line of the internal road. When the shortest path intersects with the influence range surface, a passage blocking sign is established and set to true.

[0015] In a preferred embodiment, the intersection of the facade view in S2 is to decompose the building outline into facades, project rays from the sampling midpoint of each facade to the street view collection point and intersect with the remaining building outlines, take the ratio of the number of collection points that are not blocked and whose incident angle does not exceed the incident angle threshold to the number of theoretically observable collection points of the facade, and then attenuate the facade by the vegetation shading rate to obtain the facade view accessibility.

[0016] Roof shading is calculated as the ratio of the remaining area after deducting the shading strip formed by adjacent buildings at the roof edge from the roof area to the roof area, thus obtaining roof visibility. The geometric observability of each successive state is taken as the visual accessibility of the corresponding facade in the street view mode and as the roof visibility in the orthographic mode.

[0017] In a preferred embodiment, the state mode detection rate statistics in S2 are obtained by retrospectively determining whether each successive state presents identifiable features under orthophoto and street view modes based on the ledger records with completion time. The detection rate is then subjected to hierarchical backsliding smoothing and a lower limit is set. The temporal phase coverage is obtained by the ratio of the number of effective time phases in the year interval to be determined to the interval span and by attenuation based on the longest gap window.

[0018] The visibility coefficient set is synthesized independently of the geometric observability, detection rate and temporal coverage of each building in each year and each successive state under each mode, taking the complement of the probability of missed detection in each mode.

[0019] In a preferred embodiment, the interval assignment in S3 is as follows: when the interval between two adjacent images is one year step and both can be determined, the target transfer is recorded as a confirmed sample; when the interval spans multiple years, it is recorded as an interval sample and its year interval is recorded; when a certain building does not detect a certain type of state feature in a certain year, it is recorded as a weak negative sample; and the disappearance of the roof outline corresponds to the transfer to demolition under construction, the addition of the building outline corresponds to the transfer to the renovation and completion, and the change of roof and facade features corresponds to the transfer to repair and reinforcement.

[0020] The confidence assignment for visibility-weighted samples is the upper confidence limit, the interval sample is the likelihood of at least one target shift occurring within the interval of that year, and the weak negative sample is the visibility coefficient corresponding to the visibility coefficient set. The samples are organized into layers according to state category and visibility.

[0021] In a preferred embodiment, the positive or negative modulation coefficients in S4 are generated by splicing the current state embedding of the neighbors on each edge, the current state embedding of the body, the passage blocking flag and the normalized spacing, and then passing through two fully connected layers and hyperbolic tangent. The values ​​are between negative one and positive one and are not normalized during aggregation.

[0022] The neighborhood message passing network multiplies the latent attribute vectors of each neighbor by their modulation coefficients and sums them along the edges. It then aggregates the adjacent edges and the passing edges separately and concatenates them into a neighborhood context. The latent attribute vectors, category attribute vectors, current state embeddings, years of stay embeddings, neighborhood contexts, and intervention codes are then concatenated and processed by a fully connected layer and a normalized exponent to obtain the probability distribution of transitions from the current state to each state, which serves as a row of the non-homogeneous state transition kernel.

[0023] In a preferred embodiment, the objectives of training the non-homogeneous state transition kernel in S4 include the classification loss of the confirmation samples, the interval likelihood loss of the interval samples, the negative loss of the weak negative samples weighted by the visibility coefficient, and the sparse regularization of the modulation coefficients that can be positive or negative.

[0024] Training is rolled over annually. For years with observed states, the observed states are used as the input for the next year. For years without observed states, the inference results from the previous year are used. The network parameters are updated by iterating through sample batches organized by state category and visibility.

[0025] In a preferred embodiment, in step S5, transitions with a transition probability exceeding the candidate probability threshold and a target state different from the current state are selected as candidate transitions; the urgency is composed of the security assessment level ordinal number and the number of years the current state has been in existence, and increases as the assessment level decreases or the number of years the current state has been in existence increases.

[0026] The candidates are ranked in descending order of the ratio obtained by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, and are included in the three types of accounts for pre-verification. Among them, the funds account verification shows that the cumulative amount of funds occupied by the candidates does not exceed the annual funds limit. The construction work area account verification shows that the work area occupied by the buildings under demolition and construction does not exceed the upper limit of the same period and does not constitute a traffic conflict with the candidates already included in the demolition and construction. The resettlement housing account verification shows that its balance is not negative after deduction and the number of resettlement units added by the current year is carried forward to the next year. Candidates that fail the verification are postponed to the next year and their number of years of stay is accumulated.

[0027] In a preferred embodiment, the scenario comparison assessment in S6 involves comparing the progress of dilapidated building elimination, the coverage rate of renovation, the account occupancy status, and the waiting time for relocation among different scenarios obtained by changing the annual funding amount, the upper limit of the work area, or the annual plan scope.

[0028] The transfer contribution is attributed to the input of the specified building and the annual path-by-path masked transfer kernel, and then recalculated forward after replacing it with the mean vector of that path on the training set. The probability change is taken as the contribution of each path in the attributes, current state, number of years of stay, neighborhood context and intervention code. The account constraint trigger record is summarized by the extension record according to the account type and building, to obtain the extension number of each account and the cumulative extension years of each building.

[0029] In a preferred embodiment, an intelligent simulation system for building temporal succession in urban renewal is used to execute the steps of the aforementioned intelligent simulation method for building temporal succession in urban renewal, including:

[0030] The object graph construction module is used to collect building outline vectors, multi-period images, safety assessments and project ledgers. After standardization and unique cross-year coding of entities, it constructs a building succession object graph with individual buildings as nodes and adjacent and access edges as edges, and outputs image tile indexes.

[0031] The observation quantification module is used to quantify the visibility of each successive state on the object map and image tile index by finding the intersection of facade lines, calculating roof occlusion, and statistically analyzing the state mode detection rate, and to obtain a set of visibility coefficients coupled with the state category.

[0032] The succession label construction module is used to determine changes based on the image slice index, combine the visibility coefficient set, divide the intervals and weight the visibility to make confidence assignments, and form a succession label set containing three types of samples.

[0033] The state transition kernel training module is used to construct a neighborhood message passing network with positive and negative modulation coefficients generated by the neighbor succession stage, based on the succession label set and object graph attributes, and trains a non-homogeneous state transition kernel for each building and each year.

[0034] The rolling simulation module is used to include and extend default candidates in descending order by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, based on the candidate transfer probabilities given by the transfer core, under the condition of conservation of the three types of accounts: funds, construction work surface and resettlement housing resources. This results in multi-year building-level succession paths, annual trajectories of the three types of accounts and extension records.

[0035] The assessment and attribution module is used to conduct scenario-based assessments of multi-year building-level succession paths, annual trajectories of three types of accounts, and continuation records. It calls the transfer kernel to perform transfer contribution attribution and outputs building renewal implementation sequence plans, account constraint trigger records, and a list of key driving factors.

[0036] The beneficial effects of this invention are as follows:

[0037] The visibility coefficient is synthesized from facade accessibility, roof visibility, state mode detection rate, and temporal coverage to determine whether each building's successive state could be observed in each year. This coefficient is then incorporated into the monitoring signal, ensuring that undetected states lacking observation conditions no longer serve as negative evidence, and that the probability of transitioning to remediation-type successive states is no longer systematically suppressed. Simultaneously, the likelihood of cross-year projects is monitored using interval censoring, eliminating the need to assign them to specific years, and ensuring that the projected annual rhythm does not shift with time assignment. Furthermore, samples are organized hierarchically according to state category and visibility, maintaining a stable proportion of samples with low visibility states during training.

