A method and system for identifying salinization permafrost railway by remote sensing

Abstract

This invention discloses a remote sensing identification method and system for saline-alkali permafrost along railways, relating to the field of engineering construction. The method includes: acquiring various data within a pre-defined buffer zone centered on the railway line, and constructing a multi-source data cube for the railway corridor through coordinate registration, orthorectification, and spatial resampling. Subsequently, remote sensing sensitive factors such as spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation factor, InSAR linear deformation rate, and seasonal deformation amplitude are extracted, and a comprehensive index for saline-alkali permafrost is constructed and corrected based on the heat-salt coupling relationship. The comprehensive index and sensitive factors are input into a multi-task deep learning model, outputting the probability of permafrost presence, salinity level, and engineering risk level. Finally, a risk profile is constructed along the railway mileage direction, and risk zoning identification is completed, achieving visualization and refined assessment of permafrost risks along the railway line.

Description

Technical Field

[0001] This specification relates to the field of engineering construction; more specifically, this application relates to a remote sensing identification method and system for railways in saline-alkali permafrost. Background Technology

[0002] In railway construction in high-altitude and cold regions, the thermal stability of permafrost directly affects the subgrade's settlement, cracking, frost heave, and thaw settlement. However, against the backdrop of recent global warming, the thermal condition of permafrost regions has shown a continuous deterioration trend, with the active layer deepening and ice-bearing structures being destroyed, making frost heave-thaw settlement behavior more frequent and severe. In high-altitude and saline regions, salt migration is coupled with the freeze-thaw cycle, leading to an alternating effect of "salt softening—freezing expansion—thaw settlement and collapse," exacerbating the risk of uneven settlement in railway projects. Traditional permafrost identification methods mainly rely on limited ground exploration, pore water and salt analysis, ground temperature observation, or single remote sensing data sources for judgment. These methods often suffer from poor spatial continuity, high cost, slow response, and difficulty in capturing large-scale deformation dynamics. Furthermore, existing remote sensing evaluation methods primarily rely on single-factor discrimination based on spectral, temperature, or radar scattering data, lacking a joint characterization of four key processes: salinity, thermal state, ice content, and deformation. They also lack the ability to reveal the true dynamic behavior of permafrost using deformation time series. Furthermore, traditional methods lack the ability to directly project the identification results onto the railway mileage direction for engineering expression, making it difficult to support actual railway route selection, structural type comparison, and operation and maintenance safety early warning.

[0003] Therefore, there is an urgent need for a remote sensing intelligent identification method that can integrate multi-source remote sensing data, combine the thermal-salt coupling mechanism of permafrost, and generate risk profiles that can be directly used in engineering along railway mileage directions, so as to achieve high-precision, engineering-oriented, and dynamic identification of saline permafrost. Summary of the Invention

[0004] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. This summary section is not intended to limit the key and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0005] Firstly, this application proposes a remote sensing identification method for railways in saline-alkali permafrost, including: Acquire multi-source remote sensing and monitoring data within a pre-defined buffer zone centered on the railway line. The multi-source remote sensing and monitoring data includes multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, InSAR deformation time series data, digital elevation model, ground temperature monitoring and borehole salinity profile data. The above-mentioned multi-source remote sensing and monitoring data were registered in a unified coordinate system, orthorectified and spatially resampled to construct a multi-source data cube for the railway corridor; Based on the above-mentioned multi-source data cube of the railway corridor, remote sensing sensitive factors characterizing salinity, ice content, thermal state and surface deformation are extracted. Among them, the remote sensing sensitive factors include spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation amplitude factor, InSAR linear deformation rate and seasonal deformation amplitude factor. Using the aforementioned remote sensing sensitive factors, a comprehensive index of salinized permafrost was constructed, and the comprehensive index was physically constrained and corrected based on the heat-salt coupling relationship of permafrost to obtain the corrected comprehensive permafrost index. The above-corrected permafrost comprehensive index and the above-mentioned remote sensing sensitive factors are input into the multi-task deep learning model, which outputs the probability of permafrost existence, salinization level and engineering risk level, respectively. Based on the above-mentioned engineering risk levels, a risk profile of saline permafrost was generated along the railway mileage, and risk zones were identified along the railway line.

[0006] In one feasible implementation, the above-mentioned multi-source remote sensing and monitoring data undergoes unified coordinate system registration, orthorectification, and spatial resampling to construct a multi-source data cube for the railway corridor, including: For the multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data and InSAR deformation time series data in the above multi-source remote sensing and monitoring data, a unified coordinate system registration is performed based on the ground control points of the railway line, and terrain correction and orthorectification are performed using the above digital elevation model. Spatial resampling is performed on the registered and corrected image data according to the preset spatial resolution to obtain a remote sensing image dataset with a unified spatial grid. The above-mentioned geothermal monitoring data and borehole salt profile data are mapped to the corresponding grid cells in the above-mentioned unified spatial grid by spatial interpolation or along-line projection. Using the railway line centerline as a reference, all resampled grid data are indexed according to time sequence and mileage direction to construct a multi-source data cube for railway corridors with a unified spatial grid, unified time step and railway mileage coordinate system.

[0007] In one feasible implementation, the remote sensing sensitive factors characterizing salinity, ice content, thermal state, and surface deformation features based on the aforementioned multi-source data cube of the railway corridor include: The spectral salinity index is calculated based on the above multispectral or hyperspectral images to characterize the salt distribution of the surface soil along the railway line. Based on the above synthetic aperture radar images, radar scattering coefficient combination factors are extracted to characterize the ice content and water content of the active layer along the railway line. The surface temperature sequence was retrieved based on the above thermal infrared data and the annual variation factor of surface temperature was calculated to characterize the surface thermal state along the above railway line. Based on the aforementioned InSAR deformation time series data, the InSAR linear deformation rate and seasonal deformation amplitude factor were obtained to characterize the frost heave and thaw settlement deformation features along the railway line.

