Stone cultural relics salt damage multi-modal image recognition and spatio-temporal evolution analysis method

Abstract

The present application relates to the technical field of computer vision, and discloses a stone cultural relic salt damage multi-modal image recognition and space-time evolution analysis method. The surface reference image, the infrared thermal response image, the near-infrared hyperspectral image and the shallow layer resistivity distribution image are registered to a unified pixel grid; a modal effectiveness code is generated according to a local registration residual and an imaging quality, a corresponding feature fusion sub-model is selected, and a salt damage quantitative image, a salt damage type semantic segmentation image and an identification confidence image are output; a salt return risk image and a spatial topological feature image are generated from continuous images, a salt damage analysis area is identified as a salt migration active, connected residual, discrete residual, stable convergence or to-be-rechecked state in combination with a state prior, and a space-time evolution analysis image is output. The present application can reduce the influence of moisture, mineral substrate and local modal failure on salt damage recognition and continuous state determination.

Description

Technical Field

[0001] This invention belongs to the field of image processing and computer vision technology, specifically relating to a method for multimodal image recognition and spatiotemporal evolution analysis of salt damage to stone cultural relics. Background Technology

[0002] Salt efflorescence, powdering, and flaking on the surfaces of grotto temples, cliff carvings, and ancient architectural stone components exhibit similar morphologies but differ in their causes. Infrared thermal imaging, near-infrared hyperspectral imaging, visible light texture imaging, and shallow surface resistivity imaging respectively reflect thermal response, spectral absorption, surface morphology, and conductive connectivity characteristics, but each type of image differs in spatial resolution, imaging geometry, and response scale. Moisture, mineral substrate, and surface roughness can also alter spectral and thermal responses, and a single image or threshold can easily misidentify damp areas, mineral differences, and shadows as salt damage. During continuous monitoring, existing methods typically process images at different times independently, making it difficult to maintain consistency in pixel coordinates, region identification, and recognition status, and also difficult to simultaneously describe changes in salt content, shrinkage of connected salt bands, and the presence of discrete high-salt spots. Therefore, a method for analyzing salt damage images of stone cultural relics is needed that can perform cross-modal registration, pixel-level fusion recognition, uncertainty preservation, temporal feature extraction, and spatial topological recognition. Summary of the Invention

[0003] This invention provides a method for multimodal image recognition and spatiotemporal evolution analysis of salt damage to stone cultural relics. The method registers a surface reference image, an infrared thermal response image, a near-infrared hyperspectral image cube, and a shallow surface resistivity distribution map to a unified pixel grid. Based on local registration residuals and the imaging quality of each modality, a modal validity coding map is generated, and a feature fusion sub-model corresponding to the effective modal combination is selected for each pixel. Pixel-level fusion recognition is supervised by ion chromatography detection results, generating a quantitative salt damage map, a semantic segmentation map of salt damage types, a recognition confidence map, and a pore connectivity map. A salt return risk map and a spatial topological feature map are generated from consecutive image iterations to identify the salt damage evolution state, and the state map of the previous time step constrains the region inheritance and state update of the next time step.

[0004] The technical solution adopted in this invention is as follows: A method for multimodal image recognition and spatiotemporal evolution analysis of salt damage to stone cultural relics, including the following steps: Step 1: Acquire a surface reference image, an infrared thermal response image sequence, a near-infrared hyperspectral image cube, and a shallow surface resistivity distribution map of the stone artifact surface. Register the infrared thermal response image sequence, the near-infrared hyperspectral image cube, and the shallow surface resistivity distribution map to the surface coordinate system of the surface reference image and resample them to a unified pixel grid to form a multimodal image group. Generate a modal validity coding map based on the local registration residuals, surface texture imaging quality, spectral imaging quality, thermal response imaging quality, and resistivity imaging quality of each pixel grid unit. Use sampling measurement... The ion chromatography detection results of the points serve as supervisory labels. The spectral reflectance features, thermal response features, local texture features, and resistivity connectivity features of each pixel grid unit are extracted. Using a multimodal salt damage identification model, a feature fusion sub-model corresponding to the effective mode combination is selected according to the modal effectiveness coding map to generate a salt damage quantitative map, a salt damage type semantic segmentation map, and an identification confidence map. Pixels without matching feature fusion sub-models or whose identification confidence does not meet the preset requirements are written into a low-confidence unresolved mask, and a pore connectivity map is generated based on the local texture features and the resistivity connectivity features. Step 2: Divide the salt damage analysis region according to the spatial continuity relationship between the semantic segmentation map of the salt damage type and the porosity connectivity map; repeat Step 1 for the multimodal image group at different times to obtain the salt damage quantitative map sequence of each salt damage analysis region; extract the remaining salt time series from the salt damage quantitative map sequence and establish a fractional-order salt migration model; map the predicted salt return flux to the unified pixel grid to generate a salt return risk map; construct a cubic complex with the salt damage quantitative map at the current time and perform upper level set filtering to obtain 0-persistence barcodes and 1-persistence barcodes; map the connected components and independent loops corresponding to the barcodes to the unified pixel grid to generate a spatial topological feature map; Step 3: For each salt damage analysis area, extract regional-level fusion features from the salt damage quantitative map, the salt damage type semantic segmentation map, the identification confidence map, the salt return risk map, and the spatial topological feature map. Based on the regional-level fusion features, identify the salt damage analysis area as an active salt migration state, a connected residual state, a discrete residual state, a stable convergence state, or a state awaiting verification. Backfill the identification state of each salt damage analysis area into the unified pixel grid to generate a spatiotemporal evolution analysis map with state boundaries and identification confidence. Update the spatiotemporal evolution analysis map using the multimodal image group at the next time step.

[0005] Further, in step 1, a curved coordinate model of the stone artifact surface is established using three-dimensional laser scanning. Band images with a signal-to-noise ratio that meet preset conditions are selected from the near-infrared hyperspectral image cube as the surface reference image. The infrared thermal response image sequence is continuously acquired before, during, and after controlled thermal excitation. Dark current correction and standard reflector correction are performed on the near-infrared hyperspectral image cube. Geometric coarse registration is completed based on surface control points, and cross-modal fine registration is completed based on mutual information and stable texture edges. Non-reference modes with registration residuals exceeding half the side length of the unified pixel grid unit are marked as invalid in the corresponding pixel grid unit.

[0006] Furthermore, the multimodal salt damage identification model includes multiple feature fusion sub-models corresponding to different effective modal combinations, each of which is a multi-output partial least squares regression model; the spectral reflectance features include the first derivative spectrum after standard normal variable transformation and water absorption band features; the thermal response features include thermal excitation heating amplitude, cooling slope, and thermal response time; the local texture features include local gradient and local variance; and the resistivity connectivity features include local resistivity and its distance to low resistivity connectivity channels. A feature fusion sub-model with the same modal validity encoding as the current pixel grid cell and independently verified is selected, outputting predicted values ​​and identification confidence levels for chloride ion content, sulfate ion content, and nitrate ion content. When the sum of the three anion contents is higher than the quantification limit, the one with the highest proportion is determined as the dominant anion salt damage type of the current pixel grid cell; when the sum of the contents is not higher than the quantification limit, the current pixel grid cell is determined as a salt-free background type. When no matching and independently verified feature fusion sub-model exists, the current pixel grid cell is written into a low-confidence unresolved mask.

[0007] Furthermore, the monitoring labels are jointly provided by standard stone test blocks of the same lithology and the sampling points; when the area of ​​a continuous low-confidence region exceeds a preset area or the content gradient at the boundary of a salt damage type exceeds a preset gradient, supplementary sampling points are selected at the continuous low-confidence region or the boundary of the salt damage type, and the multimodal salt damage identification model is updated using the supplementary ion chromatography detection results; for pixel grid units that have not obtained effective monitoring labels and whose identification confidence does not meet the requirements, a low-confidence unresolved mask is retained, and the predicted values ​​of adjacent pixels are not used for filling.

[0008] Furthermore, a portable nuclear magnetic resonance single-sided probe was used to collect transverse relaxation signals at sampling points. After calibration with the vacuum saturated porosity of standard stone blocks of the same lithology, the equivalent connected porosity was obtained. Gaussian process regression was used to generate an equivalent connected porosity map and an estimation uncertainty map. Linear fracture features were extracted from the surface reference image, and low resistivity connected channels were extracted from the shallow surface resistivity distribution map. The overlapping and adjacent regions of the two in a unified pixel grid were determined as equivalent conductive connected channels. The porosity connectivity map was generated based on the distance from each pixel grid unit to the equivalent conductive connected channel, and a region growing algorithm constrained by the dominant anion salt damage type and the equivalent connected porosity was used to divide the salt damage analysis area.

[0009] Further, in step 2, repeatable monitoring windows are set up in each salt damage analysis area. Infrared thermal response images and near-infrared hyperspectral images of the monitoring windows are acquired at multiple sampling times. The representative values ​​of residual salt content at each sampling time are obtained using the multimodal salt damage identification model and the residual salt content time series is formed. Using the residual salt content time series as the fitting target, the memory order, apparent migration coefficient, and boundary release term of the Caputo fractional salt migration model are inverted using the Levenburg-Marquardt nonlinear least squares method. The predicted salt return flux within the set extrapolation period is obtained using the Greenwald-Letnikov difference scheme. The predicted salt return flux is then backfilled into the salt return risk map according to the salt damage analysis area boundary and identification confidence level.

[0010] Furthermore, each pixel grid cell of the salt damage quantitative map is treated as a 2D cell of a cubic complex, the shared edges and outer boundaries of adjacent pixel grid cells are treated as 1D edges, and the corner points of pixel grid cells are treated as 0D vertices. The cubic complex is filtered by a horizontal set according to the remaining salt content from high to low. Boundary matrix reduction based on a binary finite field is used to track the birth and death of 0D connected components and 1D independent loops, generating the 0-dimensional persistent barcode and the 1-dimensional persistent barcode. The pixel positions corresponding to the finite barcodes are written into the spatial topological feature map.

[0011] Further, the determination order of the identification state is as follows: when the area proportion of the low-confidence unresolved mask in the salt damage analysis area reaches a preset proportion, it is determined to be the state to be reviewed; when it does not reach the preset proportion, if the maximum predicted salt return flux in the extrapolated time period is not lower than the preset salt return stability threshold, it is determined to be the active salt migration state; if the maximum predicted salt return flux is lower than the preset salt return stability threshold and there is a barcode in the 1-permanent barcode with a lifetime not lower than the preset connectivity convergence threshold, it is determined to be the connected residual state; if the maximum predicted salt return flux is lower than the preset salt return stability threshold, the lifetime of the 1-permanent barcode is lower than the preset connectivity convergence threshold, and the 95th percentile of the soluble salt content in the salt damage quantitative map is not lower than the preset upper limit of residual soluble salt or the longest lifetime of the 0-dimensional finite barcode is not lower than the preset residual spot convergence threshold, it is determined to be the discrete residual state; the remaining salt damage analysis areas are determined to be the stable convergence state.