[0038] The modulation coefficients generated during the neighbor succession stage are positive or negative and non-normalized, so that the inhibition of construction during the construction period of adjacent buildings and the driving force after completion are expressed in opposite directions within the same structure; and the conservation conditions of three types of accounts—funds, construction work area, and resettlement housing—replace global optimization, so that the prerequisite of construction before vacating and vacating before demolition is generated by the constraints themselves. Annual rolling and default extension ensure that buildings with low assessment levels are not suppressed for a long time, and the extrapolation can be implemented on the actual scale of the renewal unit; through the contribution attribution of inputs that are shielded one way and the account constraint triggering record, each extrapolation conclusion and its bottleneck can be traced back to the corresponding evidence and constraints. Attached Figure Description

[0039] Figure 1 This is a structural diagram of an intelligent simulation system for the temporal evolution of buildings in urban renewal, according to the present invention.

[0040] Figure 2 This is a flowchart of an intelligent deduction method for the temporal evolution of buildings in urban renewal according to the present invention;

[0041] Figure 3 This invention provides a flowchart of an intelligent deduction method for the temporal evolution of buildings in urban renewal. Figure 1 ;

[0042] Figure 4 This invention provides a flowchart of an intelligent deduction method for the temporal evolution of buildings in urban renewal. Figure 2 . Detailed Implementation

[0043] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0044] At least one embodiment of the present invention discloses an intelligent deduction method for the temporal evolution of buildings in urban renewal, such as... Figure 1The diagram shows the overall structure of the intelligent simulation system for building temporal succession in urban renewal according to the present invention. This method uses multi-source spatiotemporal data as the entry point and obtains three simulation outputs that can be reviewed by the urban renewal center through six sequentially connected processing stages. The dashed box in the figure shows the seven types of data entering the first stage, including the urban renewal unit range vector, building outline vector, multi-period orthophotos, street view images, building safety appraisal ledger, implemented project ledger, and internal road centerline and resettlement housing resources. The solid box shows the main processing chain. The scenario setting sends three types of accounts—funds, construction work surface, and resettlement housing resources—from the bypass to the fifth stage. The sixth stage outputs the building renewal implementation sequence plan, account constraint trigger records, and key driving factor list to the simulation output box.

[0045] After seven types of data are imported, the first stage constructs a building succession object graph with individual buildings as nodes and adjacent and access edges as edges, and outputs an image tile index. The second stage quantifies the visibility of each succession state on the object graph and tile index, outputting a set of visibility coefficients coupled with the state category. The third stage combines image change judgment and the visibility coefficient set to form a succession label set containing confirmed samples, interval samples, and weakly negative samples. The fourth stage uses positive and negative neighborhood message passing based on the succession label set and object graph attributes to train a non-homogeneous state transition kernel for each building and each year. The fifth stage, based on the candidate transition probabilities given by the transition kernel, rolls in or extends candidates under the condition of three types of account conservation, obtaining multi-year building-level succession paths, annual trajectories of the three types of accounts, and extension records. The sixth stage performs scenario-based evaluation of the paths, account trajectories, and extension records, calls the transition kernel to attribute transition contributions, and outputs the aforementioned three results. Specifically, as follows... Figures 2 to 4 As shown, it includes:

[0046] S1 collects building outline vectors, multi-period images, safety assessments and project ledgers. After standardization and unique cross-year coding of entities, it constructs a building succession object map with individual buildings as nodes and adjacent and access edges as edges, and outputs image slice indexes.

[0047] Specifically, the collected data includes seven categories: updated unit range vectors define the spatial boundaries of the projection; building outline vectors carry attributes such as number of floors, year of construction, structural type, building area, and property rights; multi-period orthophotos cover several years prior to the projection's starting year; street view images include the coordinates of the collection points, shooting orientation, and field of view parameters; building safety appraisal ledgers record the appraisal level and appraisal time; implemented project ledgers record the project name, scope, and implementation year; and internal road centerlines and resettlement housing lists are used for subsequent traffic relationships and account initialization.

[0048] All types of data are first unified to the same geodetic coordinate system and the same projection. Raster data are resampled to a consistent ground sampling interval using the nearest neighbor or bilinear method. Building outlines vary in terms of splitting, merging, and outline accuracy across different years. Therefore, entity matching is not judged by geometric consistency, but by whether the ratio of the overlapping area of ​​the outlines across years to their respective areas reaches a threshold. Outlines that reach this threshold are considered the same building entity and share a unique code. For buildings that disappear in intermediate years, their codes are retained and the time interval of disappearance is marked so that demolition categories can be mapped to this interval later.

[0049] Aligning the ledger data with the actual building structures is a practical challenge at this stage. Most implemented project ledgers record only the community name or plot number, not the building number. This embodiment uses the spatial inclusion relationship between the plot boundary and the building outline to sink the ledger records down to the individual building level, meaning one ledger record is associated with all buildings within its scope. Records whose scope description is insufficient to determine the inclusion relationship are marked as weak evidence at the plot level and do not enter the subsequent confirmation sample set. Building safety appraisal ledgers are usually recorded at the building level, and can be directly matched by building number or address. If a match fails, the nearest neighbor is used, combined with the number of floors and the year of construction for verification.

[0050] In one embodiment of the present invention, the node attribute vector consists of two parts: a numerical type and a categorical type. The numerical type includes the number of floors, building area, number of years since construction from the starting year of the projection, and road network distance to the unit entrance, which are linearly normalized according to their respective value ranges. The categorical type includes the structure type, property rights, safety assessment level, and whether it is included in the reserved list. After being converted into an index, the vector is obtained by embedding the table in subsequent stages. In addition, each building is assigned an initial state for the starting year of the projection. Its value is based on the latest image interpretation results and safety assessment ledger. Buildings that have been acquired but not demolished are assigned the status of being vacant and awaiting demolition, buildings under construction are assigned the status of being demolished and under construction, and the rest are assigned the corresponding status according to the rectification record and the use record.

[0051] In addition to existing building nodes generated from building outline vectors, the node set also includes pre-defined planned new construction nodes based on the resettlement housing list and planned relocation plots. Each planned relocation building corresponds to a planned new construction node at its planned plot location. Its attributes are given by planning indicators such as the planned number of units, the proposed number of floors, the structural type, and the building area. It is also labeled as planned new construction with a category attribute to distinguish it from existing buildings. The initial state of the planned new construction node is recorded as "maintaining the status quo." Here, "maintaining the status quo" means that the plot has not yet started construction, and it is subsequently transformed into a redevelopment phase after the demolition of the under-construction new construction work. The resettlement buildings located on the original sites of existing dilapidated buildings will no longer have separate nodes. Instead, the existing nodes will be successively vacated and awaiting demolition, demolished and under construction, and then converted to new construction to express the reconstruction after demolition on the original site. The planned new nodes have no historical image support, do not participate in subsequent visibility quantification and label construction, and do not participate in the training of the transfer core. That is, they are neither training samples nor neighbors during the training period. They only participate in the calculation as neighbor status and construction candidates during rolling simulation. The post-construction driving mode required by the transfer core is learned from the historical succession of the existing buildings and then applied to the reconstruction of the planned new nodes during simulation.

[0052] Edge relationships are divided into two categories: adjacent edges and passable edges. The two are constructed based on different criteria and play different roles in subsequent calculations. Adjacent edges describe spatial proximity. When the minimum distance between the outlines of two buildings is within the adjacent distance threshold, or when they share a gable wall, an adjacent edge is established and its normalized distance is recorded. Passable edges describe construction interference relationships. Their construction first determines the outer width of the construction influence range according to the building height and structural type. This outer width is obtained by multiplying a base width by a height coefficient that increases with the building height, and then by a structural coefficient obtained from a table based on the structural type. The height coefficient corresponds to the larger fall and rotation range of tall buildings during demolition or hoisting, and the structural coefficient corresponds to the differences in demolition methods for different structures such as brick-concrete and frame. Each coefficient is calibrated according to local construction safety protection regulations. After obtaining the outer width, the building outline is buffered by this width to obtain the influence range surface. Then, the shortest path from the main entrance of another building along the center line of the internal road is solved to the nearest entrance of the unit. If this path intersects with the influence range surface, a passable edge is established from the former to the latter and the passable blockage sign is set to true.