[0008] In one feasible implementation, the aforementioned remote sensing sensitive factors are used to construct a comprehensive index for salinized permafrost, and the comprehensive index is physically constrained and corrected based on the heat-salt coupling relationship of permafrost to obtain a corrected comprehensive permafrost index, including: Based on the above spectral salinity index, the above radar scattering coefficient combination factor, the above annual surface temperature variation factor, the above InSAR linear deformation rate and the above seasonal deformation amplitude factor, a comprehensive index of salinized permafrost is constructed according to the multiple regression method. The physical consistency correction of the above-mentioned comprehensive index of salinized permafrost is performed based on the thermal-salt coupling relationship of permafrost, and the corrected comprehensive index of salinized permafrost is output. When the annual variation factor of the above-mentioned surface temperature is lower than the preset temperature threshold and the above-mentioned spectral salinity index is higher than the preset salinity threshold, the above-mentioned comprehensive index is corrected to a high-salt-high-freezing state; when the above-mentioned surface temperature is close to zero degrees Celsius and the above-mentioned seasonal deformation amplitude factor exceeds the preset deformation threshold, the above-mentioned comprehensive index is corrected to a strong melt subsidence risk state.

[0009] In one feasible implementation, the aforementioned multi-task deep learning model includes a shared feature extraction module, a physical consistency gating layer, and three task decoding heads. The above-mentioned corrected permafrost comprehensive index and the above-mentioned remote sensing sensitive factors are input into a multi-task deep learning model, which outputs the probability of permafrost existence, salinization level, and engineering risk level, respectively, including: The corrected permafrost comprehensive index, along with the spectral salinity index, the radar scattering coefficient combination factor, the annual surface temperature variation factor, the InSAR linear deformation rate, and the seasonal deformation amplitude factor, are input into the shared feature extraction module. The shared feature extraction module jointly encodes spatial and temporal features through a convolutional neural network sub-layer, a long short-term memory network sub-layer, and a cross-modal attention layer to obtain the unified feature representation. The above-mentioned unified feature representation is input into the above-mentioned physical consistency gating layer. The above-mentioned physical consistency gating layer generates consistency control weights based on the above-mentioned corrected permafrost comprehensive index, the above-mentioned annual surface temperature variation amplitude factor and the above-mentioned seasonal deformation amplitude factor, and performs physical constraint renormalization on the above-mentioned unified feature representation. The features renormalized by the above-mentioned physical consistency gating layer are respectively input into the above-mentioned permafrost identification decoding head, the above-mentioned salinization level classification decoding head and the above-mentioned engineering risk assessment decoding head to generate the above-mentioned permafrost existence probability, the above-mentioned salinization level and the above-mentioned engineering risk level.

[0010] In one feasible implementation, the above-mentioned risk profile of saline permafrost is generated along the railway mileage direction based on the aforementioned engineering risk level, and risk zoning is identified along the railway line, including: By combining the above-mentioned engineering risk levels with the above-mentioned digital elevation model and the geometric information of the railway line centerline, the grid cells corresponding to the above-mentioned engineering risk levels are subjected to track-direction reprojection processing to obtain the risk mileage sequence along the above-mentioned railway line mileage direction. Based on the above risk mileage sequence, a risk profile in the mileage direction is constructed according to a preset mileage step size. In the construction process, the above InSAR linear deformation rate and the above seasonal deformation amplitude factor are integrated to enhance the ability of the risk profile to characterize the sensitivity of frost heave and thaw settlement. Dynamic grading processing is performed on the above risk profile to obtain dynamic grading results; Based on the above dynamic grading results, a risk zoning identification map along the railway line is generated, realizing the refined zoning of the above-mentioned saline permafrost risk in the direction of railway mileage.

[0011] In one feasible implementation, the above method further includes: When the risk level of multiple adjacent mileage sections continuously crosses the preset risk change threshold, the above sections are classified as continuous high-risk sections. The risk level of the aforementioned high-risk areas was further corrected based on the corrected permafrost comprehensive index.

[0012] In one feasible implementation, the aforementioned risk profile is constructed based on the aforementioned risk mileage sequence according to a preset mileage step size, and the aforementioned InSAR linear deformation rate and the aforementioned seasonal deformation amplitude factor are integrated during the construction process to enhance the characterization ability of the risk profile on the sensitivity of frost heave and thaw settlement, including: The above InSAR linear deformation rate and the above seasonal deformation amplitude factor are subjected to orbital interpolation to align the above deformation factor with the above risk mileage sequence in the same mileage coordinate system. The aligned deformation factors are fused with the aforementioned risk milestone sequence through a weighted superposition method. The weighted superposition method adaptively adjusts the weights based on the magnitude of the aforementioned seasonal deformation amplitude factors to highlight the frost heave or thaw settlement response characteristics of high deformation sections. Based on the fused risk enhancement sequence, a risk profile is constructed according to the preset mileage step size, so that the risk profile can reflect the deformation sensitivity distribution characteristics along the mileage direction of the railway line.

[0013] In one feasible implementation, the above-mentioned generation of a risk zoning identification map along the railway line according to the above-mentioned dynamic classification results includes: Based on the above risk profile, a mileage-based risk grid sequence along the above railway line is constructed, and the slope factor derived from the above digital elevation model is used to perform slope aspect correction on the above mileage-based risk grid sequence in order to suppress risk jumps caused by abrupt terrain changes. A variable threshold-based boundary detection algorithm is executed on the risk grid sequence after slope correction to extract the partition boundaries between different risk levels. The variable threshold is dynamically adjusted according to the local variation of the seasonal deformation amplitude factor to enhance the response of boundary detection to melting and subsidence sensitive sections. Based on the extracted risk zoning boundaries, a spatially continuous risk zoning identification map is generated to support the selection of structural types and the formulation of engineering treatment strategies for high-risk sections along railway lines.