[0012] Furthermore, in step 3, the spatiotemporal evolution analysis diagram of the previous moment is used as the state prior for the next moment; when the identification state of the same salt damage analysis area changes, the spatiotemporal evolution analysis diagram is updated only after the same new identification state is obtained for two consecutive moments; when the salt damage analysis area splits or merges, the original identification state is inherited according to the maximum overlap area between regions and the region-level fusion feature is recalculated; when the area ratio of the low-confidence unresolved mask reaches the preset ratio, the inherited identification state is covered by the state to be reviewed.

[0013] Furthermore, the modal validity coding map, salt damage quantitative map, salt damage type semantic segmentation map, identification confidence map, pore connectivity map, salt return risk map, spatial topological feature map, and spatiotemporal evolution analysis map are output; the model is updated only using the supervision labels.

[0014] By adopting the above technical solution, the present invention has the following beneficial effects: Infrared thermal response images, near-infrared hyperspectral images, surface reference images, and shallow surface resistivity distribution maps are registered to a unified pixel grid. Modal validity codes are generated based on local registration residuals and the imaging quality of each modality, preventing locally failed modes from entering mismatched feature fusion sub-models and reducing misidentification of salt damage caused by differences in moisture, mineral substrate, and imaging quality. Unresolved masks are retained for pixels without verified co-coded sub-models or with insufficient recognition confidence, and are not filled with predicted values ​​from adjacent pixels, preventing low-confidence results from propagating along the spatial topology analysis process. A salt return risk map is generated from the continuous salt damage quantitative map, and the connected components and independent loops corresponding to the persistent barcode are back-mapped into a spatial topology feature map, simultaneously characterizing the temporal changes of salt damage, connected salt bands, and discrete salt spots. The spatiotemporal evolution analysis map is updated using the prior state of the previous moment, confirmation from two consecutive moments, and the maximum overlap inheritance rule of split or merged regions, reducing image noise and state jumps caused by slight fluctuations in regional boundaries, and maintaining the correspondence between pixel coordinates, regional boundaries, and recognition states at each moment. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the principle of cross-modal registration and modal validity coding space correspondence provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the multimodal fusion recognition output space provided in an embodiment of the present invention; Figure 3 Fractional order salt migration fitting and salt return flux extrapolation curves provided in this embodiment of the invention; Figure 4 A schematic diagram illustrating the cross-sectional application of spatiotemporal evolution analysis results provided in an embodiment of the present invention for partition repair; Figure 5 This is a schematic diagram of the dual-scale salt migration image features of fractures and matrix provided in an embodiment of the present invention; Figure 6 This is an example diagram of horizontal set filtration and spatial topological feature extraction on the salt damage quantitative map provided in this embodiment of the invention; Figure 7 This is a schematic diagram of the time-series application of the spatiotemporal evolution state diagram provided in the embodiments of the present invention for asynchronous processing in multiple regions. Detailed Implementation

[0016] The method for multimodal image recognition and spatiotemporal evolution analysis of salt damage to stone cultural relics in this embodiment is implemented according to the following steps: Step 1: Acquire a surface reference image, an infrared thermal response image sequence, a near-infrared hyperspectral image cube, and a shallow surface resistivity distribution map of the stone artifact surface. Register the infrared thermal response image sequence, the near-infrared hyperspectral image cube, and the shallow surface resistivity distribution map to the surface coordinate system of the surface reference image and resample them to a unified pixel grid to form a multimodal image group. Generate a modal validity coding map based on the local registration residuals, surface texture imaging quality, spectral imaging quality, thermal response imaging quality, and resistivity imaging quality of each pixel grid unit. Use sampling measurement... The ion chromatography detection results of the points serve as supervisory labels. The spectral reflectance features, thermal response features, local texture features, and resistivity connectivity features of each pixel grid unit are extracted. Using a multimodal salt damage identification model, a feature fusion sub-model corresponding to the effective mode combination is selected according to the modal effectiveness coding map to generate a salt damage quantitative map, a salt damage type semantic segmentation map, and an identification confidence map. Pixels without matching feature fusion sub-models or whose identification confidence does not meet the preset requirements are written into a low-confidence unresolved mask, and a pore connectivity map is generated based on the local texture features and the resistivity connectivity features. Step 2: Divide the salt damage analysis region according to the spatial continuity relationship between the semantic segmentation map of the salt damage type and the porosity connectivity map; repeat Step 1 for the multimodal image group at different times to obtain the salt damage quantitative map sequence of each salt damage analysis region; extract the remaining salt time series from the salt damage quantitative map sequence and establish a fractional-order salt migration model; map the predicted salt return flux to the unified pixel grid to generate a salt return risk map; construct a cubic complex with the salt damage quantitative map at the current time and perform upper level set filtering to obtain 0-persistence barcodes and 1-persistence barcodes; map the connected components and independent loops corresponding to the barcodes to the unified pixel grid to generate a spatial topological feature map; Step 3: For each salt damage analysis area, extract regional-level fusion features from the salt damage quantitative map, the salt damage type semantic segmentation map, the identification confidence map, the salt return risk map, and the spatial topological feature map. Based on the regional-level fusion features, identify the salt damage analysis area as an active salt migration state, a connected residual state, a discrete residual state, a stable convergence state, or a state awaiting verification. Backfill the identification state of each salt damage analysis area into the unified pixel grid to generate a spatiotemporal evolution analysis map with state boundaries and identification confidence. Update the spatiotemporal evolution analysis map using the multimodal image group at the next time step.

[0017] This embodiment describes multimodal image acquisition and registration, pixel-level salt damage identification, salt return risk map generation, spatial topological feature extraction, and evolution state identification in the order of steps 1, 2, and 3. This embodiment uses the sandstone grotto wall as the image acquisition object. Before implementation, cultural relic protection professionals determine the areas where imaging, contact detection, and micro-sampling are permitted. Carved details, pigment layers, hollow edges, and structurally unstable areas are designated as prohibited sampling areas. The image analysis terminal saves the acquisition time, equipment calibration file, surface coordinate transformation relationship, sampling number, model version, and identification status for each image, ensuring traceability of the same pixel grid unit at all times.

[0018] First, a unified surface coordinate system is established. A handheld 3D laser scanner is used to acquire point clouds of the surface to be repaired, with a point spacing of no more than 1 mm. Near-infrared hyperspectral image blocks with high signal-to-noise ratios and minimal water absorption are selected as the surface reference images. The point cloud is parametrically unfolded according to local surface parameters, and a uniform 20 mm × 20 mm pixel grid is constructed on the unfolded surface. A 10 mm × 10 mm grid is used for carving details, while a 50 mm × 50 mm grid can be used for large, flat stone walls. Each pixel grid unit stores a surface area correction coefficient to prevent deviations caused by directly calculating the planar area after the surface is unfolded. After fine registration, a 4-bit modal validity code is generated for each pixel grid cell. These four bits correspond sequentially to surface texture, near-infrared hyperspectral data, infrared thermal response, and shallow surface resistivity. The surface texture bit is set to valid or invalid based on the local Laplacian variance, saturated pixel ratio, and occlusion ratio. The remaining bits are set to valid or invalid based on the local registration residual and corresponding imaging quality. When the local Laplacian variance is lower than the 5th percentile of a clearly defined reference area of ​​the same lithology, the saturated pixel ratio exceeds 5%, or the occlusion ratio exceeds 10%, the surface texture bit is set to invalid. When the local registration residual exceeds half the side length of the uniform pixel grid cell, the corresponding non-reference modal bit is set to invalid.

[0019] The monitoring labels are provided jointly by standard stone test blocks of the same lithology and on-site micro-sampling. The basic standard test blocks are configured with 3 dominant anions, 5 salt content levels, 4 moisture content levels, and 2 parallel samples, totaling 120 blocks. The cation combination used in the salt preparation is selected based on the on-site total ion analysis results of soluble salts. Additionally, mixed salt test blocks are prepared according to typical on-site anion-cation ratios to verify the stability of the dominant anion label under mixed salt conditions. On-site sampling points are determined by both visible disease coverage and spatially uniform coverage, typically 12 to 30 points. After approval, samples of 2 to 5 mg in depth, not exceeding 1 mm, are collected from non-carved details. Samples are extracted with deionized water at a fixed solid-liquid ratio and filtered through a membrane. The concentrations of chloride, sulfate, and nitrate ions in the solution are measured by ion chromatography, and then converted to sample mass content based on the volume of the extract and the sample mass. The detected values ​​are linked to the surface coordinates of the sampling points. The cation detection results are used to select the standard test block formulation and are not used as the output of the multimodal salt damage identification model.

[0020] Infrared thermal response images were acquired using an uncooled infrared thermal imager with a resolution of at least 640×480 and a thermal sensitivity better than 0.05 degrees Celsius. When active thermal excitation was permitted, a 30-second baseline was first recorded, followed by 120 seconds of excitation with a uniform low-energy heat source and a 300-second cooling process, controlling the maximum surface temperature rise to no more than 2 degrees Celsius. When active thermal excitation was not permitted, passive thermal response sequences were acquired during periods of natural ambient cooling. The near-infrared hyperspectral imaging device covered a range of 900 nm to 2500 nm, with a spectral resolution of 5 nm to 10 nm. Dark current correction and standard reflector correction were performed before and after acquisition. Cross-modal registration was first performed using surface control points and device pose for coarse registration, followed by fine registration using stable texture edges and normalized mutual information. If the fine registration residual exceeded half the side length of a uniform pixel grid unit, the region was re-acquired. If the re-acquisition still exceeded the specified limit, the corresponding non-reference modal bit was invalidated. The infrared thermal response mode bit is valid when the baseline drift of the infrared thermal response sequence is no greater than 0.10 degrees Celsius and the spatial variation coefficient of temperature rise in the excitation region is no greater than 10%; the near-infrared hyperspectral mode bit is valid when the signal-to-noise ratio of the near-infrared hyperspectral pixel reaches the lower limit of the equipment calibration, saturation has not occurred, and the water absorption feature is within the coverage of the calibration sample.

[0021] Figure 1 This demonstrates cross-modal registration and modal validity coding in a unified surface coordinate system. Figure 1(a) Taking a local area of ​​a 20 mm × 20 mm grid as an example, the solid lines are near-infrared hyperspectral salt feature contour lines, the dashed lines are infrared thermal response contour lines before registration, and the dotted lines are shallow surface low resistivity connected channels; the hollow circles are surface reference image control points, the triangles are corresponding control points of the modes to be registered, and the arrows indicate the local registration residual vector. If the 6 mm residual is less than the 10 mm limit, the corresponding modal bit remains valid; if the 13 mm residual is greater than the 10 mm limit and still exceeds the limit after re-acquisition, the corresponding modal bit is set to invalid. Figure 1 (b) Write the 4-bit code of each grid cell into the same coordinate position, with the bit order being surface texture, near-infrared hyperspectral, infrared thermal response, and shallow surface resistivity; 1 indicates valid, and 0 indicates invalid. During recognition, only the feature fusion sub-model that is consistent with the 4-bit code and has been verified is called.