[0053] The two types of edges constructed in this way provide a physical carrier for subsequent neighborhood calculations; the adjacent edges carry the demonstration and driving relationship brought about by spatial proximity, while the access edges carry the relationship of construction restrictions caused by traffic obstruction during construction; the final output of this stage is a building succession object map, whose nodes carry attribute vectors and initial states, and whose edges carry type markers, access obstruction markers and normalized spacing, while also outputting the slice index of each building in the orthophoto and street view images of each period.

[0054] S2, on the object map and image tile index, through facade line intersection, roof occlusion calculation and state mode detection rate statistics, the visibility of each successive state is quantified to obtain a set of visibility coefficients coupled with the state category;

[0055] Specifically, after the object map and image slice index output from the previous stage enter this stage, the first thing to do is to solve a problem that repeatedly occurs in actual interpretation: traces of remediation projects only appear on building facades or roofs, and whether these two places can be captured depends on whether there is obstruction between the building and the street view acquisition point, and whether the adjacent buildings cast obstruction strips on the edge of the roof; if this condition is not quantified first, the non-detection in the interpretation results cannot be distinguished from non-occurrence.

[0056] The visibility of a facade is calculated separately for each facade rather than taking a single value for the entire building. This is because elevator shafts and equipment pipelines are often located on facades that are not facing the street, while street view acquisition is conducted along main urban roads, resulting in significant differences in the photographable conditions between facades facing the street and those facing away from it. The calculation first decomposes the building outline into several facades along its edges. At equal intervals, sampling midpoints are taken on each facade. Rays are projected from each sampling midpoint to all street view acquisition points within the block area where the building is located. Each ray is then intersected with other building outline polygons. If an intersection point exists, the view is considered blocked. Unblocked rays must also meet the incident angle condition, meaning the angle between the ray and the facade normal does not exceed the incident angle threshold. If the angle exceeds this threshold, the facade is compressed into a width that is difficult to interpret in the image. The visibility of a facade is calculated as the ratio of the number of acquisition points that meet the above two conditions to the theoretically observable number of acquisition points for that facade. The theoretically observable number of acquisition points refers to the total number of acquisition points within the block area located in the front half-space of that facade.

[0057] In street view images, vegetation can obscure the lower part of a facade without blocking the view, so its influence needs to be deducted separately. In this embodiment, semantic segmentation is performed on the street view slice corresponding to each facade, and the proportion of vegetation pixels in the lower part of the facade's imaging area is counted. This proportion is recorded as the vegetation occupancy rate of the facade. The aforementioned viewability is then multiplied by the difference between the aforementioned viewability rate and the occupancy rate. The result is used as the viewability of the facade after vegetation attenuation in the street view modality. The calculation basis for roof visibility is different. Orthophotos are approximately vertically observed but not strictly orthophotos. Adjacent taller buildings will form occupancy bands at the roof edges. The width of the occupancy band is estimated by multiplying the height difference between the adjacent building and the building itself, the ratio of the distance between the two buildings, and the tangent of the imaging zenith angle. The roof visibility is the ratio of the roof area after deducting the area of ​​the occupancy bands in each direction from the roof area to the roof area.

[0058] In one embodiment of the present invention, geometric observability only answers whether the part can be seen on the image, and also needs to answer whether the state exhibits identifiable features in the modality. For this purpose, this embodiment establishes a detection rate matrix of state and modality, with its rows corresponding to six successive states and its columns corresponding to two modalities: orthophoto and street view image. The values ​​of this matrix are not externally set, but are obtained by retrospective statistics from the ledger records of this unit and similar units that can be down to a single building and have completion time. The statistical method is to take image slices before and after the completion time of each confirmed record, determine whether the state presents identifiable features in the modality and count it as a hit, and then summarize the number of hits and the total number of records according to state and modality respectively. The ratio of the two is the initial detection ratio.

[0059] The confirmed records for repair, reinforcement, and functional replacement are already scarce. Directly calculating the proportions is unstable when the sample size is insufficient. Therefore, a graded backsliding smoothing process is applied to the detection rate of each state and mode. The smoothing method is to first take the hit count and total number of records for that state and mode in the current unit, and then introduce a prior detection ratio obtained from the aggregation of similar units and an equivalent prior sample number. The product of the hit count in the current unit plus the prior detection ratio and the equivalent prior sample number is used as the numerator, and the total number of records in the current unit plus the equivalent prior sample number is used as the denominator. The two are divided to obtain the smoothed detection rate. When the samples in the current unit and similar units are insufficient to support stable statistics, the prior detection ratio is further backslid to the aggregation result of all available units, and a lower limit is set for the final value to avoid the detection rate of a certain state in a certain mode being zero, which would cause the visibility to be erroneously suppressed to zero. After this processing, the detection rate matrix is ​​close to the actual situation of the current unit when the sample size is sufficient, and smoothly transitions to similar units and global results when the sample size is insufficient.

[0060] This statistical process itself reveals the differences between the states; the detection rate of demolition under construction and renovation completed in the orthophoto mode is close to the upper limit because the disappearance of the roof and the addition of the outline are obvious features; the detection rate of repair and reinforcement and functional replacement in the orthophoto mode is close to the lower limit because their features do not appear on the roof; the detection rate of repair and reinforcement and functional replacement in the street view mode is affected by the facade orientation. If the project occurs on the back street facade, it is difficult to present even if the street view covers it; this matrix, together with the aforementioned geometric observability, constitutes the basis for the coupling of state categories.

[0061] Temporal coverage answers whether there are sufficient images to support interpretation for the year; its value is determined by the ratio of the number of effective temporal phases for that mode within the interval of the year to be determined to the interval span. Effective temporal phases refer to the number of usable observations for that mode within the interval with acceptable image quality. Attenuation is then applied based on the longest window interval within the interval; the degree of attenuation increases with the ratio of the longest window to the interval span, and the longer the window, the lower the temporal coverage value. The three factors are combined into a visibility coefficient based on the assumption that missed detections for each mode are independent, using the following formula:

[0062]

[0063] in, For architecture In the year For state categories Observable visibility, To observe the total number of modes, For architecture In modality Next to state categories Geometric observability For state category In modality The detection rate is below For modality In the year Temporal phase coverage;

[0064] Geometric observability in the formula In the orthophoto mode, rooftop visibility is measured, with the same value for each state category. In the street view mode, the visual accessibility of the corresponding facade for that state category is measured after vegetation attenuation; since the facade varies with state category, this value differs for each state category. Detection rate. Take the elements of the detection rate matrix after the aforementioned backtracking statistics and hierarchical backtracking smoothing, and the temporal coverage. Take the value calculated based on the number of effective time phases and the longest gap window; the multiplication term in the formula represents the probability that each mode will miss detection at the same time, and its complement is the probability that at least one mode can observe the state;

[0065] The correspondence between geometric observability and state category is determined according to the actual location of the engineering traces in each state. This correspondence is the specific landing point of the coupling between visibility and state category. Repair and reinforcement involve structure and pipeline network, and its construction facade is not fixed. In the street view modality, the maximum value of the visual accessibility of each facade of the building is taken. As long as one side can be photographed, it is considered to have the observation conditions. The traces of facade renovation appear on the main facade facing the street, and the visual accessibility of the facade facing the street is taken. The shaft and external pipeline of the elevator are mostly arranged on the non-street facade or gable wall, and the visual accessibility of the corresponding rear facade is taken. Functional replacement is mostly manifested as the renovation of the storefronts and shops on the ground floor facing the street, and the visual accessibility of the facade facing the street is taken. Vacant buildings waiting to be demolished do not have significant facade engineering features, and their geometric observability in the street view modality is taken as a lower value, mainly supported by the vacancy and pile-up features in the orthographic modality. Demolition under construction and renovation are based on the disappearance and addition of the roof outline. The geometric observability is taken as the roof visibility, and its street view modality does not participate in the synthesis.