[0014] Secondly, the present invention also proposes a remote sensing identification and control system for railways in saline-alkali permafrost, used to execute the remote sensing identification method for railways in saline-alkali permafrost as described in any one of the first aspects, including: The acquisition unit is used to acquire multi-source remote sensing and monitoring data within a preset buffer zone centered on the railway line. The multi-source remote sensing and monitoring data includes multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, InSAR deformation time series data, digital elevation model, ground temperature monitoring and borehole salinity profile data. The construction unit is used to perform unified coordinate system registration, orthorectification and spatial resampling on the above-mentioned multi-source remote sensing and monitoring data in order to construct a multi-source data cube for the railway corridor; The extraction unit is used to extract remote sensing sensitive factors characterizing salinity, ice content, thermal state and surface deformation based on the multi-source data cube of the railway corridor mentioned above. The remote sensing sensitive factors include spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation amplitude factor, InSAR linear deformation rate and seasonal deformation amplitude factor. The correction unit is used to construct a comprehensive index of salinized permafrost using the aforementioned remote sensing sensitive factors, and to perform physical constraint correction on the comprehensive index based on the heat-salt coupling relationship of permafrost, so as to obtain the corrected permafrost comprehensive index. The output unit is used to input the above-corrected permafrost comprehensive index and the above-mentioned remote sensing sensitive factors into the multi-task deep learning model, and output the probability of permafrost existence, salinization level and engineering risk level, respectively. The identification unit is used to generate a risk profile of saline permafrost along the railway mileage based on the above-mentioned engineering risk level, and to identify risk zones along the railway line.

[0015] In summary, compared with traditional permafrost identification methods that rely on single data sources or local surveys, this invention, by constructing a multi-source data cube along a railway corridor, achieves spatiotemporal integration of spectral, radar scattering, thermal infrared temperature, InSAR deformation, and ground monitoring information. This eliminates the limitations of single-data bias or the randomness of observation periods in the identification of saline permafrost. At the sensitive factor level, this invention simultaneously introduces four types of indicators characterizing salinity, ice content, thermal state, and deformation behavior. Through the synergistic characterization of spectral, thermal, radar, and deformation fields, the description of the permafrost's thermal-water-salt state is more comprehensive and physically consistent, significantly improving the reliability of the identification results. By constructing a comprehensive index for saline permafrost and introducing the thermal-salt coupling relationship of permafrost for physical consistency correction, this invention effectively avoids misjudgments that occur under extreme conditions with purely statistical methods. The comprehensive index possesses both data-driven discriminative power and adheres to the physical mechanisms of permafrost regions, thus making the comprehensive judgment on salt enrichment, temperature fluctuations, and thaw settlement risk more accurate. At the intelligent identification level, this invention employs a multi-task deep learning model, simultaneously outputting the probability of permafrost presence, salinization level, and engineering risk level. It achieves joint expression of multi-dimensional features through shared feature extraction and cross-modal attention mechanisms, and sets up a physical consistency gating layer to ensure consistency between the deep feature flow and the freeze-thaw-saltification-deformation coupling relationship, making the model predictions more engineering interpretable and verifiable. Furthermore, this invention is the first to propose projecting the engineering risk identification results onto the railway mileage direction, forming a "mileage-risk" profile. Combined with InSAR deformation enhancement, dynamic grading, and variable threshold boundary detection mechanisms, it achieves the identification and secondary correction of continuous high-risk sections along the railway line, thereby generating a risk zoning identification map with spatial continuity and direct engineering significance. Through this innovative process, this invention can directly support railway route optimization, subgrade structure design, monitoring deployment of key mileage sections, and operation and maintenance strategy formulation, realizing a complete closed loop from "remote sensing identification" to "engineering decision-making," significantly improving the quantification and intelligent level of risk in railway construction and operation in saline permafrost areas.

[0016] Other advantages, objectives and features of this application will be partly apparent from the description below, and partly understood by those skilled in the art through study and practice of this application. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating a remote sensing identification method for railways in saline-alkali permafrost, provided by the present invention.

[0018] Figure 2 A flowchart illustrating a method for constructing a multi-source data cube for a railway corridor, provided by this invention. Figure 3 This invention provides a flowchart illustrating a method for extracting data from a multi-source data cube of a railway corridor. Figure 4 This is a flowchart illustrating a method for obtaining a corrected comprehensive index of frozen soil provided by the present invention. Figure 5 This is a flowchart illustrating the specific execution method of a multi-task deep learning model provided by the present invention. Figure 6 This is a flowchart illustrating a method for identifying risk zones along railway lines, provided by the present invention. Figure 7 This invention provides a flowchart illustrating a method for secondary correction of risk levels. Figure 8 A flowchart illustrating a method for enhancing the characterization ability of risk profiles to resist frost heave and thaw settlement, provided by this invention. Figure 9 This invention provides a flowchart illustrating a method for generating a risk zoning identification map along a railway line. Figure 10 This invention provides a structural schematic diagram of a remote sensing identification and control system for railways in saline-alkali permafrost. Detailed Implementation

[0019] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0020] Please see Figure 1 This is a flowchart illustrating a remote sensing identification method for railways in saline-alkali permafrost, provided in an embodiment of this application. Specifically, it may include: S110. Acquire multi-source remote sensing and monitoring data within a preset buffer zone centered on the railway line. The multi-source remote sensing and monitoring data includes multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, InSAR deformation time series data, digital elevation model, ground temperature monitoring and borehole salinity profile data. S120. Perform unified coordinate system registration, orthorectification and spatial resampling on the above multi-source remote sensing and monitoring data to construct a multi-source data cube for the railway corridor. S130. Based on the above-mentioned multi-source data cube of the railway corridor, extract remote sensing sensitive factors that characterize salinity, ice content, thermal state and surface deformation. The remote sensing sensitive factors include spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation amplitude factor, InSAR linear deformation rate and seasonal deformation amplitude factor. S140. Construct a comprehensive index of salinized permafrost using the above remote sensing sensitive factors, and perform physical constraint correction on the comprehensive index based on the heat-salt coupling relationship of permafrost to obtain the corrected permafrost comprehensive index. S150. Input the above-corrected permafrost comprehensive index and the above-mentioned remote sensing sensitive factors into the multi-task deep learning model, and output the probability of permafrost existence, salinization level and engineering risk level respectively. S160. Based on the above-mentioned engineering risk levels, generate a risk profile of saline permafrost along the railway mileage and identify risk zones along the railway line.