[0022] For each calibration sample, the first derivative spectrum of the corresponding pixel grid cell after standard normal transformation, water absorption band characteristics, thermal excitation heating amplitude, cooling slope, thermal response time, local gradient, local variance, local resistivity, and the distance to the lowest resistivity connected channel are concatenated as input features and paired with the content of the three anions given by ion chromatography. The multi-output partial least squares regression model is written as follows: .in, Input a feature matrix for the samples, with each row corresponding to a calibration sample; A matrix showing the content of three anions; This is the regression coefficient matrix; The residual matrix is ​​used. Standard stone test blocks are grouped by test block number for 5-fold cross-validation. Repeated images of the same test block are only assigned to the same fold. The number of latent variables is selected from 1 to 10, and the stopping condition is that the prediction residual of the validation fold in the grouped 5-fold cross-validation is minimized and continuously increasing the latent variables no longer significantly reduces the residual. The 4-bit modality validity code generates a maximum of 15 valid modality combinations, excluding all invalid combinations. Each combination corresponds to a feature fusion sub-model, and each feature fusion sub-model is an independently trained multi-output partial least squares regression model. Only when the same combination has no less than 20 mutually independent sample groups and the 5-fold cross-validation error passes the preset limit will it enter independent validation. Only after passing independent validation will it be registered as a callable sub-model. Fully modally valid calibration samples are formed by selecting feature blocks of the corresponding modality to form the training input for each combination. Combination samples derived from the same standard test block are always assigned to the same fold in cross-validation to prevent information leakage. During recognition, the same encoding sub-model is called according to the modal validity encoding of the pixel, and the mean of other combinations is not used to fill in invalid modalities; if there is no verified same encoding sub-model, the pixel is written into the low confidence unresolved mask.

[0023] The multimodal salt damage identification model was trained before being deployed for on-site identification. The training process included, in sequence, establishing calibration records, supervising label conversion and spatial binding, training data quality control, fixed physical scale feature extraction, input-output matrix construction, stratified partitioning of sample groups, training of sub-models with validity codes for each modality, grouped 5-fold cross-validation, refitting of the entire model set, independent validation, and saving training results. The above training process was performed separately for each 4-bit modality validity code except for the all-invalid code 0000; the sub-models with different codes formed input matrices, preprocessing parameters, and regression parameters respectively, without using the independent validation set to determine features, calculate preprocessing parameters, select the number of latent variables, or fit regression coefficients.

[0024] Each standard stone test block number and each field sampling point number constitute an indivisible sample group. Records obtained from the same test block or the same sampling point in different imaging rounds, different repeated exposures, or at different times retain the same sample group identifier and are only used as repeated observations within the same group, not as independent calibration samples. Each calibration record must at least include the sample group identifier, stone lithology, data version, acquisition time, ambient temperature and relative humidity, uniform pixel grid coordinates, sampling influence range, 4-bit modal validity code, original data index of each valid mode, physical size of the feature window, input feature vector, ion chromatography label vector, quantitation limits for each anion, and quality control indicators. After calibration records are generated, they are assigned a unique record number, which is used for subsequent data partitioning, cross-validation, regression parameter fitting, independent validation, and saving training results.

[0025] The training sample consisted of standard stone blocks of the same lithology, mixed salt blocks configured according to typical anion-cation ratios in the field, salt-free control blocks, and approved field micro-sampling points. The basic standard blocks were configured into 120 independent blocks according to 3 dominant anions, 5 salt content levels, 4 moisture content levels, and 2 parallel samples, to cover the content gradients and moisture states when chloride, sulfate, and nitrate ions were dominant, respectively. Salt-free control blocks were used to verify the salt-free background, mixed salt blocks were used to cover situations where two or more anions coexisted and the dominant anion ratios were similar, and field sampling points were used to cover the actual mineral substrate, surface roughness, weathering degree, and moisture state. The standard blocks, mixed salt blocks, and salt-free control blocks were subjected to multimodal imaging using the same imaging equipment and acquisition parameters as the stone artifacts to be identified.

[0026] Ion chromatography monitoring labels are generated in the following order: sample collection, soluble salt extraction, blank correction, ion chromatography quantification, and mass content conversion. For each sampling point, the sample dry mass, extract volume, dilution factor, and instrument-provided chloride, sulfate, and nitrate ion concentrations are recorded. The detection concentration is corrected using the corresponding ion concentration from the blank sample. The product of the corrected concentration, the extract volume, and the dilution factor is then divided by the sample dry mass to convert the sample mass content of the corresponding anion. The label vector for each sample is recorded in a fixed order of chloride, sulfate, and nitrate ions, with the label unit uniformly being sample mass content. Detections below the corresponding limit of quantitation (LOQ) retain both the converted value and the LOQ indicator; these detections are not included in the relative error statistics for exceeding the LOQ target. Salt-free background determination is performed separately according to the LOQ of the three anions. If the ion chromatography spiked recovery rate is not within the range of 80% to 120%, the relative deviation of parallel samples exceeds 10%, or the blank sample is abnormal, the corresponding label will not be included in the monitoring training.

[0027] To ensure a unique correspondence between chemical detection tags and image features, the coordinates of the sampling center in the 3D curved surface coordinate model are projected onto a unified pixel grid. The sampling influence range is determined based on the actual contact size of the sampling tool, sampling depth, and registration residual. Feature statistics are calculated only for pixels within the sampling influence range that are not occluded and pass the corresponding modal quality control. An imaging feature record is generated from this influence range and bound to an ion chromatography tag vector. The same ion chromatography detection value is not replicated as independent tags for multiple surrounding pixels. When sampling influence ranges overlap, their respective sample group identifiers are retained. When dividing the model construction set and independent validation set, records from the same sampling operation are placed in the same data subset to avoid duplicate tag counting and data leakage.

[0028] The 4-bit modal validity encoding uses a fixed bit order: surface texture, near-infrared hyperspectral, infrared thermal response, and shallow surface resistivity. A valid mode is recorded as 1, and an invalid mode as 0. Surface texture mode validity is determined based on sharpness, exposure range, and occlusion ratio; near-infrared hyperspectral mode validity is determined based on dark current correction, standard reflector correction, saturated pixel ratio, effective band ratio, and signal-to-noise ratio; infrared thermal response mode validity is determined based on baseline stability, thermal excitation uniformity, effective temperature rise, and the number of effective cooling sampling points; shallow surface resistivity mode validity is determined based on electrode contact impedance, effective measurement combination ratio, and inversion residual. If the local registration residual of any non-reference mode in the corresponding pixel grid cell exceeds half the side length of the grid cell, this bit is still recorded as 0, even if the original image quality meets the requirements.

[0029] Before calibration records are included in training, they undergo original file integrity checks, equipment calibration checks, imaging quality control, spatial binding quality control, and chemical detection quality control. Records with equipment calibration failures, incomplete original files, registration residuals between sampling coordinates and the unified pixel grid exceeding half the grid cell side length, insufficient number of effective pixels within the sampling influence range, sampling influence range crossing restricted areas, or failing ion chromatography quality control are excluded from training. Records with only partial image modalities failing quality control are retained, encoded with a 4-bit modality validity code according to the actual effective modalities, and used only for combinations of effective modalities or low-dimensional combination sub-models that can be constructed from their effective feature blocks; missing or invalid modalities are not filled in with zero values, mean values, predicted values ​​of adjacent pixels, inversion values ​​of other modalities, or manual interpolation results. Quality control results, exclusion reasons, and original file verification values ​​are saved together with the calibration records to prevent excluded records from re-entering the input matrix in subsequent training steps.

[0030] For near-infrared hyperspectral data that has passed quality control, saturated pixels and pixels with a signal-to-noise ratio below the device calibration limit are first removed within the sampling influence range. The median reflectance of the remaining effective pixels is calculated point by point according to wavelength and interpolated to the fixed wavelength sequence used in the training data. A standard normal variable transformation is performed on this median spectrum, that is, the average value of all effective wavelengths of the spectrum is used as the center and its standard deviation is used for scaling transformation; then the first derivative spectrum is formed by dividing the reflectance difference between adjacent wavelengths by the wavelength interval, and the endpoints are calculated using one-sided difference. A local continuous baseline is established between the pre-registered left and right shoulder wavelengths, centered on the absorption valleys near 1450 nm and 1940 nm. The maximum absorption depth of the absorption valley relative to the baseline is extracted, and the absorption area is calculated using trapezoidal integral. The fixed wavelength sequence used during training, the excluded bands, the left and right boundaries of the water absorption band, and the difference method are all written into the feature definition file. Each cross-validation fold and independent validation set of the model building set uses this fixed definition, and the bands should not be reselected or the absorption band boundaries changed based on the validation results.

[0031] For the infrared thermal response image sequence, the average temperature of the first 30 seconds before thermal excitation is used as the pixel baseline. After subtracting the baseline, the maximum temperature rise is calculated within the sampling influence range. The cooling fitting interval is defined from the moment of the maximum temperature rise to the moment when the temperature rise decays to 37% of the maximum temperature rise. Within this interval, a least-squares linear fit is performed with time as the independent variable and temperature rise as the dependent variable to obtain the cooling slope. The time required to reach the 37% temperature rise is taken as the thermal response time. When using a passive thermal response sequence, the start time of natural cooling is used instead of the end time of thermal excitation, while the rest of the calculation order remains unchanged. For the surface reference image, the brightness image is smoothed with a fixed-scale Gaussian smoothing method within the sampling influence range, and the horizontal and vertical gradients are calculated. The median gradient magnitude and the local variance of brightness are extracted. For the shallow surface resistivity distribution map, the median resistivity within the sampling influence range is extracted, and the surface distance from the center of this range to the nearest low-resistivity connected channel is calculated on the curved grid using the shortest path. All distances are converted to millimeters, all temperature differences are converted to degrees Celsius, and all times are converted to seconds.

[0032] The statistical range of training features uses a fixed physical size. For standard stone blocks, mixed salt blocks, and salt-free control blocks, the pre-marked sampling center is used as the window center; for on-site micro-sampling points, the position where the sampling center is projected onto a uniform pixel grid is used as the window center; the surface area of ​​each window is consistent with the sampling influence range formed by the sampling tool. All training records are formed into features according to the same rules for effective pixel selection, median statistics, gradient calculation, cooling fitting, and surface shortest path calculation. When a window crosses an image boundary, occluded edge, or prohibited sampling area and its effective area is less than the specified minimum area, the corresponding modality is marked as invalid and grouped according to the actual modality validity. Features with different physical scales from other training samples must not be formed by reducing the window size. The window side length, the allowed minimum effective area ratio, and the surface area conversion rules are written into the feature definition file and remain unchanged during cross-validation and independent validation.

[0033] The input feature vector for each calibration record is concatenated in a fixed order: first derivative spectrum, water absorption band characteristics, maximum temperature rise, cooling slope, thermal response time, median local gradient, local variance, median resistivity, and surface distance to the lowest resistivity connected channel. The output label vector is arranged in a fixed order: chloride ion, sulfate ion, and nitrate ion mass content. Feature names, units, data types, calculation windows, missing value criteria, and arrangement positions are all written into the feature definition file. When the model training program reads the output labels, it does not read the salt information contained in the sample name. The standard test block formulation information is only used for sample component stratification and training result verification and is not used as model input to prevent the output labels from being directly leaked through file names, formulation numbers, or manual category fields.