[0066] Output a set of visibility coefficients, indexed by building number, year, and status category. The significance of this coefficient set is to separate the condition of whether or not the building is visible from the judgment result, so that the next stage can distinguish between the two meanings of "not detected". For example, if the facade renovation of a building facing the street is not detected in the same year, its visibility value is relatively high, and the negative information provided by "not detected" is credible. On the other hand, if the elevator installation of a building located deep in the unit and whose back facade is not covered by street view is not detected, its visibility value is close to the lower limit, and "not detected" provides almost no negative information.

[0067] S3, based on the image slice index, makes a change determination, combines the visibility coefficient set, divides the intervals and weights the visibility to make a confidence assignment, and forms a succession label set containing three types of samples;

[0068] Specifically, this stage receives the visibility coefficient set output from the previous stage, and extracts multi-period image slices of each building based on the image slice index output from the object graph construction stage, and performs building-level change judgment on them; the judgment items for the orthophoto mode include three categories: disappearance of roof outline, appearance of new building outline, and change of roof material and texture; the judgment items for the street view mode include two categories: change of facade color and material, and appearance of additional components, among which additional components include elevator shafts, canopies and external pipelines; the judgment is made by comparing each building slice, and the judgment result is recorded as a certain type of feature change of a certain building between two periods;

[0069] The correspondence between each judgment item and the transition to the succession state is determined according to its engineering meaning. This correspondence places the changes in image features into specific target states. The disappearance of the roof outline corresponds to the transition to demolition and construction, the appearance of the new building outline corresponds to the transition to renovation and completion, the change of roof material and texture corresponds to the repair and reinforcement of the roof, the change of facade color and material corresponds to facade renovation and repair, and the appearance of additional components corresponds to the installation of elevators and equipment upgrades. The image features of functional replacement are relatively weak and are usually not determined by a single judgment item, but by the facade changes of the ground floor storefront and shop front combined with the use change records in the ledger. Through this correspondence, the direction of transfer and the year of occurrence can be determined for confirmed samples, and the direction of transfer and the year interval can be determined for interval samples.

[0070] The correspondence between the judgment result and the year is handled in three cases, which is the key division of labor in this stage. The first case is that the interval between two adjacent images is exactly one year step, and the building status can be determined in both the previous and next images. In this case, the year in which the target transfer occurred is determined, and this sample is recorded as a confirmatory sample. The second case is that the interval between two adjacent images spans multiple years. It can only be confirmed that the target transfer falls within the interval but not in which year. In this case, this sample is recorded as an interval sample, that is, the interval censored sample mentioned in the title, and its annual interval set is recorded. The third case is that a certain type of status feature of a building is not detected in a certain year. This sample is recorded as a weak negative sample.

[0071] The ledger evidence is incorporated into the above three scenarios according to its time precision; records with the year of completion and acceptance are incorporated into the confirmation sample; records with only the year of project implementation but spanning multiple years are incorporated into the interval sample, with the interval being the start and end years of the project; weak evidence at the plot level is not incorporated into any sample set and is only used for subsequent result verification; the purpose of this treatment is to avoid using evidence with insufficient time precision as confirmation, because once a project spanning multiple years is assigned to a specific year, the annual rhythm learned by the model will be shifted as a whole.

[0072] In one embodiment of the present invention, the confidence levels of the three types of samples are assigned according to different rules; the confidence level of the confirmatory samples is taken as the upper limit, and its supervision form is the standard classification loss; the interval samples are not assigned a specific year, and their supervision form is the likelihood that at least one target shift occurs within the interval, that is:

[0073]

[0074] in, For interval samples, This represents the set of annual intervals corresponding to this sample. For architecture In the year From state Turning status The probability of;

[0075] The multiplication term in the formula represents the probability that the transition will not occur in any year within the interval, and its complement is the probability that it will occur at least once within the interval. This form allows the model to automatically allocate the transition tendency to a certain year or several years within the interval without mandatory assignment.

[0076] The confidence level of weak negative examples is directly taken as the visibility coefficient of the building and the state category in that year. This assignment links the negation strength provided by the non-detection to the observation conditions: when the observation conditions are sufficient, the negation strength is close to the confirmatory negation, and when the observation conditions are insufficient, the negation strength is close to none. At this point, the result of non-detection is no longer regarded as not occurring. The visibility quantified in the previous stage enters the supervision signal here.

[0077] The sample organization also needs to address a structural skew; there is an overabundance of confirmed samples from demolished, under-construction, and renovated projects, and an underabundance of confirmed samples from repaired, reinforced, and functionally replaced projects, while there is an overabundance of weak negative samples. If the samples are mixed and trained according to the original proportions, the first two categories will dominate the parameter updates. In this embodiment, the samples are organized in layers according to state category and visibility. Each layer is extracted according to a preset proportion within each training batch, so that samples with low visibility states maintain a stable proportion in the batch. The succession label set is output in this stage, which contains the set of three types of samples and their respective confidence levels.

[0078] S4, using the succession label set and object graph attributes as training basis, constructs a neighborhood message passing network that generates positive and negative modulation coefficients from the neighbor succession stage, and trains to obtain a building-by-building and year-by-year non-homogeneous state transition kernel.

[0079] Specifically, this stage receives the succession label set output from the previous stage and extracts node attributes and edge relationships from the object graph to construct a transition kernel network. The network input is divided into five paths, and the data flow is as follows: The first path is a numerical attribute vector, which is mapped to the hidden layer dimension through a fully connected layer and then nonlinear activation is applied, followed by another fully connected layer to obtain the attribute latent vector; the second path is a category attribute index, which includes structural type, property rights, security assessment level, and reserved list flag, and each is concatenated after extracting vectors from independent embedding tables; the third path is the current state index, which is extracted from a state embedding table, which contains six rows, each corresponding to a succession state; the fourth path is the number of years the data has remained in the current state, which is first binned by the number of years and then processed by the embedding table. The vector is used to distinguish between two situations: those that have just entered the vacant and demolition phase and those that have been vacant and demolition for a long time. The latter often corresponds to buildings where the expropriation has been completed but the relocation conditions have not been met. The fifth vector is the scenario intervention vector, which includes three components: annual funding intensity, whether it is included in the annual plan, and policy inclination. It is encoded by a fully connected layer. This vector is given by the scenario setting in the extrapolation phase and taken from the historical records in the training phase. The annual funding intensity is obtained by normalizing the annual investment amount of the implemented project ledger. Whether it is included in the annual plan is obtained by the inclusion record of the annual plan document. The policy inclination is mapped from the updated policy level of the year to an ordered code, so that the training and extrapolation use the same form of intervention expression.

[0080] The above five paths are each mapped to a preset hidden layer dimension. The dimension after mapping is determined during network initialization and remains unchanged during training to ensure the stability of the dimension of the subsequent concatenated vector. The purpose of binning by the number of years of stay is to express the influence of the number of years as non-linear. The difference between entering a certain state and staying in the same state for a long time does not change linearly with the number of years. After binning, each interval of the number of years can occupy an embedding vector to carry this non-linearity.