[0021] For example, in step S110, a pre-defined buffer zone of a certain width is set on both sides of the proposed or existing railway line as the centerline, for example, several kilometers on each side of the centerline. Within this buffer zone, multi-source remote sensing and ground monitoring data are systematically acquired. The multi-source remote sensing and monitoring data includes not only multispectral or hyperspectral images reflecting the spectral characteristics of the Earth's surface, but also synthetic aperture radar images reflecting the dielectric properties and roughness of the surface and shallow layers. It also includes thermal infrared data for retrieving the surface temperature field, InSAR deformation time series data for monitoring surface deformation processes, and a digital elevation model for characterizing topographic relief. Simultaneously, to establish a correspondence between remote sensing information and actual permafrost-saline conditions, ground temperature monitoring data and borehole salinity profile data are also collected along the railway line, providing ground-based measurement data for subsequent model calibration and correction.

[0022] In step S120, a unified spatiotemporal framework is applied to the aforementioned data. Since data from different sensors exhibit variations in coordinate systems, imaging geometry, and spatial resolution, this embodiment employs unified coordinate system registration, orthorectification using a digital elevation model, and resampling of image data according to a unified spatial resolution. This aligns multispectral, hyperspectral, radar, thermal infrared, and InSAR data within the same spatial grid. Simultaneously, discretely distributed ground temperature monitoring points and borehole salinity profile data are mapped to the same grid cells via spatial interpolation or along-line projection. Ultimately, a multi-source data cube with a unified spatial grid and time step is constructed within the railway line centerline and its buffer zone, achieving integrated spatiotemporal organization of spectral, radar, thermal, deformation, and ground monitoring information.

[0023] In step S130, based on the multi-source data cube of the railway corridor, a set of remote sensing sensitive factors with engineering orientation are extracted for four key physical processes: salinity, ice content, thermal state, and surface deformation. On the one hand, a spectral salinity index is constructed using multispectral or hyperspectral imagery to characterize the degree of salt enrichment in the surface soil. On the other hand, a radar scattering coefficient combination factor is constructed by combining the backscattering characteristics of synthetic aperture radar imagery to characterize the difference between ice content and water content in the active layer. Simultaneously, the surface temperature time series is retrieved based on thermal infrared data, and the annual variation amplitude factor of surface temperature is extracted to reflect the surface thermal state and freeze-thaw intensity. Finally, InSAR deformation time series analysis is used to obtain the InSAR linear deformation rate and seasonal deformation amplitude factor to characterize the deformation pattern caused by frost heave and thaw settlement in permafrost regions. Through the extraction of the above multi-source factors, this embodiment improves the identification mode of the traditional single data source to a four-field synergistic characterization of "salt-ice-heat-deformation".

[0024] In step S140, a comprehensive index for salinized permafrost is constructed using the aforementioned remote sensing sensitive factors. Specifically, this can be achieved by combining the spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation factor, InSAR linear deformation rate, and seasonal deformation amplitude factor using a preset weighting method or multiple regression method, resulting in an initial comprehensive index reflecting the comprehensive heat-water-salt-deformation state of each grid cell. Considering the clear physical coupling relationship between geothermal field, salt migration, and freeze-thaw deformation in permafrost regions, this embodiment further introduces the heat-salt coupling relationship of permafrost to perform physical constraint correction on the comprehensive index: for example, in areas with low annual surface temperature variation and high spectral salinity index, the comprehensive index is corrected towards a high-salt-high-freeze state. In areas with surface temperature close to 0℃ and large seasonal deformation amplitude, the comprehensive index is adjusted to a state with a strong risk of thaw settlement. Through such physical consistency correction, a corrected permafrost comprehensive index that more accurately reflects the state of salinized permafrost is obtained, reducing the misjudgment that may occur if only statistical combinations are relied upon.

[0025] In step S150, the corrected permafrost comprehensive index and various remote sensing sensitive factors are input into a multi-task deep learning model. This model uses a shared feature extraction module to jointly encode spatial-temporal-multi-source features through convolutional neural networks, temporal networks, and cross-modal attention mechanisms, obtaining a unified deep feature representation for each location along the railway corridor. Subsequently, the model outputs the corresponding permafrost presence probability, salinization level, and engineering risk level through three task decoding heads: permafrost identification, salinization level classification, and engineering risk assessment. During model training, physical consistency constraints based on the relationship between ground temperature, salinity, and deformation can be introduced to ensure physical coordination between the multi-task outputs, thereby improving the correlation and interpretability between the identification results and actual engineering defects.

[0026] In step S160, based on the spatial distribution results of the aforementioned engineering risk levels, a risk profile of saline permafrost is generated along the mileage direction, using the railway line centerline as a reference. Specifically, the engineering risk levels on the planar grid can be projected or summarized into the railway mileage coordinate system, forming a continuous "mileage-risk" curve or strip distribution at the monthly or multi-year scale. On this basis, the risk profile can be dynamically graded and its boundaries identified by combining topographic conditions, deformation characteristics, etc., dividing the line into risk zones of different levels, such as general risk zones, higher risk zones, and high risk zones. This results in the identification of saline permafrost risk zones for railway alignment optimization, subgrade structure selection, and operation and maintenance strategy formulation, providing direct support from remote sensing identification to engineering decision-making.