[0034] For each training fold, the mean and standard deviation of each input feature and each output label are calculated using only the records from that fold, and the centering and scaling transformations of the input and output matrices are performed respectively. When the standard deviation of an input feature sample is zero, that feature is removed from the currently encoded training matrix and recorded in the feature deletion mask. If any output label does not exceed the effective range of the quantitation limit in the training fold, no sub-model involving that label distribution is trained. The feature order, feature deletion mask, input mean, input standard deviation, output mean, and output standard deviation obtained from the training fold are used as is in the corresponding validation fold. The prediction results are restored to the mass content units based on the output mean and output standard deviation of the training fold before the validation error is calculated. After determining the number of latent variables and refitting using the entire model construction set, the independent validation set only uses the above parameters calculated using the final model construction set, without recalculating the mean, standard deviation, or feature deletion mask separately.

[0035] After completing the calibration records and feature vector construction, the data was partitioned according to sample groups rather than individual image records. First, a joint stratification was performed based on dominant anion type, salt content level, moisture content level, and sample source, dividing the sample groups into model building sets and independent validation sets, with the model building set accounting for 80% and the independent validation set accounting for 20%. The 120 basic standard test blocks were divided into 96 model building sample groups and 24 independent validation sample groups; mixed salt test blocks, salt-free control test blocks, and field sampling points were also divided as a whole according to sample groups, without splitting the same group for repeated observations. After partitioning, the difference in the marginal sample ratio between the model building set and the independent validation set for the three dominant anions, five salt content levels, and four moisture content levels did not exceed 10 percentage points, and each dominant anion, each salt content level, each moisture content level, and salt-free background had a sample group in the independent validation set. If the above retention conditions were not met, the sub-models involving that category were not registered until the corresponding independent validation samples were added. Once the independent validation set is determined, it is sealed and does not participate in band selection, feature definition adjustment, calculation of centering and scale parameters, selection of the number of latent variables, determination of stopping conditions, or fitting of regression coefficients.

[0036] The model build set is then grouped and cross-validated in 5-fold according to the test block number and the field sampling point number. A fixed grouping list is used to distribute the number of sample groups as evenly as possible among the 5 folds, while maintaining similar distributions of dominant anion type, salt content level, water content level, and sample source in each fold; all imaging rounds, repeated exposures, and repeated calculation records of the same sample group are always assigned to the same fold. Combined records formed by selecting different feature blocks from the same full-modality effective records inherit the same fold number, and records from the same source are not allowed to appear in the training fold and validation fold respectively. Each cross-validation uses 4 folds as the training fold and the remaining 1 fold as the validation fold, until each fold is used as a validation fold; each cycle recalculates the preprocessing parameters and regression coefficients from the original calibration records, without reusing parameters obtained from validation folds in other cycles. The grouping list, fixed pseudo-random number seed, and data version number are saved together with the model training results, so that the same data version can repeatedly obtain the same model build set, independent validation set, and 5-fold partition.

[0037] For each 4-bit modal validity code except for the completely invalid code 0000, calibration records with the required valid modalities are selected from the model building set. Only the feature blocks corresponding to the code are retained, and a sub-model input matrix is ​​formed according to the aforementioned fixed feature order. The sub-model output matrix is ​​formed using the three anion labels of the same record. Records with additional valid modalities can be used for corresponding low-dimensional combinations by selecting the required feature blocks, but they are still grouped according to the original sample groups. For example, code 1011 indicates that the near-infrared hyperspectral is invalid according to the position order of surface texture, near-infrared hyperspectral, infrared thermal response, and shallow surface resistivity. During training, only local texture, thermal response, and resistivity connectivity features are selected, and the spectral features are not set to zero and are retained in the input matrix. A sub-model is trained only when a code has no less than 20 independent model building sample groups, and samples of the three anions exceeding the corresponding quantitative limits are distributed in all sample groups. If there are fewer than 20 sample groups or the output distribution is insufficient, a sub-model for the code is not built, and the code status is recorded as insufficient samples in the training result list.

[0038] Each code to be trained employs a nonlinear iterative partial least squares (PLLS) algorithm to fit a multi-output PLLS regression model. For the centered and scaled input and output matrices, the anion column with the largest variance is first selected from the current output residual matrix as the initial output score. The input weights are calculated using the covariance between the output score and the input residual matrix, and then normalized. The input residual matrix is ​​projected onto the normalized input weights to obtain the input score. The output weights are then calculated using the covariance between the input score and the output residual matrix, and the output score is updated using these output weights. This process repeats the calculation of input weights, input scores, output weights, and output scores until the relative change between two adjacent input scores satisfies the convergence condition. This iterative process utilizes three types of anion output information simultaneously, avoiding the training of three independent models with different data partitions for each output.

[0039] The current latent variable is determined when the relative change between two consecutive input score vectors is no greater than one billionth. If the convergence condition is not met after 500 iterations, the candidate model is marked as non-convergent and does not participate in the latent variable selection. After the current latent variable converges, the input loading and output loading are calculated, and the parts explained by the current latent variable are subtracted from the input residual matrix and output residual matrix, respectively. The next latent variable is then extracted from the updated residual matrix. After reaching the required number of candidate latent variables, the multi-output regression coefficients are calculated based on all input weights, input loadings, and output loadings. The number of candidate latent variables starts from 1, and its upper limit is the minimum of 10, the rank of the training fold input matrix, and the number of independent sample groups in the training fold minus 1, to avoid the number of latent variables exceeding the actual rank of the input matrix or approaching the number of independent sample groups.

[0040] For each candidate latent variable, a full grouped 5-fold cross-validation was performed. Each time, the input / output preprocessing parameters and regression coefficients were fitted using a 4-fold record, and the root mean square (RMS) prediction errors for chloride, sulfate, and nitrate ions were calculated using the remaining 1-fold record. When the same sample group had multiple rounds of repeated observations, the arithmetic mean of the predicted values ​​from each round was calculated first, and then the error was calculated using a single count for each sample group to prevent sample groups with more repeated imaging from receiving higher weights. The standardized RMS prediction error was obtained by dividing the RMS prediction error of each anion by the larger of the label range and the corresponding ion chromatography limit of quantitation in the model building set. The arithmetic mean of the standardized RMS prediction errors of the three anions was used as the latent variable selection criterion. After calculating the number of all candidate latent variables, the minimum number of latent variables was selected from the candidate values ​​whose difference from the minimum selection criterion did not exceed 1% of the minimum selection criterion. If all candidate models failed to converge, the sub-model for that encoding was not established.

[0041] After determining the number of latent variables, the input-output centering parameters and scaling parameters are recalculated using all model construction set samples, and the final regression intercept and regression coefficients of the code are refitted. The refitted sub-model is only sent for independent validation if the selection index corresponding to the number of selected latent variables is no greater than 0.15; otherwise, the sub-model of the code is not sent for independent validation. During refitting, the training residuals are still checked using the sample group as the statistical unit. If the direction of the repeated observation residuals of a single sample group is consistently consistent, the sample group is retained and marked as pending verification, and is not manually deleted after reviewing the independent validation results due to its large residuals. Only records that are clearly unqualified in terms of original files, equipment calibration, spatial binding, or chemical testing quality control are excluded according to pre-defined quality control conditions.

[0042] Independent validation uses the feature definitions, fixed order, mask removal, and preprocessing parameters of the final fitted candidate sub-model as input. It outputs predicted values ​​for the mass content of three anions without adjusting any model parameters using the independent validation label. Validation error is calculated before negative predicted values ​​reach zero to prevent manual error reduction during truncation. For target items with ion chromatography reference values ​​higher than the corresponding limits of quantitation, the relative prediction error is calculated by dividing the absolute value of the difference between the predicted value and the reference value by the reference value; the average relative prediction error for all valid target items should not exceed 15%. When the sum of the predicted values ​​of the three anions exceeds the total quantitation limit, the anion with the highest predicted proportion is selected as the dominant anion type, and the accuracy rate for determining the dominant anion type should not be less than 90%. For salt-free control samples where the sum of the reference contents of the three anions does not exceed the total quantitation limit, the proportion of predicted total contents not exceeding the total quantitation limit should not be less than 95%. The total quantitation limit is determined by the sum of the mass content limits of quantitation obtained by converting the three anions under the same sampling quality and extraction conditions. Independent validation passes when all three conditions are met.

[0043] After independent validation, save the training data version, sample group list of the model building set and independent validation set, 5-fold group list, feature names and fixed order, physical size of local feature windows, fixed wavelength sequence, water absorption band boundary, feature deletion mask, input and output mean and standard deviation, number of latent variables, regression intercept and regression coefficient, group cross-validation error, distribution of residuals for each anion, and independent validation results for each validated sub-model. Model training uses only quality-controlled ion chromatography detection results as supervision labels, and does not use the model's own predicted values, neighboring pixel predicted values, manual interpolation results, presumed salt damage type in low-confidence regions, or subsequent remediation status as training labels. Save the data list, exclusion record list, fixed pseudo-random number seed, feature definition file, and training log for each training session. Re-execute training when the data version, group list, fixed pseudo-random number seed, and training program version are all the same; this should result in the same sample group division, number of latent variables, and regression parameters. This forms a complete and repeatable training process, from establishing supervised labels, forming training features, grouping, fitting sub-models, selecting latent variables to independent validation.

[0044] The 900 nm to 2500 nm range simultaneously covers water absorption information near 1450 nm and 1940 nm. The influence of moisture on salt damage prediction is not fixed based on the condition of being below the average temperature of the entire surface in a single thermal image, but is characterized by the heating amplitude, cooling slope, thermal response time, and water absorption band. The model outputs effective content predictions only when the water content characteristics are within the coverage area of ​​the calibration test block; pixel grid cells outside the coverage area are marked as low confidence. Negative predicted values ​​of the three anion contents are treated as zero values, and each pixel grid cell forms three non-negative content layers for chloride ions, sulfate ions, and nitrate ions. The soluble salt content referred to in this application is the sum of the mass contents of the above three target anions, used as a quantitative indicator of salt damage; when the sum of the mass contents of the three target anions is higher than the quantification limit of ion chromatography, the anion with the highest proportion is used to generate a semantic segmentation map of salt damage type; when the sum of the contents is not higher than the quantification limit, the corresponding pixel is marked as a salt-free background type. The grouped cross-validation residuals and the Mahalanobis distance from the latent variable score space to the calibration sample distribution are mapped to confidence scores of 0 to 1, and the lower of the two is taken as the recognition confidence.

[0045] After the initial training is completed, a confidence level check is performed. In this embodiment, pixel grid cells with a confidence level below 0.80 are marked as low confidence, and this limit is predetermined by the standard test block validation set. When the area of ​​a continuous low confidence region is greater than 0.04 square meters, or when the content gradient between adjacent salt damage types is higher than the 95th quantile of the content gradient in the non-type boundary region of the same lithology validation set, one pixel grid cell far away from the existing sampling point is selected as a supplementary sampling candidate point for each continuous region. The supplementary sampling must be confirmed again by cultural relic protection professionals, and the obtained ion chromatography detection results are added to the training data and refitted. When the relative prediction error of group cross-validation is not greater than 15%, and at least 90% of the processable pixel grid cells have a Mahalanobis distance in the latent variable score space that does not exceed the 95th quantile of the calibration sample, the identification results are used for subsequent partitioning; regions that do not meet the above conditions retain an unresolved label and remain in a pending review state in the spatiotemporal evolution analysis diagram. For pixel grid cells with retained low confidence unresolved masks, the predicted values ​​of adjacent pixels are not used for filling.