[0081] Neighborhood message passing is the core structure of this stage and what distinguishes it from conventional graph network aggregation. For each edge, the embedding vector of the current state of the neighbor and the embedding vector of the current state of the ontology are taken first. Then, the passage blocking flag and normalized distance of the edge are taken. These four are concatenated into a vector and fed into a two-layer fully connected structure. The first fully connected layer performs a linear transformation on the concatenated vector and applies a nonlinear activation. The second fully connected layer maps the activation result into a single scalar. This scalar is then compressed to between negative and positive one by the hyperbolic tangent function. The resulting value is the modulation coefficient of the neighbor on the ontology in that year. A positive modulation coefficient indicates that the neighbor has an uplifting effect on the corresponding transition of the ontology, and a negative one indicates a downlifting effect. The absolute value indicates the strength of the effect.

[0082] The modulation coefficient ranges from negative one to positive one and is not normalized during aggregation. This setting makes it possible for the same neighbor to have opposite effects on the entity at different stages. When the neighbor is in the demolition and construction stage and the passage blockage flag on that side is true, the network can learn a negative modulation coefficient, thereby reducing the probability that the entity will be transferred to repair and reinforcement or demolition and construction in the same year, corresponding to the situation where construction fences occupy passage paths and work spaces on site. When the same neighbor is transferred to the renovation and completion stage, its state embedding changes, and the network can learn a positive modulation coefficient, thereby increasing the probability that the entity will be transferred to repair and reinforcement and functional replacement, corresponding to the situation where the residents of this building vote to approve the neighboring building after it is delivered for use.

[0083] It is important to emphasize the generation method of the modulation coefficients here, as it determines that this structure cannot be understood as simply pre-marking the positive and negative signs on the edges. The modulation coefficients are not fixed attributes of the edges, but are generated annually by the aforementioned small fully connected structure after splicing the current state of the neighbors, the current state of the entity, the passage blocking flag, and the normalized spacing. Therefore, the modulation coefficients of the same edge will change their sign and amplitude year by year depending on the succession stage of the neighbors. This is different from the aggregation method that fixes the edge polarity on the graph or takes non-negative weights based on similarity. The latter, once marked, does not change with the year and the stage of the neighbors, and cannot express the reversal of direction between the suppression during the construction period and the driving force after completion. Normalization is deliberately not performed in the aggregation stage in order to preserve the positive and negative directions and relative amplitudes of the modulation coefficients. If normalization is performed before aggregation, the directional information will be compressed, and the negative suppression and positive driving force will degenerate into a difference of strength.

[0084] The neighborhood context vector is obtained by multiplying the latent attribute vectors of each neighbor by their modulation coefficients and then summing them along the edges. Adjacent edges and access edges are aggregated separately and then concatenated, so that the driving relationship brought by spatial proximity and the inhibitory relationship brought by access obstruction are each formed separately and will not cancel each other out in the summation. Then, the six components of latent attribute vector, category attribute vector, state embedding, years of stay embedding, neighborhood context and intervention code are concatenated and mapped to the state dimension through two fully connected layers to obtain the score of each target state. Then, the probability distribution of the building transitioning from the current state to each state in the current year is obtained by normalizing the exponential function, which is one row of the transition kernel. The normalizing exponential function is to first take the exponent of each target state score and then divide it by the sum, so that the output is a set of non-negative probabilities that sum to one. Since this row is generated by the above six components, and the state embedding, years of stay embedding, neighborhood context and intervention code all change with the building and the year, the transition kernel is not homogeneous between buildings and between years, which is different from the fixed transition matrix of the lookup table.

[0085] In one embodiment of the present invention, the training objective consists of four components, as specified in the formula:

[0086]

[0087] in, For the total loss, To confirm the cross-entropy loss on the samples, For interval samples, the interval likelihood loss is... The visibility coefficient for weak negative samples. This represents the transition probability of the state corresponding to this sample. The regularity of the modulation coefficient, The modulation coefficient;

[0088] The third term needs separate explanation; when the visibility coefficient approaches the lower limit, the value of this term approaches zero, and its gradient with respect to network parameters also approaches zero. Therefore, undetected cases without observation conditions will not provide a negative signal to the network. When the visibility coefficient approaches the upper limit, this term degenerates into a regular negative sample log loss. The fourth term applies a sparsity constraint to the modulation coefficients. Its function is to suppress the network from excessively pushing the prediction responsibility to neighborhood paths while ignoring ontological attributes and the number of years of residence, thus avoiding neighborhood-dominated degeneration. The value of the regularization strength can be adjusted on the validation set so that the contribution of neighborhood paths to the transition probability does not exceed a preset proportion, while maintaining that the interval sample likelihood does not decrease significantly. The value selected accordingly is used as the default. For example, the regularization strength in this embodiment... The default value is one in a thousand;

[0089] The training process unfolds on a rolling annual basis. For years with observed states, the input state for the next year is taken from the observed values ​​rather than the model's inferred values ​​to avoid the accumulation of errors over the years. For years without observed states, the input state for the next year is taken from the inferred results of the previous year. Each training round traverses the hierarchically organized sample batches. First, the transition probability rows for each building and each year within the batch are calculated forward. Then, the first three losses are calculated separately according to the sample type, and a regularization term is added. Finally, backpropagation updates the parameters of each embedding table and the fully connected layer. The number of years of residence accumulates or returns to zero during the rolling process depending on whether the state changes. Its value is synchronized with the state inference.

[0090] After training, the network is fixed as a non-homogeneous state transition kernel. Given the attributes, current state, number of years of stay, neighbor states, and intervention settings of any building in any year, the transition kernel outputs the probability distribution of its transition to six states. The transition kernel is output in this stage and is subsequently called by the next stage under resource constraints.

[0091] S5. Based on the candidate transfer probabilities given by the transfer core, under the condition of conservation of the three types of accounts (funds, construction work area, and resettlement housing), the candidates for default are included in descending order by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, and the multi-year building-level succession path, the annual trajectory of the three types of accounts, and the extension record are obtained.

[0092] Specifically, this stage receives the non-homogeneous state transition kernel output from the previous stage, and also receives the annual funding amount, the upper limit of the construction work area during the same period, and the number of initial resettlement housing units set by the scenario. The deduction is carried out step by step on an annual basis. In each year, the transition kernel first calculates the transition probability row for all buildings, and takes the non-self transitions whose probability exceeds the candidate probability threshold as candidate transitions for that year, forming a candidate set. Non-self transitions refer to transitions where the target state is different from the current state.

[0093] The candidate set cannot be included as a whole; each candidate must be verified in order of priority. The ranking is based on a ratio, the numerator of which is the product of the probability of the candidate being transferred and the urgency of the building in that year, and the denominator is the amount of funds occupied by the transfer. That is, the ratio is equal to the transfer probability multiplied by the urgency and then divided by the amount of funds occupied.

[0094] The urgency level is jointly determined by the safety assessment level and the number of years the building has remained in its current state. The safety assessment level is first mapped to an ordinal number that increases with the degree of danger and then normalized to the highest level. The number of years remaining is processed by a saturation function whose rate of increase slows down with the number of years. The two are weighted and summed according to non-negative weights to form the urgency level. The urgency level increases when the assessment level is low or the number of years remaining increases. The number of years remaining is in a saturated form to ensure that the urgency level of buildings that have been staying for a long time is close to the upper limit and does not increase indefinitely with the number of years. Its weight can be determined by the renovation center according to the guidance of safety priority or progress priority. The amount of funds occupied here refers only to the funds expenditure for the transfer, and its value is converted according to the actual expenditure of the operation corresponding to the target state of the transfer. The amount of funds occupied for the transfer is converted according to the nature of the operation. For demolition operations, it is converted according to the building area and the demolition unit price. For newly constructed nodes, the cost of new construction work is calculated based on the planned number of units and the unit construction cost; for those undergoing repair and reinforcement, the cost is calculated based on the list of renovation contents and the building area; for those undergoing functional replacement, the cost is calculated based on the renovation contents and the building area; for those vacated and awaiting demolition, the cost is calculated based on the expropriation compensation and relocation expenses; for those undergoing renovation and construction, the cost is calculated based on the completion confirmation, and no new capital expenditure will occur in the current year, therefore, it is not separately included in the capital occupation and is not included in the ratio ranking; its inclusion is based on the number of years the building remains in the demolition and construction state to reach the construction cycle calculated based on its engineering volume, which is a definite successor after demolition and construction; the resettlement housing account and the construction work area account are not included in the denominator here, but are separately verified as feasibility conditions before inclusion; the meaning of this ratio is the product of the transfer tendency and urgency obtained by unit capital investment, and those with higher ratios are given priority for inclusion.