[0027] In summary, compared with traditional permafrost identification methods that rely on single data sources or local surveys, this invention, by constructing a multi-source data cube along a railway corridor, achieves spatiotemporal integration of spectral, radar scattering, thermal infrared temperature, InSAR deformation, and ground monitoring information. This eliminates the limitations of single-data bias or the randomness of observation periods in the identification of saline permafrost. At the sensitive factor level, this invention simultaneously introduces four types of indicators characterizing salinity, ice content, thermal state, and deformation behavior. Through the synergistic characterization of spectral, thermal, radar, and deformation fields, the description of the permafrost's thermal-water-salt state is more comprehensive and physically consistent, significantly improving the reliability of the identification results. By constructing a comprehensive index for saline permafrost and introducing the thermal-salt coupling relationship of permafrost for physical consistency correction, this invention effectively avoids misjudgments that occur under extreme conditions with purely statistical methods. The comprehensive index possesses both data-driven discriminative power and adheres to the physical mechanisms of permafrost regions, thus making the comprehensive judgment on salt enrichment, temperature fluctuations, and thaw settlement risk more accurate. At the intelligent identification level, this invention employs a multi-task deep learning model, simultaneously outputting the probability of permafrost presence, salinization level, and engineering risk level. It achieves joint expression of multi-dimensional features through shared feature extraction and cross-modal attention mechanisms, and sets up a physical consistency gating layer to ensure consistency between the deep feature flow and the freeze-thaw-saltification-deformation coupling relationship, making the model predictions more engineering interpretable and verifiable. Furthermore, this invention is the first to propose projecting the engineering risk identification results onto the railway mileage direction, forming a "mileage-risk" profile. Combined with InSAR deformation enhancement, dynamic grading, and variable threshold boundary detection mechanisms, it achieves the identification and secondary correction of continuous high-risk sections along the railway line, thereby generating a risk zoning identification map with spatial continuity and direct engineering significance. Through this innovative process, this invention can directly support railway route optimization, subgrade structure design, monitoring deployment of key mileage sections, and operation and maintenance strategy formulation, realizing a complete closed loop from "remote sensing identification" to "engineering decision-making," significantly improving the quantification and intelligent level of risk in railway construction and operation in saline permafrost areas.

[0028] In one feasible implementation, such as Figure 2 As shown, step S120 above performs unified coordinate system registration, orthorectification, and spatial resampling on the above multi-source remote sensing and monitoring data to construct a multi-source data cube for the railway corridor, including: S1201. For the multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data and InSAR deformation time series data in the above multi-source remote sensing and monitoring data, perform unified coordinate system registration based on the ground control points of the railway line, and use the above digital elevation model to perform terrain correction and orthorectification. S1202. Spatial resampling is performed on the registered and corrected image data according to the preset spatial resolution to obtain a remote sensing image dataset with a unified spatial grid. S1203. Map the above-mentioned geothermal monitoring data and borehole salt profile data to the corresponding grid cells in the above-mentioned unified spatial grid by spatial interpolation or along-line projection. S1204. Using the railway line centerline as a reference, all resampled grid data are indexed according to time sequence and mileage direction to construct a railway corridor multi-source data cube with a unified spatial grid, unified time step and railway mileage coordinate system.

[0029] For example, the process of performing unified coordinate system registration, orthorectification and spatial resampling on the above-mentioned multi-source remote sensing and monitoring data aims to integrate data from different sensors, different sampling scales and different acquisition methods into a multi-source data cube of railway corridors suitable for railway engineering analysis.

[0030] Specifically, firstly, ground control points along the railway line are selected as spatial references. Then, multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, and InSAR deformation time series data from the aforementioned multi-source remote sensing and monitoring data are registered in a unified coordinate system, ensuring strict alignment of all types of image data under the same geographic coordinate reference system. After completing coordinate system one, the aforementioned digital elevation model is further used to perform terrain correction and orthorectification on the various image data, correcting spatial distortions caused by terrain undulations, sensor perspective, and imaging geometry, thereby obtaining an orthorectified image sequence that accurately reflects the positional relationships of the Earth's surface.

[0031] After registration and orthorectification, to eliminate the impact of spatial resolution differences between different sensors on subsequent analysis, this embodiment spatially resamples the registered and corrected image data according to a preset spatial resolution. This unifies multispectral, hyperspectral, radar, and thermal infrared images onto the same grid size and spatial grid, thus forming a remote sensing image dataset with a unified spatial grid. Simultaneously, considering that geothermal monitoring data and borehole salinity profile data typically exist as discrete points or along profiles, to enable joint analysis with the aforementioned remote sensing image data within the same data structure, this embodiment maps the geothermal monitoring data and borehole salinity profile data to corresponding grid cells in the unified spatial grid using spatial interpolation or along-line projection. This achieves simultaneous image features and ground-measured attributes within each grid cell. Finally, using the railway line centerline as a reference benchmark, all resampled grid data are indexed in chronological order and mileage direction. This forms a multi-temporal sequence structure in the time dimension and establishes an index relationship with railway mileage in the spatial dimension, thereby constructing a multi-source data cube for the railway corridor with a unified spatial grid, a unified time step, and a railway mileage coordinate system.

[0032] Through this data cube, this embodiment achieves a high degree of integration and standardization of spectral information, radar scattering information, thermal infrared temperature information, InSAR deformation information, and ground temperature and salinity monitoring information within the railway corridor, providing accurate, unified, and railway-application-oriented basic data support for subsequent remote sensing sensitive factor extraction, comprehensive index construction, and engineering risk identification.