[0046] Figure 2 The spatial correspondence between multimodal features and recognition outputs in a unified pixel grid is shown. Figure 2 (a) Solid lines represent spectral salt feature contour lines, dashed lines represent thermal response time contour lines, dotted lines represent stable texture edges, dotted-dash lines represent low resistivity connected channels, and hollow squares represent ion chromatography calibration points. Figure 2 (b) Maintaining the same surface coordinates, different diagonal lines are used to represent chloride-dominated, sulfate-dominated, and nitrate-dominated regions; the percentage of solid contour lines represents the sum of the mass contents of the three target anions, the dashed lines represent the 0.80 confidence level boundary, and the dotted areas represent low-confidence unresolved masks. Pixels within the unresolved mask are not filled with the predicted values ​​of adjacent pixels, nor are they spatially connected to the effective pixels on both sides of the mask.

[0047] Equivalent connected porosity was measured at sampling points using a portable single-sided nuclear magnetic resonance (NMR) probe. With the permission of cultural heritage professionals, additional measurement points were added at non-sampling locations with higher estimation uncertainty. The probe operated at frequencies from 8 MHz to 13 MHz, with an effective detection depth of approximately 5 mm. Transverse relaxation attenuation signals were acquired using a Carr–Purcell–Meiboom–Gill multi-echo sequence. For standard specimens of the same lithology, the porosity was first determined using a vacuum saturation method. Then, NMR signals were acquired at four water content levels and under the same probe pressure to establish a calibration relationship between water content, the integral value of the movable water signal in the transverse relaxation spectrum, and the porosity. The on-site water content was determined by combining the hyperspectral water absorption characteristics and infrared thermal response at the same location to determine its calibration interval. The on-site NMR signal was converted to equivalent connected porosity using the corresponding calibration relationship. The correspondence between transverse relaxation time and pore size is affected by surface relaxation rate and mineral composition; therefore, this embodiment does not directly equate a fixed relaxation time to a fixed pore size.

[0048] The two-dimensional coordinates of the on-site NMR measurement points are used as input to the Gaussian process regression, and the equivalent connectivity porosity is used as the output. A Matérn-type kernel function is used, with the smoothness parameter selected from 1.5 and 2.5 through cross-validation. The length scale and noise variance are determined by maximizing the edge likelihood. The posterior mean of each pixel grid cell is used as the estimated equivalent connectivity porosity, and the posterior standard deviation is used as the estimation uncertainty. For continuous regions where the estimation uncertainty exceeds a preset upper limit, NMR measurement points are preferentially added; if no more measurement points can be added, the partition uncertainty of that region is increased without changing the recognition confidence given separately by the image recognition model. The posterior mean and posterior standard deviation are written into a unified pixel grid to generate the equivalent connectivity porosity map and the estimation uncertainty map.

[0049] The shallow resistivity distribution map was obtained using a portable electrical impedance tomography probe. The probe was equipped with 32 flexible dry contact electrodes (4×8), and an AC current with an amplitude less than 1 mA and a frequency of 1 kHz was applied using a Winner-Schlumberger combined array. Contact impedance was checked before each measurement; data with contact impedance exceeding the equipment's calibration range were not included in the inversion. The probe moved along a uniform pixel grid, and the boundary potential response was inverted into a shallow resistivity distribution map using Gauss-Newton iteration and Tikhonov regularization, and then registered to a surface reference image based on a surface coordinate model. The current amplitude, contact pressure, and single-point dwell time were all verified on specimens of the same lithology; electrodes were not used in the pigment layer or loose surface layer. The shallow resistivity modal bit was valid when the contact impedance was within the equipment's calibration range and the inversion residual was not higher than the 95th quantile of the verified data for the same lithology; otherwise, the bit was set invalid, and the corresponding resistivity feature was not input into the fusion sub-model.

[0050] Multi-scale linear fracture features were extracted from the surface reference image and morphologically refined. In the shallow surface resistivity distribution map, low-resistivity pixels were extracted based on the lower resistivity limit of a dry reference area with the same lithology, and connected components with at least 5 four-connected members were retained. Linear fracture features and low-resistivity connected components that coincide within the same pixel grid or are adjacent to each other within one grid cell were identified as equivalent conductive connected channels. The Euclidean distance from the center of each pixel grid cell to the nearest equivalent conductive connected channel was calculated. and with As an indicator of pore connectivity. The Euclidean distance is given in millimeters. This is a positive bias term; in this embodiment, it is set to 1 mm to avoid... When the denominator is zero; It can be calibrated within 0.5 mm to 5 mm according to a uniform pixel grid scale. Low resistivity connected components that do not coincide with or are not adjacent to linear crack features are not considered as equivalent conductive connected channels.

[0051] First, the salt-free background pixels in the semantic segmentation map of salt damage types are excluded; each remaining pixel grid unit has one of the three dominant anionic salt damage types, an equivalent connectivity porosity level, and a pore connectivity level. The latter two are divided into low, medium, and high levels according to the 33rd and 67th percentiles of the lithological calibration samples, respectively, and combined with the three dominant anionic salt damage types to form... Labeling is used. Adjacent cells with the same label are merged using four-connected region growth. Regions with an area less than 0.04 square meters are only merged with adjacent regions of the same dominant anionic salt damage type, and the target region is selected according to the sum of the equivalent connectivity porosity and the standardized difference of pore connectivity; if there are no adjacent regions of the same salt damage type, they are retained as independent salt damage analysis regions.

[0052] When the quantitative sequence of salt damage is taken from the actual desalination process, a repeatable monitoring window is set for each salt damage analysis area. Two monitoring windows are set for salt damage analysis areas with an area no larger than 0.25 square meters, and one monitoring window is added for every additional 0.25 square meters. The monitoring windows avoid the tips of cracks and loose surfaces. The desalination dressing is made into 50 mm × 50 mm segments that can be peeled off individually at the monitoring windows. Only this segment is peeled off during intermediate sampling. After close-range imaging is completed, a new segment with the same formula is inserted. Full-width imaging of the salt damage analysis area is performed at the beginning and end of the entire cycle. This arrangement allows the remaining salt time series to be obtained mainly from non-destructive images, with ion chromatography only used for initial and final calibration or verification of low-confidence areas.

[0053] Imaging conditions can be replaced according to cultural relic protection restrictions. When active heating is prohibited, natural ambient thermal processes should be used. When space is insufficient, a multi-band imaging device filtered using the same training data can be used instead of a full-band hyperspectral device. When electrode contact is not permitted on the surface, continuous high-equivalent connectivity porosity regions are extracted from nuclear magnetic resonance interpolation maps and fused with surface linear fissure features to generate a porosity connectivity map, while simultaneously increasing the zoning uncertainty of the corresponding regions. Portable Raman spectroscopy or Fourier transform infrared spectroscopy can be used for salt site verification, but quantitative monitoring labels still rely on ion chromatography results. When the sample size is sufficient to support nonlinear regression, multi-output partial least squares regression can be replaced with salt-specific support vector regression, maintaining the input features, output content, and confidence verification conditions unchanged after the replacement.

[0054] Step 2 is performed independently within each salt damage analysis region. The multimodal salt damage identification model provides a quantitative salt damage map at each sampling time, while the fractional-order salt migration model describes the change of residual salt over time within the same salt damage analysis region. The full-scale quantitative salt damage map at the current time is used for persistent coherence analysis. The salt return risk map, absolute residual salt level, and spatial topological feature map represent temporal changes, content levels, and morphological connectivity, respectively. If any image feature fails to meet the corresponding identification criteria, the salt damage analysis region is not identified as a stable convergent state. Figure 5The thicker connected lines represent the rapid migration path of the fissures, the small round spots represent the microporous matrix, and the hexagons represent salt crystal particles; the dual-scale structure of the fissures and matrix allows the image time series to simultaneously contain both rapid migration and slow replenishment.

[0055] This embodiment selects the actual desalination process as the time-series image acquisition condition. Before this round of desalination, a pulp and diatomaceous earth composite dressing with a thickness of 15 mm to 25 mm was laid on the surface of the salt damage analysis area. The desalination start time is recorded as... The duration of a single round is denoted as In this embodiment Take 96 hours. Within the interval. Inner time interval set up Each sampling time In this embodiment Take 12 hours, Take 8. At each intermediate sampling time, only the monitoring window segment is peeled off, and after standing for 2 minutes, the near-infrared thermal response image and near-infrared hyperspectral image are acquired. After completion, a new dressing segment with the same formula is inserted; at the end of a single round, the full image is acquired.

[0056] At each sampling time The monitoring window image is input into the multimodal salt damage identification model. Pixels that meet the required identification confidence level are averaged according to their confidence level. Then, the area-weighted average of multiple monitoring windows in the same salt damage analysis area is taken to obtain the representative value of remaining salt content. Arranged according to sampling time. And record the initial value at the start of this round. When trace sampling is permitted, one ion chromatography point is used to verify the image prediction at the first and last sampling times; sampling is not repeated at intermediate times. If the ion chromatography detection value is higher than the limit of quantitation, the relative deviation is calculated using this detection value as the denominator; if the relative deviation exceeds 15%, the identification model is updated first, and then the residual salt time series is generated. If the detection value is not higher than the limit of quantitation, the predicted value is checked against the limit of quantitation to see if it is within the acceptable range.

[0057] A migration control equation was established using the time series of residual salinity. The representative values ​​of residual salinity were then used. As a state variable, the process of salt migration from shallow pores to the dressing under the action of the desalination dressing, and its continued replenishment from deeper pores, is taken as the evolutionary object. The migration control equation is:

[0058] in, Indicates self The initial fractional time derivative of the Caputo type; The order of memory takes a value greater than 0 and less than 1. The smaller the value, the longer the influence of historical conditions on current changes lasts; The apparent migration coefficient is expressed in negative time units. Power of; This is the boundary release term, with units of remaining salt content divided by time. The power of 1 indicates the salt supply from deeper pores to the monitored shallow layer.

[0059]

[0060] in, It is a gamma function; The remaining salt content is represented by the integral variable. The first derivative at that point; In For the current moment, For integration variables, For memory order. When When, the derivative approaches the ordinary first-order time derivative; when When the value is between 0 and 1, historical terms participate in current changes according to power-law weights. The combination of faster migration in fractures and slower replenishment in the microporous matrix can produce a long-tailed release, therefore... Used to describe memory strength. The smaller the size, the more pronounced the long tail; It represents the combined migration coefficient under the current dressing and porosity conditions, and its value varies with the water content, dressing absorption capacity, and equivalent interconnected porosity. The physical meaning and units remain consistent across all rounds.

[0061] Boundary release item It can be implemented in either a constant form or a humidity modulation form. When environmental changes are minor, let... Equal to a constant When environmental changes are significant, adopt .in, The time series is obtained by normalizing the infrared thermal response and the on-site relative humidity, with values ​​ranging from 0 to 1. This is a dimensionless humidity sensitivity coefficient. Insufficient environmental data may cause this. In this embodiment, a constant form is used.