[0095] The prioritization of planned new construction nodes needs special explanation because these are resettlement buildings that have not yet been built and do not have a safety assessment rating. The probability of them being transferred to demolition and new construction work is determined by the transfer core based on the planning attributes of the node and whether it is included in the annual plan. The urgency is not determined by the safety assessment rating, but by the number of years the node has been waiting for construction and the housing shortage in the current year. The housing shortage refers to the portion of the resettlement demand of the building to be vacated in the current year that exceeds the housing supply. The larger the shortage or the longer the waiting period, the higher the urgency. In this way, when the relocation chain is waiting due to insufficient housing, the urgency of the corresponding resettlement building construction candidate increases and moves forward in the priority, so that the production and consumption of housing resources drive each other within the same priority mechanism.

[0096] Candidates are included one by one in descending order of their ratios, with verification of three types of accounts before inclusion. For the funds account, the cumulative amount of funds already included in the candidates for the year is verified to ensure it remains within the annual funding limit. The construction work area account is verified on two levels: firstly, the work area occupied by buildings under demolition or construction in the current year, accumulated by building area or section, does not exceed the upper limit of the work area for the same period. Here, "demolition under construction" encompasses both demolition and new construction operations. Secondly, whether the candidate conflicts with already included candidates in the demolition under construction category. The conflict is determined when the construction impact areas of the two intersect and one's impact area cuts off the other's path. The required path edge for this determination has been constructed in the object graph. The process involves phased generation; verification of the resettlement housing account balance to ensure it remains non-negative after the candidate is included; the account is carried over annually, with the resettlement housing balance at the start of the next year equal to the current year's starting balance, plus the sum of the number of resettlement units available from all buildings that were upgraded and designated for resettlement in the current year, minus the sum of the number of households to be resettled from all buildings that were vacated and awaiting demolition in the current year, and the starting balance of any year must not be negative; when candidates are included one by one, the candidates that were vacated and awaiting demolition are verified against the current year's starting balance and are still non-negative after deduction, while the number of newly added resettlement units in the current year is included in the starting balance of the following year and does not participate in the verification of vacated candidates in the current year, thus ensuring that the housing built in the current year can be used for vacation in the following year;

[0097] The condition of a non-negative balance serves to express prior dependencies; transferring candidates to vacate and await demolition will consume resettlement housing resources. If the balance is insufficient, the candidate cannot be included. The increase in the balance can only come from buildings that have been completed and designated for resettlement in previous years. Therefore, the order of completion before vacating is generated by the account conditions themselves; the land released after demolition returns to the planned resettlement land. The planned new construction nodes that have been pre-set in the object map construction stage will start construction in subsequent years according to resettlement needs. Therefore, the order of demolition before construction is reflected by the state of demolition under construction on the same land before its completion. In this way, the order of vacating, demolition, and new construction is jointly generated by the resettlement housing resource account and the state order, without the need for additional logical rules.

[0098] In one embodiment of the present invention, candidates that fail account verification are not invalidated, but are carried over to the next year and retain their candidate status; while being carried over, the number of years the building has been in existence is accumulated, and as can be seen from the composition of urgency, its value in the next year increases, thus obtaining a higher ranking in the next year's ranking; this process ensures that buildings that are postponed due to insufficient resources will not be suppressed for a long time, especially for dilapidated buildings with low assessment levels and whose funds have been fully occupied in the current year, their priority after being postponed for one year will be significantly moved forward;

[0099] The waiting phase of the relocation chain is directly provided by the account trajectory. If the account balance of resettlement housing is zero in a certain year and there are candidates to be transferred to vacate and demolished in the candidate set, no vacating transfer will be generated in that year, and the simulation will enter the waiting phase until a building is transferred to be upgraded and completed, increasing the account balance. The year in which the waiting occurs and the buildings involved are the bottleneck phase of the relocation chain. To prevent the simulation from not terminating under abnormal input, a deadlock judgment is set in this stage. That is, if the balances of the three types of accounts do not change and no candidates are included in a certain number of consecutive years, the simulation will terminate and the bottleneck account identifier and the list of buildings involved will be output. The number of consecutive years can be slightly larger than the typical cycle from vacating to resettlement of a single unit in this unit, so that normal waiting will not be misjudged as deadlock. For example, in this embodiment, the default value of the number of consecutive years for deadlock judgment is three.

[0100] The simulation progresses year by year until the set termination year. At the end of each year, the status of buildings included in the candidate list is updated to their target status, while those not included remain in their original status and their years of residence are accumulated. The income and expenditure of the three types of accounts are carried forward to the next year. This stage outputs the multi-year building-level succession path, which includes the succession status of each building in each year. It also outputs the annual trajectory of the three types of accounts and all the follow-up records. This output is both the object of evaluation and the basis for attribution.

[0101] S6 performs scenario-based evaluation of multi-year building-level succession paths, annual trajectories of three types of accounts and continuation records, calls the transfer kernel to attribute transfer contributions, and outputs building renewal implementation sequence plan, account constraint trigger records and a list of key driving factors.

[0102] Specifically, this stage receives the succession path and account trajectory output from the previous stage; the scenario is defined by the differences set by the intervention. Under the same object map and the same transfer core, changing the annual funding amount, the upper limit of the work area, or the scope of the annual plan will result in different succession paths; the evaluation is carried out by comparing between scenarios, without giving absolute scores, because absolute scores are affected by local construction costs and the number of households and are not comparable across units.

[0103] The comparison dimensions are fourfold: The progress of dilapidated building elimination examines the proportion of buildings with low assessment levels that have been cumulatively transferred to renovation or completed repair and reinforcement in each year, reflecting the pace of achieving safety goals; the renovation coverage ratio examines the proportion of all buildings that have moved away from their current state of maintenance, reflecting the overall progress; the account occupancy balance examines the occupancy rate of the fund account and the construction work area account, as well as the balance level of the resettlement housing account in each year; accounts with occupancy rates consistently close to the upper limit or balances consistently low are considered constrained resources; the waiting time for relocation examines the annual interval from when each building first enters the candidate set for vacating and awaiting demolition, to when it is actually included in the vacating and awaiting demolition due to the availability of resettlement housing, and the number of waiting years due to insufficient housing resources during the simulation process; these four dimensions together constitute the comparison conclusions between scenarios.

[0104] The purpose of this comparison is to answer specific questions in the implementation plan preparation. When the funding amount is adjusted in a certain year, the adjusted amount is used as the intervention component for that year to be re-analyzed. By comparing the new path with the original path building by building and year by year, a list of buildings that need to be moved forward and a list of buildings that need to be moved backward can be obtained, as well as the position changes of the waiting links in the relocation chain. This comparison does not rely on manual rearrangement of tables, and due to the existence of account conditions, there will be no unexecutable arrangements such as arranging vacating when there are no houses available for vacating.