[0033] In one feasible implementation, such as Figure 3 As shown, step S130 above extracts remote sensing sensitive factors characterizing salinity, ice content, thermal state, and surface deformation features based on the multi-source data cube of the railway corridor, including: S1301. Calculate the spectral salinity index based on the above multispectral or hyperspectral images to characterize the salt distribution of the surface soil along the railway line. S1302. Based on the above synthetic aperture radar image, extract the radar scattering coefficient combination factor to characterize the ice content and water content of the active layer along the railway line. S1303. Based on the above thermal infrared data, the surface temperature sequence is retrieved and the annual variation factor of surface temperature is calculated to characterize the surface thermal state along the above railway line. S1304. Based on the above InSAR deformation time series data, obtain the InSAR linear deformation rate and seasonal deformation amplitude factor to characterize the frost heave and thaw settlement deformation characteristics along the above railway line.

[0034] For example, based on the aforementioned multi-source data cube of the railway corridor, remote sensing sensitive factors characterizing salinity, ice content, thermal state, and surface deformation are extracted. The aim is to transform multi-source, multi-temporal remote sensing and monitoring information into a set of characteristic quantities that directly reflect key processes of salinized permafrost, providing physically meaningful input for subsequent comprehensive index construction and deep learning recognition. Regarding salinity characterization, this embodiment utilizes multispectral or hyperspectral imagery to construct a spectral salinity index, characterizing the spectral response of surface crystalline salt and saline soil through the reflectance difference between near-infrared and short-wave infrared bands. For example, the spectral salinity index can be expressed as... in, This represents the surface reflectance in the shortwave infrared band. This represents the surface reflectance in the near-infrared band. A very small positive number is introduced to avoid numerical instability when the denominator is close to zero. The shortwave infrared band is sensitive to salt crust, moisture, and clay minerals, while the near-infrared band is sensitive to vegetation and moisture. By using the normalized difference between the two, the brightening effect and spectral characteristics of saline surfaces can be highlighted, thereby achieving a qualitative or semi-quantitative characterization of the salt distribution in the surface soil along the railway.

[0035] In characterizing ice content and water content, this embodiment extracts a radar scattering coefficient combination factor based on synthetic aperture radar imagery, utilizing the sensitivity of radar backscattering to dielectric constant and roughness to reflect changes in the ice-water ratio in the active layer. The following combination factor can be constructed: in, and These represent the normalized radar backscattering coefficients for the vertical-vertical polarization and vertical-horizontal polarization channels, respectively. These are weighting coefficients obtained through fitting or training based on ground-measured ice and water content samples. Generally, increased ice content alters the effective dielectric constant and scattering mechanism of the soil, leading to differences in scattering responses across different polarization channels. By combining these linear or near-linear parameters, these differences can be summarized into a comprehensive index to characterize the permafrost structure and water-ice state of the active layer along the railway line.

[0036] In characterizing the surface thermal state, this embodiment retrieves the surface temperature time series based on thermal infrared data and calculates the annual variation amplitude factor of surface temperature to depict the seasonal oscillation intensity of the surface thermal environment in permafrost regions. For example, the surface temperature series of a certain grid cell over a year can be denoted as... The annual variation factor can then be expressed as: in, Indicates time The surface temperature, and These represent the maximum and minimum values ​​within a one-year time window, respectively. A larger annual variation generally indicates a thicker active layer and stronger seasonal freeze-thaw cycles. However, in high-altitude or permafrost regions, the annual amplitude of surface temperature typically converges. This can be achieved by comparing values ​​at different mileage intervals. The range can be used to identify frozen soil units with significant differences in thermal state. In practical applications, the root mean square amplitude or other forms can be used instead of the range; this embodiment is not limited to the above formula.

[0037] In terms of characterizing surface deformation features, this embodiment is based on Deformation time series data, extract respectively Two factors, linear deformation rate and seasonal deformation amplitude, are used to reflect the changes in surface elevation along the railway line due to frost heave, thaw settlement, and salinization. For the linear deformation rate, the set of observation times can be denoted as... The corresponding deformation along the radar line of sight is denoted as The long-term deformation rate can then be estimated using linear fitting or simplified averaging, for example: in, This represents the average linear deformation rate along the radar line of sight. For a moment Cumulative deformation, This is a reference time. If it needs to be converted to vertical deformation rate, it can be geometrically projected in conjunction with the incident angle. The linear rate reflects the long-term settlement or uplift trend and can indicate potential permafrost degradation or soft soil compression processes.

[0038] For seasonal deformation amplitude, this embodiment can approximate the deformation time series as a simplified model containing a linear trend and an annual cycle component, for example: in, Represents the initial deformation reference term. For the aforementioned linear deformation rate, This represents the amplitude of seasonal deformation. The cycle is typically one year (e.g., using days or months as the time unit). The phase is used to describe the offset between the deformation peak and the time axis. This is obtained through fitting. This refers to the seasonal deformation amplitude factor. A larger factor indicates a stronger response to frost heave and thaw settlement within a freeze-thaw cycle, making the subgrade or structure overlying permafrost more prone to periodic deformation and cracking. By combining... and This embodiment can simultaneously capture long-term degradation trends and short-term freeze-thaw responses, thereby more comprehensively characterizing the deformation behavior of saline permafrost along railway lines.

[0039] In summary, this embodiment extracts remote sensing sensitive factors such as spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation factor, and InSAR linear deformation rate and seasonal deformation amplitude from a multi-source data cube of a railway corridor. This allows for a unified quantitative characterization of salinity enrichment, active layer ice content and water content, surface thermal environment, and frost heave-thaw deformation processes. This approach preserves the specificity of each physical process and provides standardized, multi-dimensional feature inputs for subsequent comprehensive index construction and multi-task deep learning recognition. The above formula is only one feasible implementation method; those skilled in the art can make equivalent modifications based on data sources and accuracy requirements without affecting the overall concept of this embodiment.