[0062] Parameter inversion uses the residual salinity time series as the fitting target, and the undetermined parameter vector is denoted as . Given Then, the model values ​​at each sampling time are calculated using the same difference scheme as in the extrapolation stage. And construct the residual vector The summation range covers all sampling times, and the objective function is the sum of squared residuals:

[0063] The parameters are updated iteratively using the Levenberg-Marquardt nonlinear least squares method:

[0064] in, For the first The parameter vector for the next iteration; For the residual vector in Jacobian matrix for the parameter vector; For the first Damping coefficient of the next iteration; This indicates that a diagonal matrix is ​​formed by taking the diagonal elements of the matrix. The transpose superscript in the formula indicates matrix transpose, and -1 power indicates matrix inversion. Since the Jacobian matrix is ​​differentiated with respect to the parameters using the residuals, a negative sign is used before the update term. The perturbation of the numerical differentiation is taken as the larger of 0.001 times the absolute value of the parameter and the preset minimum scale, to avoid the perturbation being zero when the parameter is zero. The damping coefficient is multiplied by 0.5 when the objective function decreases, and by 2 otherwise.

[0065] Initial value of iteration , and Numerical time is in hours; the initial value of the apparent migration coefficient (0.1) is calculated as a negative power of the hourly memory order, and the initial value of the boundary release term (0.01) is also calculated as a negative power of the hourly memory order. Execution is performed after each iteration. , and Projective constraints. The relative change of the objective function between two adjacent orders is less than... The iteration may terminate after 100 iterations. In the numerical verification example, the initial value was normalized to 1.000, and the eight subsequent values ​​obtained at 12-hour intervals were 0.7761, 0.6823, 0.6220, 0.5777, 0.5430, 0.5146, 0.4908, and 0.4703, respectively. Using unrounded sequence inversion, the result is... , The boundary release term inversion value is approximately 0.005, consistent with the three parameters used to generate the sequence.

[0066] After completing the parameter inversion, , and Substitute back the migration control equations to predict the set extrapolation time period. Internal changes. In this embodiment... Take 168 hours. Prediction from... Start button Discretized The number of steps satisfies After removing the desalination dressing, the dressing conditions should be... Replace with natural evaporation conditions The natural volatility coefficient was determined by a dressing removal test on specimens of the same lithology; this embodiment uses... This is for illustrative purposes only and is not a fixed percentage. It is retained when environmental conditions do not change significantly. When environmental changes are significant, the boundary release terms are updated according to the thermal response and relative humidity sequences.

[0067] Discrete Caputo-type fractional time derivatives using the Greenwald-Retnikov difference scheme:

[0068] in, negative of time step Power of; The Greenwald-Retnikov weight coefficients are derived from... and recursive formula generate; For summation indicators; This is the current time step; This represents the residual salt content at the current time step. The weighting coefficients decay algebraically with historical distance. With the boundary release term constant and the natural volatility coefficient remaining unchanged, each time step is recursively calculated using the following formula:

[0069] Depend on according to Calculate the salinity flux per unit time; a positive value indicates an increase in residual salt in the superficial layer after dressing removal. Connect the salinity fluxes at each step in chronological order to form a predicted salinity flux curve, and record the maximum value as [value missing]. The unit of salt return flux is the remaining salt content divided by time. Using the aforementioned unrounded numerical sequence, taking the natural volatility coefficient of 0.008, the boundary release term of 0.005, and retaining all historical terms, the first three extrapolated values ​​are 0.576370, 0.622389, and 0.652307, with a maximum salt return flux of 0.008839 per hour. Substituting these values ​​back into the recursive formula, the differences on both sides are less than one part per billion. The predicted maximum salt return flux for each salt damage analysis area is backfilled into a unified pixel grid according to the area boundary. Each pixel within the area retains its original identification confidence level, while low-confidence pixels remain as unresolved masks, thus generating a salt return risk map.

[0070] Figure 3 The fitting and extrapolation results of the above numerical verification examples are shown. Figure 3In (a), the nine hollow circles from 0 hours to 96 hours represent the normalized residual salt obtained from the salt damage quantitative map sequence, and the solid line is the fitting curve of the fractional salt migration model; after removing the dressing at 96 hours, the dashed line extrapolates from 0.470306 to 0.764089 at 264 hours, with the first three extrapolated values ​​being 0.576370, 0.622389 and 0.652307. Figure 3 In (b), the first positive salt return flux 12 hours after dressing removal is 0.008839 per hour. According to the rule of taking a fixed reference salt return flux of 5% for the salt return stability threshold in this embodiment, the salt return stability threshold is 0.000442 per hour. The predicted salt return flux 156 hours after removal is 0.000425 per hour, which is lower than this threshold for the first time. The predicted value at 168 hours is 0.000381 per hour. Since the maximum predicted salt return flux within the extrapolated period is still set at 0.008839 per hour, this example does not meet the salt return flux convergence condition. Figure 3 Each point in the equation is obtained by recursive calculation that retains all historical items.

[0071] Fractional-order difference calls up all historical states. The storage required to store all historical values ​​is proportional to the number of time steps, and the total computational cost of stepwise extrapolation is of the same order as the square of the number of time steps. For long-term extrapolation, short-memory truncation can be used, with the truncation length denoted as . Truncating is only allowed when the weighted sum of omitted historical items does not exceed 1% of the current item, and the truncation length is determined round by round according to this error condition. This embodiment uses a 12-hour interval and extrapolates 14 steps, retaining all history during calculation. L1 implicit differencing can also be used to improve the numerical stability over long periods.

[0072] After extrapolating the salt return flux, spatial topological features are extracted from the full-scale salt damage quantitative map at the current time. The salt damage quantitative map uses the unified pixel grid from step 1, with each pixel grid unit serving as a pixel cell. Its scalar value is the soluble salt content at that location after confidence verification. Low-confidence pixels are first supplemented with detection; if no valid value can still be obtained, they are used as unresolved masks and are not directly filled with neighboring low values. Connected regions participating in persistent cohomology calculations must not cross unresolved masks.

[0073] The pixel cell is used as the 2D cell of the cube complex, the shared edges of adjacent pixel cells and the edges of the array's outer edge are used as 1D edges, and the corner points of the pixel cell are used as 0D vertices. The filter value of a low-dimensional cell is the maximum value of the filter values ​​of its adjacent high-dimensional cells, ensuring that the boundary cell enters the complex no later than its own cell. Figure 6In the diagram, (a) shows the quantitative map of salt damage; (b) shows two isolated 0-dimensional connected components appearing at a threshold of 0.80; (c) shows the connected region and 1-dimensional independent loop formed at a threshold of 0.50; and (d) shows the expanded upper level set at a threshold of 0.25, where the central low-value cell has not yet entered, therefore the loop still exists. When the threshold is further reduced to 0.10, all cells enter, and the loop disappears; this termination state is not shown in the diagram. Figure 6 It is drawn repeatedly in the middle.

[0074] Let the maximum value in the current salt damage analysis area's quantitative salt damage map be... The minimum value is The initial threshold is and take The current threshold is adjusted by step size. Decrease step by step, denoted as until Below This embodiment takes equal The step size sensitivity was verified using a dynamic range of 0.005 to 0.02 times. The quantitative resolution of the model was set to the 95th percentile of the absolute prediction error of the grouped cross-validation. When the dynamic range of the current image was not greater than this quantitative resolution, a zero-step loop was not performed; instead, all effective pixel cells were included in the complex at once. In this case, the finite barcode lifetime was recorded as 0, and the absolute residual salt content was still determined separately by the 95th percentile condition. For each Scalar value not less than The pixel cells and their boundary units constitute the current upper level set.

[0075] As the threshold decreases progressively, the upper level set monotonically expands. Cells with high residual salt content enter first, and subsequent cells may establish new connected components, merge existing connected components, form loops, or fill loops. The persistent homology calculation of the cube complex records these topological changes.

[0076] Zero-dimensional connected components can be tracked using a disjoint-set data structure. When a newly entering pixel cell is not adjacent to an already entered cell, a new connected component is established, and its birth threshold is recorded. When a newly entering cell bridges two connected components, the connected component with the higher birth threshold is retained, and the death threshold of the other connected component is recorded as the current threshold. The lifetime of each finite barcode is equal to the difference between the birth threshold and the death threshold. The essential connected component throughout the entire filtering process is not used for speckle convergence determination; its corresponding absolute salt level is constrained by the 95th percentile condition.

[0077] One-dimensional independent loops are tracked using a cubic complex boundary matrix reduction. Elements are arranged from highest to lowest filter value; when filter values ​​are the same, 0-dimensional vertices precede 1-dimensional edges, and 1-dimensional edges precede 2-dimensional cells, ensuring that lower-dimensional boundary elements enter the matrix before higher-dimensional cells bounded by them. A boundary matrix is ​​established over a binary finite field, and reduction is performed column-by-column. After reduction, new one-dimensional loops corresponding to the lowest non-zero position in the previous column are created. When the reduced column of a subsequent 2-dimensional cell is paired with this loop, the corresponding loop disappears. The matrix columns are... Figure 6 Consistent order of expansion of the upper level set.

[0078] The aforementioned persistent cohomology calculation can be implemented using GUDHI, Cubical Ripser, or a program with equivalent boundary matrix reduction capabilities. The program inputs a masked two-dimensional salt damage quantitative array, upper level set directions, and pixel grid adjacency relationships. The output is 0D and 1D barcode endpoint pairs. When downsampling large-area salt damage analysis regions, the scalar value of the downsampling window is taken as the maximum soluble salt content within the window. A conservative union rule is used for the unresolved mask to prevent downsampling from eliminating local high-salt spots or low-confidence areas. To map the barcode back to a unified pixel grid, the 0D calculation synchronously saves the member pixel set of the connected components before each merging; the 1D calculation saves the reduced column indices paired with the finite barcodes and records the edge indices required to backtrack the representative loops along the column operations. The 0D member pixel set and the 1D representative loop edge set are rasterized and written into the spatial topological feature map. The barcode endpoints and corresponding pixel support sets are associated using the same numbering. The 0-dimensional and 1-dimensional barcode endpoint pairs output are used to generate 0-dimensional and 1-dimensional permanent barcodes, respectively.

[0079] In step 3, for each salt damage analysis area, region-level fusion features are extracted from each image: the 95th percentile of soluble salt content is extracted from the salt damage quantification map; the dominant anion salt damage type is statistically analyzed from the salt damage type semantic segmentation map, and the corresponding upper limit of residual soluble salt is selected accordingly; the proportion of low-confidence unresolved mask area is statistically analyzed from the identification confidence map; the maximum predicted salt return flux is extracted from the salt return risk map; and the longest lifetime of 0-dimensional finite barcodes and the lifetime of 1-dimensional durable barcodes are extracted from the spatial topology feature map. The evolution state classifier reads the above region-level fusion features according to a fixed priority and outputs the active salt migration state, connected residual state, discrete residual state, stable convergence state, or state pending verification.