[0105] In one embodiment of the present invention, the evaluation answers the difference between different scenarios, but does not answer why a certain building enters a certain state in a certain year. Attribution addresses this question by using a conventional sensitivity analysis method of sequentially masking the inputs and then comparing the outputs. The purpose is to make the inference conclusions traceable back to specific inputs, rather than training a new model. Attribution calls the non-homogeneous state transition kernel obtained during the training phase and the mean vector of each input recorded during that phase on the training set, without changing the network parameters. The method is to mask the input of the transition kernel sequentially for a specified building and year and recalculate the transition probability forward, taking the change in probability as the contribution of that input. When masking, the path is not deleted, but replaced by the mean vector of that path on the training set, so as to keep the input dimension and network structure unchanged. During attribution, numerical attributes and categorical attributes are merged into one attribute path, and together with the current state, number of years of stay, neighborhood context, and intervention code, a total of five paths are sequentially masked to obtain five contribution values.

[0106] The interpretation of contribution volume corresponds to the on-site judgment; if the contribution volume of the neighboring context is positive and ranks first, it indicates that the relocation of the building in that year was mainly driven by the driving relationship after the completion of the neighboring building; if the contribution volume of the neighboring context is negative, it indicates that the construction of the neighboring building restricts the conditions for the commencement of the building in that year; if the contribution volume of the number of years of stay ranks first, it indicates that the building was included after the urgency increased due to the long-term stay of the building vacant and awaiting demolition; if the contribution volume of several pre-coded factors ranks first, it indicates that the relocation of the building was mainly driven by the investment of funds in that year or the inclusion in the annual plan.

[0107] The constraint trigger tracing is based on the extension records output in the previous stage; each extension record includes the building being extended, the original year, the triggering account type, and the number of extension years. By summarizing by account type, we can obtain the number of extensions caused by each account and the buildings involved. By summarizing by building, we can obtain the cumulative number of extension years for each building. This summary corroborates the aforementioned account occupancy rate distribution and together points out the resource links that restrict the succession rhythm.

[0108] The final output includes three items: the building renovation implementation timeline, which provides the evolution status and corresponding work content of each building in each year; the account constraint trigger record, which provides the occupancy status, extension list and waiting stage position of the three types of accounts in each year; and the key driving factor list, which provides the five contribution rankings of key buildings. These three items together constitute the deduction results that can be reviewed and discussed by the renovation center, and each conclusion can be traced back to the corresponding input evidence and constraints.

[0109] In one embodiment of the present invention, the above method is applied to the preparation of a five-year implementation plan for a contiguous renovation and renewal unit in a central urban area; the unit adopts the on-site rolling resettlement model, and the unit contains both dilapidated buildings that need to be demolished and rebuilt, as well as buildings that only need to be repaired, reinforced and have elevators installed. All resettlement housing comes from the resettlement buildings that were built in the unit in the early stage.

[0110] Following the above process, the building outline vector, five phases of orthophotos, two phases of street view images, building safety appraisal ledgers and implemented project ledgers of the unit are first collected to construct a building succession object map; some initial building data obtained after benchmark unification and entity matching are shown in Table 1, where the succession status is the initial assignment of the starting year of the projection, the number of years of residence is the number of years the building has been in the current state, the visual accessibility is the value of the non-street facade of the building after vegetation attenuation, and the roof visibility is the value after deducting the obstruction belt of adjacent buildings.

[0111] Table 1, Example of initial building data for updated unit.

[0112]

[0113] As shown in Table 1, the visibility values ​​of buildings B015 and B031 are relatively low, as both buildings are located inside the unit and have no street view coverage on their non-street-facing facades. During the visibility quantification stage, the visibility coefficients of the two buildings for the two states of repair and reinforcement and functional replacement approach the lower limit, while the visibility coefficients for the two states of demolition under construction and renovation completed remain at a high level due to the support of roof visibility. During the label construction stage, the two buildings were assigned a negative intensity approaching the lower limit for the situation where repair and reinforcement were not detected in the historical years, and therefore no longer provided a strong negative signal to the transfer core. If this treatment is not done, the probability of repair and reinforcement transfer for buildings like B015 and B031 will be pushed to zero, and the extrapolation results will lock them in the status quo for a long time, which is inconsistent with the actual pipeline renovation records of the unit.

[0114] During the nuclear transfer training phase, B031, having been classified as Level D and maintained in its current state for nine years, has its years of residence, along with its safety assessment level, significantly increasing the probability of it being vacated for demolition. B007, already vacated for demolition and having remained for two years, has a high probability of being demolished and under construction in the candidate set. There is a passageway between B022 and B007, meaning that the construction impact range of B007 cuts off the passageway from B022 to the unit entrance. Therefore, when B007 is transferred to demolition and under construction in a certain year, the modulation coefficient of B022 for that year is negative, and the probability of it being transferred to subsequent repair and reinforcement procedures is reduced.

[0115] During the rolling simulation phase, in the first year's candidate set, B007, which transitioned from vacant and awaiting demolition to demolition and under construction, and B031, which transitioned from maintaining the status quo to vacant and awaiting demolition, both had high ranking ratios. Due to limited funds remaining after prioritizing the construction of the first batch of resettlement buildings and the renovation of existing buildings, B007's demolition candidate was postponed due to insufficient funds, and B031's vacant status candidate was postponed due to a zero balance in the resettlement housing account. The first batch of resettlement buildings were constructed on initial vacant land within the unit, transitioning from planned new construction to demolition and under construction. Also included in the same year were B022 (continued repair and reinforcement) and B044 (completed functional placement). In the second year, the first batch of resettlement buildings were converted to new buildings and designated for resettlement. The number of resettlement units they could provide was carried over to the remaining resettlement housing stock at the beginning of the third year. In the third year, when the resettlement housing stock was in place and funds could support the demolition, B031 was converted to vacant land awaiting demolition and resettled by the first batch of resettlement buildings. B007 was converted to demolition under construction and demolition was completed in the same year. The land released was returned to the planned resettlement land for the construction of subsequent batches of resettlement buildings. B007 entered demolition under construction in the same year, cutting off the passage from B022 to the unit entrance. The modulation coefficient of B022 turned negative, and its subsequent repair and reinforcement procedures were reduced in the same year.

[0116] During the assessment and attribution phase, the funding amount for the third year was reduced and re-analyzed. The comparison results showed that B031 was delayed by one year from the year it was transferred to the vacating and demolition phase. The relocation chain experienced a wait in the third year, and the reason for the wait was that the account trigger record pointed to the funding account rather than the resettlement housing account. The result of B022 being transferred to the subsequent work of repair and reinforcement in the fourth year was attributed one by one. The contribution of the neighboring context turned from negative to positive, which corresponds to the process of B007 completing demolition and removing construction fences in the third year. The deduction results of this unit and the account trigger record were submitted to the update center for consultation to determine the priority funding direction and the batch of resettlement buildings that need to be started in advance in that year.

[0117] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An intelligent deduction method for the temporal evolution of buildings in urban renewal, characterized in that, Includes the following steps: S1 collects building outline vectors, multi-period images, safety assessments and project ledgers. After standardization and unique cross-year coding of entities, it constructs a building succession object map with individual buildings as nodes and adjacent and access edges as edges, and outputs image slice indexes. S2, on the object map and image tile index, through facade line intersection, roof occlusion calculation and state mode detection rate statistics, the visibility of each successive state is quantified to obtain a set of visibility coefficients coupled with the state category; S3, based on the image slice index, makes a change determination, combines the visibility coefficient set, divides the intervals and weights the visibility to make a confidence assignment, and forms a succession label set containing three types of samples; S4, using the succession label set and object graph attributes as training basis, constructs a neighborhood message passing network that generates positive and negative modulation coefficients from the neighbor succession stage, and trains to obtain a building-by-building and year-by-year non-homogeneous state transition kernel. S5. Based on the candidate transfer probabilities given by the transfer core, under the condition of conservation of the three types of accounts (funds, construction work area, and resettlement housing), the candidates for default are included in descending order by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, and the multi-year building-level succession path, the annual trajectory of the three types of accounts, and the extension record are obtained. S6 performs scenario-based assessments of multi-year building-level succession paths, annual trajectories of three types of accounts, and continuation records. It calls the transfer kernel to attribute transfer contributions and outputs a building renewal implementation timeline, account constraint trigger records, and a list of key driving factors.

2. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 1, characterized in that, In S1, the unique code for entities spanning different years is determined by the ratio of the overlapping area of ​​the building outlines of different years to their respective areas. If the ratio reaches the threshold of the overlapping area ratio, they are identified as the same building entity and share the same code. Adjacent edge is established and recorded as normalized spacing when the minimum distance between two building outlines is within the adjacent distance threshold or when they share a gable wall; Passage edge is the outer width of the construction influence range determined according to the building height and structural type. The influence range surface is obtained by buffering the building outline according to the outer width. The shortest path to the unit entrance is solved along the center line of the internal road. When the shortest path intersects with the influence range surface, a passage blocking sign is established and set to true.

3. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 1, characterized in that, The S2 facade sight line intersection is to decompose the building outline into facades, project rays from the sampling midpoint of each facade to the street view collection point and intersect with the rest of the building outlines, take the ratio of the number of collection points that are not blocked and whose incident angle does not exceed the incident angle threshold to the number of theoretically observable collection points of the facade, and then attenuate the facade by the vegetation shading rate to obtain the facade sight line accessibility. Roof shading is calculated as the ratio of the remaining area after deducting the shading strip formed by adjacent buildings at the roof edge from the roof area to the roof area, thus obtaining roof visibility. The geometric observability of each successive state is taken as the visual accessibility of the corresponding facade in the street view mode and as the roof visibility in the orthographic mode.

4. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 3, characterized in that, The state mode detection rate statistics in S2 are obtained by retrospectively judging whether each successive state presents identifiable features under orthophoto and street view modes based on the ledger records with completion time. The detection rate is then subjected to hierarchical backsliding smoothing and a lower limit is set. The temporal coverage is obtained by the ratio of the number of effective time phases in the interval to be judged to the interval span and by attenuation based on the longest gap window. The visibility coefficient set is synthesized independently of the geometric observability, detection rate and temporal coverage of each building in each year and each successive state under each mode, taking the complement of the probability of missed detection in each mode.

5. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 1, characterized in that, In S3, the interval assignment is divided as follows: when the interval between two adjacent images is one year step and both can be determined, the target transfer is recorded as a confirmed sample; when the interval spans multiple years, it is recorded as an interval sample and its year interval is recorded; when a certain building does not detect a certain type of state feature in a certain year, it is recorded as a weak negative sample; and the disappearance of the roof outline corresponds to the transfer to demolition under construction, the addition of the building outline corresponds to the transfer to the renewal and completion, and the change of the roof and facade features corresponds to the transfer to repair and reinforcement. The confidence assignment for visibility-weighted samples is the upper confidence limit, the interval sample is the likelihood of at least one target shift occurring within the interval of that year, and the weak negative sample is the visibility coefficient corresponding to the visibility coefficient set. The samples are organized into layers according to state category and visibility.

6. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 1, characterized in that, The positive and negative modulation coefficients in S4 are generated by splicing the current state embedding of the neighbors on each edge, the current state embedding of the body, the passage blocking flag and the normalized spacing, and then passing through two fully connected layers and hyperbolic tangent. The values ​​are between negative one and positive one and are not normalized during aggregation. The neighborhood message passing network multiplies the latent attribute vectors of each neighbor by their modulation coefficients and sums them along the edges. It then aggregates the adjacent edges and the passing edges separately and concatenates them into a neighborhood context. The latent attribute vectors, category attribute vectors, current state embeddings, years of stay embeddings, neighborhood contexts, and intervention codes are then concatenated and processed by a fully connected layer and a normalized exponent to obtain the probability distribution of transitions from the current state to each state, which serves as a row of the non-homogeneous state transition kernel.

7. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 6, characterized in that, The objectives of training the non-homogeneous state transition kernel in S4 include the classification loss of the confirmation samples, the interval likelihood loss of the interval samples, the negative loss of the weak negative samples weighted by the visibility coefficient, and the sparse regularization of the modulation coefficients that can be positive or negative. Training is rolled over annually. For years with observed states, the observed states are used as the input for the next year. For years without observed states, the inference results from the previous year are used. The network parameters are updated by iterating through sample batches organized by state category and visibility.

8. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 1, characterized in that, In S5, transitions with a transition probability exceeding the candidate probability threshold and a target state different from the current state are selected as candidate transitions; the urgency is composed of the security assessment level ordinal number and the number of years the current state has been in existence, and increases as the assessment level decreases or the number of years the current state has been in existence increases. The candidates are ranked in descending order of the ratio obtained by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, and are included in the three types of accounts for pre-verification. Among them, the funds account verification shows that the cumulative amount of funds occupied by the candidates does not exceed the annual funds limit. The construction work area account verification shows that the work area occupied by the buildings under demolition and construction does not exceed the upper limit of the same period and does not constitute a traffic conflict with the candidates already included in the demolition and construction. The resettlement housing account verification shows that its balance is not negative after deduction and the number of resettlement units added by the current year is carried forward to the next year. Candidates that fail the verification are postponed to the next year and their number of years of stay is accumulated.

9. The intelligent deduction method for the temporal evolution of buildings in urban renewal according to claim 1, characterized in that, The scenario comparison assessment in S6 involves comparing the progress of dilapidated building elimination, the renewal coverage ratio, the account occupancy status, and the waiting time for relocation among different scenarios obtained by changing the annual funding amount, the upper limit of the work area, or the annual plan scope. The transfer contribution is attributed to the input of the specified building and the annual path-by-path masked transfer kernel, and then recalculated forward after replacing it with the mean vector of that path on the training set. The probability change is taken as the contribution of each path in the attributes, current state, number of years of stay, neighborhood context and intervention code. The account constraint trigger record is summarized by the extension record according to the account type and building, to obtain the extension number of each account and the cumulative extension years of each building.

10. An intelligent simulation system for the temporal evolution of buildings in urban renewal, used to execute the steps in the intelligent simulation method for the temporal evolution of buildings in urban renewal as described in any one of claims 1-9, characterized in that, include: The object graph construction module is used to collect building outline vectors, multi-period images, safety assessments and project ledgers. After standardization and unique cross-year coding of entities, it constructs a building succession object graph with individual buildings as nodes and adjacent and access edges as edges, and outputs image tile indexes. The observation quantification module is used to quantify the visibility of each successive state on the object map and image tile index by finding the intersection of facade lines, calculating roof occlusion, and statistically analyzing the state mode detection rate, and to obtain a set of visibility coefficients coupled with the state category. The succession label construction module is used to determine changes based on the image slice index, combine the visibility coefficient set, divide the intervals and weight the visibility to make confidence assignments, and form a succession label set containing three types of samples. The state transition kernel training module is used to construct a neighborhood message passing network with positive and negative modulation coefficients generated by the neighbor succession stage, based on the succession label set and object graph attributes, and trains a non-homogeneous state transition kernel for each building and each year. The rolling simulation module is used to include and extend default candidates in descending order by multiplying the transfer probability by the urgency and dividing by the amount of funds occupied, based on the candidate transfer probabilities given by the transfer core, under the condition of conservation of the three types of accounts: funds, construction work surface and resettlement housing resources. This results in multi-year building-level succession paths, annual trajectories of the three types of accounts and extension records. The assessment and attribution module is used to conduct scenario-based assessments of multi-year building-level succession paths, annual trajectories of three types of accounts, and continuation records. It calls the transfer kernel to perform transfer contribution attribution and outputs building renewal implementation sequence plans, account constraint trigger records, and a list of key driving factors.

Images