[0040] In one feasible implementation, such as Figure 4 As shown, step S140 above utilizes the aforementioned remote sensing sensitive factors to construct a comprehensive index for salinized permafrost, and performs physical constraint correction on the comprehensive index based on the heat-salt coupling relationship of permafrost to obtain the corrected comprehensive permafrost index, including: S1401. Based on the above spectral salinity index, the above radar scattering coefficient combination factor, the above annual surface temperature variation factor, the above InSAR linear deformation rate and the above seasonal deformation amplitude factor, a comprehensive index of salinized permafrost is constructed according to the multiple regression method. S1402. Based on the thermal-salt coupling relationship of permafrost, perform physical consistency correction on the above-mentioned comprehensive index of salinized permafrost and output the corrected comprehensive index of salinized permafrost. When the annual variation factor of the above-mentioned surface temperature is lower than the preset temperature threshold and the above-mentioned spectral salinity index is higher than the preset salinity threshold, the above-mentioned comprehensive index is corrected to a high-salt-high-freezing state. When the above-mentioned surface temperature is close to zero degrees Celsius and the above-mentioned seasonal deformation amplitude factor exceeds the preset deformation threshold, the above-mentioned comprehensive index is corrected to a strong melt subsidence risk state.

[0041] For example, a comprehensive index of salinized permafrost is constructed using the aforementioned remote sensing sensitive factors, and the comprehensive index is physically constrained and corrected based on the heat-salt coupling relationship of permafrost. The purpose is to compress the multidimensional sensitive factors obtained in the previous step into a scalar index that can comprehensively reflect the degree of salt enrichment, the thermal state of permafrost, and the risk of deformation. At the same time, by introducing the known heat-salt-deformation mechanism in the permafrost region, the results of pure data-driven analysis are physically corrected, thereby obtaining a corrected comprehensive permafrost index that is more in line with engineering practice.

[0042] Specifically, based on the aforementioned spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation factor, InSAR linear deformation rate, and seasonal deformation amplitude factor, an initial comprehensive index for salinized permafrost is constructed using a multiple regression approach. In a typical implementation, a multiple regression model can be established between the comprehensive index and labels such as whether it is salinized permafrost and the permafrost stability level, using existing ground temperature observations, borehole salinity profiles, and engineering defect investigation results within the sample area. This yields a set of regression coefficients. The aforementioned multiple sensitive factors are then combined linearly or nonlinearly with certain weights into a comprehensive index, which can be formally written as: Comprehensive Index = b0 + b1 × spectral salinity index + b2 × radar scattering coefficient combination factor + b3 × annual surface temperature variation factor + b4 × InSAR linear deformation rate + b5 × seasonal deformation amplitude factor, where b0 is the intercept term, and b1 to b5 are the weight coefficients obtained through regression. Through such multiple regression, the comprehensive index numerically reflects the importance of each factor in the identification of saline permafrost, realizing the mapping from a high-dimensional feature space to a one-dimensional risk characterization space.

[0043] After obtaining the initial comprehensive index of salinized permafrost, this embodiment further introduces the thermal-salt coupling relationship of permafrost to perform physical consistency correction on the comprehensive index and outputs the corrected comprehensive index of salinized permafrost. Although multiple regression can statistically fit an index that is highly correlated with the sample label, it may give abnormal results in some physically unreasonable scenarios if unconstrained. Therefore, it is necessary to use the empirical laws between thermal field, salinity and deformation behavior in permafrost areas to perform secondary screening and correction of the index. For example, when the annual surface temperature variation factor of a certain grid is significantly lower than the preset temperature threshold, it indicates that the temperature fluctuation in this area is small throughout the year and close to a continuous low temperature state. However, its spectral salinity index is higher than the preset salinity threshold, indicating that the surface soil is significantly enriched with salt. Under this situation of low thermal amplitude and high salinity, according to the understanding of permafrost thermal-salt coupling, this location is more likely to belong to a stable or semi-stable permafrost unit with high salinity and high freezing. Therefore, this embodiment corrects the comprehensive index to the numerical range corresponding to the high salinity and high freezing category, so that it is closer to the high salinity and high freezing state in the comprehensive index space. Conversely, when the annual variation of the surface temperature of a certain grid is close to the freezing-thawing critical range near zero degrees Celsius, and the seasonal deformation amplitude factor exceeds the preset deformation threshold, it means that there is a significant frost heave and thaw settlement response in the surface of the area within a one-year cycle. This phenomenon usually corresponds to structural degradation caused by significant thermal disturbance, thickening of the active layer, or softening of salt in the overlying or adjacent areas of permafrost. Under this combination of critical thermal state and large amplitude deformation, this embodiment corrects the comprehensive index to a state of strong thaw settlement risk, that is, moves the initial regression result to a numerical range representing high risk and easy degradation.

[0044] Through the two types of typical rules mentioned above and other scalable physical constraints, this embodiment essentially constructs a dual constraint mechanism of data-driven results and a physical rule base. Multiple regression ensures that the comprehensive index can effectively distinguish different types of saline-deformation permafrost units in a statistical sense, while the heat-salt-deformation coupling correction ensures that the index still follows the basic laws of permafrost engineering geology under extreme conditions or in areas with insufficient samples, thus obtaining a more stable and physically interpretable corrected permafrost comprehensive index. Ultimately, this corrected comprehensive index can not only serve as an important input feature for subsequent multi-task deep learning models to improve the consistency of the outputs of permafrost existence probability, salinization level, and engineering risk level, but it can also be directly used to construct risk profiles along railway mileage, providing quantitative basic indicators for engineering route selection, subgrade structure selection, and long-term operational monitoring.