[0080] The evolutionary state identification uses four convergence criteria. First, the extrapolation time period is set. Maximum salt return flux within Below the salt return stability threshold Second, the 95th percentile of soluble salt content in the current quantitative salt damage map is lower than the upper limit of residual soluble salt determined by wet-dry cycle tests for this stone mineral type and the dominant anionic salt damage type combination; in this embodiment, the sandstone mixed salt test block uses a mass content of 0.20% as the upper limit for testing. Third, the entire lifetime of the 1-permanent barcode is lower than the connectivity convergence threshold. In this embodiment, the threshold is fixed at 10% of the dynamic range of the first-round image. Fourth, the longest lifetime in a 0-dimensional finite barcode is lower than the speckle convergence threshold. In this embodiment, the threshold is fixed at 20% of the dynamic range of the first-round image. When the proportion of the low-confidence unresolved mask area reaches 10%, the pending verification state is output first; when it does not reach 10%, if the first condition is not met, the salt migration active state is output; if the first condition is met but the third condition is not met, the connected residual state is output; if the first and third conditions are met but the second or fourth condition is not met, the discrete residual state is output; if all four conditions are met, the stable converged state is output. When there is no 1D barcode or 0D finite barcode, the corresponding maximum lifetime is counted as 0. Essentially connected components do not participate in the fourth comparison.

[0081] Salt return stability threshold Connectivity convergence threshold and residual speckle convergence threshold The state is predetermined by the image sequence and ion chromatography verification results of the same lithological specimens, and the dynamic range of subsequent time steps is not widened as the dynamic range of the current image decreases. The spatiotemporal evolution analysis diagram of the previous time step serves as the state prior; the state label is only updated when the same new state is obtained for two consecutive time steps. When a region splits or merges, the overlap area between the new region and the original region is calculated, and the state of the region with the largest overlap is used as the inherited state. Then, the new region is re-identified based on the fusion characteristics of the new region. When the proportion of the low-confidence unresolved mask area reaches 10%, the state to be verified covers the inherited state. After obtaining a new multimodal image set at the next time step, steps 1 and 2 are executed again to update the boundary of the salt damage analysis region and the regional fusion characteristics. Then, the spatiotemporal evolution analysis diagram is updated according to the above state prior and continuous confirmation rules. After processing at each time step, the image analysis terminal outputs a modal validity coding map, a salt damage quantitative map, a salt damage type semantic segmentation map, a recognition confidence map, a pore connectivity map, a salt return risk map, a spatial topological feature map, and a spatiotemporal evolution analysis map. Each salt damage analysis region in the spatiotemporal evolution analysis map has a state boundary and retains the recognition confidence of the corresponding pixel. Each image is associated and saved with the acquisition time, surface coordinate transformation relationship, feature fusion sub-model version, and source of supervision labels. The multimodal salt damage recognition model is updated only using newly added supervision labels confirmed by ion chromatography.

[0082] In another embodiment, during the restoration of stone cultural relics, a spatiotemporal evolution analysis diagram is used. The on-site operation terminal uses the stable convergence state as the image criterion for entering the reinforcement review stage. The active salt migration state, the connected residual state, and the discrete residual state are used as references for adjusting the desalination sequence, maintaining desalination, and supplementing local detection, respectively. The state to be reviewed serves as a prompt for supplementary imaging or sampling. Each restoration operation is still independently determined by cultural relic conservation professionals in conjunction with material compatibility tests.

[0083] The reinforcement materials were determined jointly by pre-defined material compatibility rules and test area results. The material compatibility rules used the stone mineral type as the primary index, followed by the dominant anionic salt damage type and the construction moisture content, recording the candidate material system, solvent, allowable pressure range, single application dosage, and curing conditions. The stone mineral type was obtained from the petrographic identification archive; if the archive could not cover the area to be repaired or the surface had undergone replacement repairs, mineral identification was supplemented at permissible locations on the edges. Any candidate material was first verified on specimens of the same lithology and in inconspicuous test areas; the chemical reaction was not directly assumed to have eliminated residual salt based on the salt damage type label.

[0084] Mineral compatibility takes precedence over salt type labeling. Siliceous sandstones should be prioritized for selection from tetraethyl orthosilicate or validated silica sol systems, while carbonate stones should be prioritized from nano-lime or phosphate mineralization systems; specific systems still require color, capillary water absorption, and water vapor permeability tests. When chloride or nitrate ions are dominant, reduce the amount of water introduced during construction and confirm that the hygroscopic salt content is below the upper limit of residual soluble salts; when sulfate ions are dominant, first confirm that sulfates no longer form connected high-salt areas, and then assess whether calcium-containing aqueous phase treatment will re-dissolve salts. Figure 4 The ammonium oxalate conversion layer shown is a surface treatment example that can be used after the carbonate stone has been verified in the test area. It is only used to form a surface calcium oxalate layer and is not used to directly convert sulfate ions into calcium oxalate.

[0085] Low-pressure permeability reinforcement (see) Figure 4 In two adjacent desalination zones, the left zone enters the reinforcement intervention window, while the right zone continues to be covered with desalination dressing. A reversible isolation layer is set at the zone boundary. Before construction, the surface water content is verified using infrared thermal response and nuclear magnetic resonance. The liquid-absorbing capillary medium is laid only after the upper limit of construction water content determined by material testing is reached. Low-pressure infusion bottles are used to supply the capillary medium through conduits, with the pressure gradually adjusted from 0.02 MPa, not exceeding 0.08 MPa, and must not exceed the pressure confirmed by the same lithology test block and test area to be free of cracks and efflorescence. Figure 4 The 8 mm to 15 mm values ​​indicated are the apparent penetration depths measured by tracer imaging on test blocks of the same lithology in this embodiment, and are not fixed values ​​for all types of stone. The number of applications and the amount applied per application are determined by the absorption curve of the test area; the next application can only be performed after the previous application of material has stopped absorbing and there is no residual film on the surface.

[0086] The reinforcement effect is first verified in the test area and then extended to the corresponding desalination zone. Verification items include ultrasonic wave propagation speed, capillary water absorption coefficient, water vapor permeability, and visible color difference. Conventional rebound hammers are not used to repeatedly impact fragile artifact surfaces. The acceptance criteria for the test area in this embodiment are: a relative increase in ultrasonic wave propagation speed of no less than 15%, a decrease in capillary water absorption coefficient of 30% to 60%, a water vapor permeability retention rate of no less than 70%, and a CIEDE2000 color difference value of no more than 3. If any item fails to meet the criteria, the material compatibility, penetration uniformity, and surface moisture content are first assessed; increasing the number of applications is not used as a substitute for root cause analysis. The material system is discontinued if abnormal color, a hardened surface shell, or excessively low permeability occurs.

[0087] For desalination zones that have not entered the reinforcement intervention window, the maximum predicted salt return flux will be used. Relative to a fixed reference salt return flux The proportions are used to divide the grades. This is during peak hours; This is during the mid-peak period; This is during off-peak hours; It will no longer engage in gear adjustment. The first positive salt return flux exceeding the quantitative limit of return flux was recorded under natural volatilization conditions after the removal of the first round of desalination dressing, and remained unchanged in subsequent rounds. The quantitative limit of return flux was taken as the 95th percentile of the absolute value of the extrapolated flux obtained from blank specimens of the same lithology under the same sampling sequence. If no positive salt return flux exceeding this quantitative limit appeared in the first round of extrapolation, the minimum positive reference value pre-calibrated for specimens of the same lithology was used to ensure that the fixed reference salt return flux was always greater than zero. To determine the stable salt return threshold, this embodiment takes... .fixed To avoid the gear boundaries changing synchronously with the prediction results due to the use of new reference values ​​in each round.

[0088] when When the peak value is reached, the dressing replacement interval is 12 hours and the single-cycle dwell time is 96 hours; when the peak value is reached, the intervals are 24 hours and 72 hours respectively; and when the peak value is reached, the intervals are 48 hours and 48 hours respectively. When the salt return flux is high, the dressing reaches its adsorption capacity quickly, thus shortening the replacement interval; when deeper pores are still being replenished, the total single-cycle duration is extended to achieve a complete decay phase. When the salt return flux is low, the number of replacements and the continuous wetting time are reduced to avoid further extraction of soluble components related to matrix cementation. Before each replacement, the image content and identification confidence level of the monitoring window are checked; if the content is abnormally high or the identification confidence level is insufficient, the adjustment is paused and a recheck is performed first.

[0089] After each round of desalination, infrared thermal response images and near-infrared hyperspectral images are reacquired; shallow surface resistivity distribution maps are acquired when the dressing is removed, the surface reaches uniform water content, and there is no liquid phase construction in adjacent zones. Ion chromatography is not repeated for all sampling points in each round, but only in low-confidence areas, type boundary migration areas, or termination verification points. Nuclear magnetic resonance measurements are generally performed once every 3 rounds, or earlier when resistivity connectivity and thermal response change significantly simultaneously. Updated dominant anion salt damage type, equivalent connectivity porosity level, and pore connectivity level are used to verify zone boundaries; small area merging is still constrained by the condition of consistent dominant anion salt damage type.

[0090] The fractional-order salt migration model is updated after a new round of data arrives. When parameter changes are gradual and image partition boundaries remain unchanged, the previous round's parameters are used as initial values ​​for iteration. When partition boundaries change, equivalent connectivity porosity changes by more than 30%, or the previous round's fitting fails to converge, the preset initial values ​​are reapplied. The multimodal salt damage identification model is updated only using new samples confirmed by ion chromatography and cannot use its own low-confidence predictions as supervisory labels. Persistent cohomology calculation is performed independently for each round of salt damage quantification map. The barcode threshold is referenced to the fixed dynamic range before the first round of desalination and does not automatically widen as the current image's dynamic range shrinks.

[0091] For the timing relationship of multiple desalination zones, see [link to relevant documentation]. Figure 7 Zones A, B, C, and D reached the reinforcement intervention window on days 4, 9, 14, and 25, respectively, corresponding to different rounds of testing, desalination, and assessment. When reinforcement begins in one zone, adjacent zones can continue desalination; if a zone needs additional desalination rounds, it does not affect the commencement of maintenance in zones that have already met the standards. This arrangement avoids premature reinforcement in zones whose salt content has not yet reached the joint criteria, and also avoids unnecessary wet compresses being applied to zones that have already met the standards.

[0092] When adjacent zones are in different states, a polyethylene film and a reversible isolation liner are laid along the zone boundary. The isolation liner is first covered with acid-free fiber paper or a proven reversible cellulose material, and neutral silicone is not used directly on the artifact surface for fixation. The isolation layer is removed after solvent migration on the reinforced side has stopped and the prescribed curing has been completed. The shallow resistivity acquisition is not performed simultaneously with the wet dressing or reinforcement liquid supply of adjacent zones; if it cannot be staggered, an unresolved mask is set for the affected pixels, and the resistivity results of this acquisition are not used for zone updates.

[0093] The on-site operation terminal maintains five states for each desalination zone: desalination in progress, awaiting verification, reinforcement pending, reinforcement and curing in progress, and reinforcement completed. Once the desalination determination meets the joint criteria, it first enters the awaiting verification state, confirmed by independent image verification and necessary ion chromatography points; after passing verification, it enters the reinforcement pending state. Images are acquired again after reinforcement and curing are completed; if new high-salt areas are found, an abnormal verification mark is set, and the status is returned to awaiting verification. Wet application is not directly repeated on the existing reinforcement layer; instead, cultural relic protection professionals determine localized treatment based on the material penetration depth and the location of new salt damage. The restoration process terminates when all zones are in the reinforcement completed state and pass the overall verification.

[0094] Low-pressure permeation can be replaced with vacuum-assisted permeation after verification with lithological test blocks. The negative pressure amplitude, sealing material, and duration should all be limited to avoid causing surface displacement. Material adaptation rules can include construction temperature, surface moisture content, and equivalent interconnected porosity as indexes, but new indexes must have corresponding experimental data. Fractional-order models can use L1 implicit difference or predictive correction schemes, and persistent cohomology can use a cube complex procedure equivalent to boundary matrix reduction; after algorithm replacement, the same input, output, and joint decision meanings must be maintained. The isolation liner can be made of acid-free fiber paper, methylcellulose reversible membrane, or other reversible materials confirmed by cultural relic protection tests.

[0095] This embodiment establishes a continuous image processing flow encompassing multimodal image acquisition and registration, pixel-level salt damage identification, identification confidence verification, salt return risk map generation, spatial topological feature extraction, and evolution state identification. The output of the multimodal salt damage identification model directly constitutes the input for time series analysis and persistent cohomology analysis; the image at the next time step is used to update the region boundary and state map. Figure 4 and Figure 7 This only shows one field application of the spatiotemporal evolution analysis results and does not change the technical nature of this invention, which takes image data as input and outputs recognition maps and state maps. Figures 1 to 3 Spatial or numerical relationships are given for cross-modal registration and modal validity coding, multimodal fusion recognition output, and fractional-order temporal extrapolation.

[0096] The above embodiments are used to illustrate the technical solution of the present invention. Those skilled in the art, while maintaining the data relationships between multimodal image registration, pixel-level fusion recognition, back-salt risk map generation, spatial topological feature extraction, and spatiotemporal evolution state recognition, can make equivalent substitutions for the imaging device, grid size, regression method, and persistent coherence procedure; the equivalent substitution scheme should still satisfy the image input, processing procedure, and image output as defined in the claims.

Claims

1. A method for multimodal image recognition and spatiotemporal evolution analysis of salt damage to stone cultural relics, characterized in that, Includes the following steps: Step 1: Obtain the surface reference image, infrared thermal response image sequence, near-infrared hyperspectral image cube, and shallow surface resistivity distribution map of the stone artifact surface. Register the infrared thermal response image sequence, the near-infrared hyperspectral image cube, and the shallow surface resistivity distribution map to the surface coordinate system of the surface reference image and resample them to a unified pixel grid to form a multimodal image group. Based on the local registration residuals, surface texture imaging quality, spectral imaging quality, thermal response imaging quality, and resistivity imaging quality of each pixel grid unit, a modal validity coding map is generated. Using the ion chromatography detection results of the sampling points as supervision labels, the spectral reflectance features, thermal response features, local texture features, and resistivity connectivity features of each pixel grid unit are extracted. Using a multimodal salt damage identification model, a feature fusion sub-model corresponding to the effective modal combination is selected according to the modal validity coding map to generate a salt damage quantitative map, a salt damage type semantic segmentation map, and an identification confidence map. Pixels without matching feature fusion sub-models or whose identification confidence does not meet the preset requirements are written into a low-confidence unresolved mask, and a porosity connectivity map is generated based on the local texture features and the resistivity connectivity features. Step 2: Divide the salt damage analysis region according to the spatial continuity relationship between the semantic segmentation map of the salt damage type and the porosity connectivity map; repeat Step 1 for the multimodal image group at different times to obtain the salt damage quantitative map sequence of each salt damage analysis region; extract the remaining salt time series from the salt damage quantitative map sequence and establish a fractional-order salt migration model; map the predicted salt return flux to the unified pixel grid to generate a salt return risk map; construct a cubic complex with the salt damage quantitative map at the current time and perform upper level set filtering to obtain 0-persistence barcodes and 1-persistence barcodes; map the connected components and independent loops corresponding to the barcodes to the unified pixel grid to generate a spatial topological feature map; Step 3: For each salt damage analysis area, extract regional-level fusion features from the salt damage quantitative map, the salt damage type semantic segmentation map, the identification confidence map, the salt return risk map, and the spatial topological feature map. Based on the regional-level fusion features, identify the salt damage analysis area as an active salt migration state, a connected residual state, a discrete residual state, a stable convergence state, or a state awaiting verification. Backfill the identification state of each salt damage analysis area into the unified pixel grid to generate a spatiotemporal evolution analysis map with state boundaries and identification confidence. Update the spatiotemporal evolution analysis map using the multimodal image group at the next time step.

2. The method as described in claim 1, characterized in that, In step 1, a curved coordinate model of the stone artifact surface is established using three-dimensional laser scanning. Band images with a signal-to-noise ratio that meet preset conditions are selected from the near-infrared hyperspectral image cube as the surface reference image. The infrared thermal response image sequence is continuously acquired before, during, and after controlled thermal excitation. Dark current correction and standard reflector correction are performed on the near-infrared hyperspectral image cube. Geometric coarse registration is completed based on surface control points, and cross-modal fine registration is completed based on mutual information and stable texture edges. Non-reference modes with registration residuals exceeding half the side length of the unified pixel grid cell are marked as invalid in the corresponding pixel grid cell.

3. The method as described in claim 2, characterized in that, The multimodal salt damage identification model includes multiple feature fusion sub-models corresponding to different effective modal combinations. Each feature fusion sub-model is a multi-output partial least squares regression model. The spectral reflectance features include the first derivative spectrum after standard normal variable transformation and water absorption band features. The thermal response features include thermal excitation heating amplitude, cooling slope, and thermal response time. The local texture features include local gradient and local variance. The resistivity connectivity features include local resistivity and its distance to low resistivity connectivity channels. The feature fusion sub-model with the same modal validity encoding as the current pixel grid cell and which has passed independent verification is selected. The predicted values ​​of chloride ion content, sulfate ion content, and nitrate ion content, as well as the identification confidence level, are output. When the sum of the three anion contents is higher than the quantification limit, the one with the highest proportion is determined as the dominant anion salt damage type of the current pixel grid cell. When the sum of the contents is not higher than the quantification limit, the current pixel grid cell is determined as a salt-free background type. If no matching and independently validated feature fusion sub-model exists, write the current pixel grid cell into the low-confidence pending mask.

4. The method as described in claim 3, characterized in that, The monitoring labels are provided by standard stone test blocks of the same lithology and the sampling points. When the area of ​​a continuous low-confidence region exceeds a preset area or the content gradient at the boundary of a salt damage type exceeds a preset gradient, supplementary sampling points are selected at the continuous low-confidence region or the boundary of the salt damage type, and the multimodal salt damage identification model is updated using the supplementary ion chromatography detection results. For pixel grid units that have not obtained effective monitoring labels and whose identification confidence does not meet the requirements, a low-confidence unresolved mask is retained, and the predicted values ​​of adjacent pixels are not used for filling.

5. The method as described in claim 4, characterized in that, A portable nuclear magnetic resonance single-sided probe was used to collect transverse relaxation signals at sampling points. After calibration with the vacuum saturated porosity of standard stone blocks of the same lithology, the equivalent connected porosity was obtained. Gaussian process regression was used to generate an equivalent connected porosity map and an estimation uncertainty map. Linear fracture features were extracted from the surface reference image, and low resistivity connected channels were extracted from the shallow surface resistivity distribution map. The overlapping and adjacent areas of the two in a unified pixel grid were determined as equivalent conductive connected channels. The porosity map is generated based on the distance from each pixel grid unit to the equivalent conductive connected channel, and the salt damage analysis region is divided using a region growing algorithm constrained by the dominant anion salt damage type and the equivalent connected porosity.

6. The method as described in claim 1, characterized in that, In step 2, repeatable monitoring windows are set up in each salt damage analysis area. Infrared thermal response images and near-infrared hyperspectral images of the monitoring windows are acquired at multiple sampling times. The representative values ​​of residual salt at each sampling time are obtained using the multimodal salt damage identification model and the residual salt time series is formed. Using the residual salt time series as the fitting target, the memory order, apparent migration coefficient and boundary release term of the Caputo fractional salt migration model are inverted using the Levenburg-Marquardt nonlinear least squares method. The predicted salt return flux within the set extrapolation period is obtained using the Greenwald-Retnikov difference scheme. The predicted salt return flux is then backfilled into the salt return risk map according to the salt damage analysis area boundary and identification confidence level.

7. The method as described in claim 6, characterized in that, Each pixel grid cell of the salt damage quantitative map is taken as a 2D cell of a cubic complex, the shared edge and outer boundary of adjacent pixel grid cells are taken as 1D edges, and the corner points of pixel grid cells are taken as 0D vertices. The cubic complex is filtered by the upper level set according to the remaining salt content from high to low. The boundary matrix reduction based on the binary finite field is used to track the birth and death of 0D connected components and 1D independent loops respectively, generating the 0-persistent barcode and the 1-persistent barcode, and the pixel position corresponding to the finite barcode is written into the spatial topological feature map.

8. The method as described in claim 7, characterized in that, The determination order of the identification status is as follows: when the area proportion of the low-confidence unresolved mask in the salt damage analysis area reaches a preset proportion, it is determined to be the state to be reviewed; if it does not reach the preset proportion, if the maximum predicted salt return flux in the extrapolated time period is not lower than the preset salt return stability threshold, it is determined to be the active salt migration state; if the maximum predicted salt return flux is lower than the preset salt return stability threshold and there is a barcode in the 1-permanent barcode with a lifetime not lower than the preset connectivity convergence threshold, it is determined to be the connected residual state; if the maximum predicted salt return flux is lower than the preset salt return stability threshold, the lifetime of the 1-permanent barcode is lower than the preset connectivity convergence threshold, and the 95th percentile of the soluble salt content in the salt damage quantitative map is not lower than the preset upper limit of residual soluble salt or the longest lifetime of the 0-dimensional finite barcode is not lower than the preset residual spot convergence threshold, it is determined to be the discrete residual state; the remaining salt damage analysis areas are determined to be the stable convergence state.

9. The method as described in claim 8, characterized in that, In step 3, the spatiotemporal evolution analysis diagram of the previous moment is used as the state prior of the next moment; when the identification state of the same salt damage analysis area changes, the spatiotemporal evolution analysis diagram is updated only after the same new identification state is obtained for two consecutive moments; when the salt damage analysis area splits or merges, the original identification state is inherited according to the maximum overlap area between the areas and the regional fusion feature is recalculated. When the area ratio of the low-confidence unresolved mask reaches the preset ratio, the inherited identification state is overwritten by the pending verification state.

10. The method as described in claim 9, characterized in that, The output includes the modal validity coding map, salt damage quantitative map, salt damage type semantic segmentation map, identification confidence map, pore connectivity map, salt return risk map, spatial topological feature map, and spatiotemporal evolution analysis map; the model is updated only using the supervision labels.

Images