[0045] In one feasible implementation, the aforementioned multi-task deep learning model includes a shared feature extraction module, a physical consistency gating layer, and three task decoding heads. like Figure 5As shown, step S150 above inputs the corrected permafrost comprehensive index and the remote sensing sensitive factors into the multi-task deep learning model, and outputs the probability of permafrost existence, salinization level, and engineering risk level, including: S1501. Input the above-corrected permafrost comprehensive index, the above-mentioned spectral salinity index, the above-mentioned radar scattering coefficient combination factor, the above-mentioned annual surface temperature variation factor, the above-mentioned InSAR linear deformation rate, and the above-mentioned seasonal deformation amplitude factor into the above-mentioned shared feature extraction module. The above-mentioned shared feature extraction module jointly encodes spatial features and temporal features through convolutional neural network sub-layers, long short-term memory network sub-layers, and cross-modal attention layers to obtain the above-mentioned unified feature representation. S1502. Input the above-mentioned unified feature representation into the above-mentioned physical consistency gating layer. The above-mentioned physical consistency gating layer generates consistency control weights based on the above-mentioned corrected permafrost comprehensive index, the above-mentioned annual surface temperature variation amplitude factor and the above-mentioned seasonal deformation amplitude factor, and performs physical constraint renormalization on the above-mentioned unified feature representation. S1503. The features after being renormalized by the above-mentioned physical consistency gating layer are respectively input into the above-mentioned permafrost identification decoding head, the above-mentioned salinization level classification decoding head and the above-mentioned engineering risk assessment decoding head to generate the above-mentioned permafrost existence probability, the above-mentioned salinization level and the above-mentioned engineering risk level respectively.

[0046] For example, in this embodiment, the multi-task deep learning model is designed to consist of a shared feature extraction module, a physical consistency gating layer, and three task decoding heads connected in series. The purpose is to simultaneously complete three types of tasks—permafrost presence identification, salinization level discrimination, and engineering risk assessment—within the same network framework. The model also explicitly introduces the constraints of the permafrost heat-salt-deformation mechanism into the network, so that the model output not only has statistically significant distinguishing ability but also physical consistency and interpretability.

[0047] Specifically, in the feature input stage, the corrected permafrost comprehensive index, along with the spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation factor, InSAR linear deformation rate, and seasonal deformation amplitude factor, are used as input features of the model. The corrected permafrost comprehensive index, having already incorporated the combined effects of multiple factors and undergone thermal-salt coupling physical correction in the preceding steps, represents the overall thermal-water-salt-deformation state of each grid cell. The spectral salinity index primarily reflects the salt enrichment degree of the surface soil. The radar scattering coefficient combination factor reflects the changes in ice and water content in the active layer. The annual surface temperature variation factor characterizes the surface thermal environment and freeze-thaw intensity. The InSAR linear deformation rate and seasonal deformation amplitude describe frost heave and thaw settlement behavior from both long-term trend and seasonal response perspectives. These features are first concatenated into a multi-dimensional feature vector sequence and then fed into the shared feature extraction module in temporal and spatial order.

[0048] In the shared feature extraction module, the spatial distribution pattern of features is first extracted through a convolutional neural network sublayer. The input can be represented as a two-dimensional or one-dimensional raster sequence, and spatial correlations between factors such as salinity, temperature, and deformation in the local neighborhood are extracted through multiple convolutional operations. For example, suppose the input feature tensor is... in, Indicates spatial dimension (such as the number of grids in the mileage direction). Indicates the optional horizontal dimension (such as the number of grids along the width of the line buffer). This represents the number of input channels, including the aforementioned comprehensive index and various remote sensing sensitivity factors. After passing through the convolution kernel weights... The convolution operation yields intermediate feature representations: in, This represents the convolution operation. For bias terms, This is a linear rectified activation function. The purpose of this step is to extract the correlation between spatially adjacent grids, such as patterns within a continuous segment that commonly exhibit high salinity, high deformation, or anomalous thermal states.

[0049] Subsequently, for features with explicit time dimensions (such as InSAR deformation time series or multi-temporal surface temperature data), temporal encoding is performed using a Long Short-Term Memory (LSTM) sublayer. The features of each raster at multiple observation times are denoted as... in, For time steps, This represents a single-time-step feature dimension. The recursive process of LSTM can be represented as: in, For a moment The hidden state, For the state of the memory unit, and These represent the hidden state and the memory state from the previous moment, respectively. Through this process, the model can capture the deformation and temperature evolution characteristics of permafrost regions with seasonal and multi-year scale changes, compressing the original time series into a set of time series features with memory.

[0050] To better integrate spatial convolutional features and temporal LSTM features, this embodiment further incorporates a cross-modal attention layer within the shared feature extraction module. This layer utilizes an attention mechanism to perform weighted fusion between different feature types. For example, a simple attention weight can be constructed: in, Indicates the first Convolutional spatial features of each grid cell This represents the LSTM temporal features at the corresponding position. A trainable weight matrix or vector. For unnormalized attention scores, These are the normalized attention weights. The final fused features are: in, This indicates a feature concatenation operation. That is, the first A unified feature representation of each grid. Through this attention fusion mechanism, the model can adaptively balance spatial pattern information with temporal evolution information, thereby more accurately characterizing the overall state of saline permafrost.

[0051] After obtaining a unified feature representation, the model inputs it into a physically consistent gating layer. The core design of this gating layer utilizes a combination of the corrected permafrost composite index, the annual surface temperature variation factor, and the seasonal deformation amplitude factor to generate a weighting factor that modulates the features, thereby physically constraining and renormalizing the unified features. For example, a gating function of the following form can be used: in, For the first The gating weights of each grid cell, For the Sigmoid function, This represents the corrected composite index of frozen soil for this grid. This is the annual variation range factor for surface temperature. The seasonal deformation amplitude factor. These are the weight parameters learned through training. Subsequently, the unified feature representation is renormalized: in, These are features renormalized by the physical consistency gating layer. Intuitively, when a grid cell exhibits significant permafrost characteristics (high comprehensive index, small temperature amplitude, and large seasonal deformation), the gating weight will increase, thereby strengthening the features at that location; conversely, it will suppress audible or anomalous features that do not conform to physical laws.

[0052] In the decoding phase, this embodiment sets up three task decoding heads, respectively targeting the three output tasks of permafrost existence probability, salinization level, and engineering risk level. Each decoding head can employ one to multiple layers of fully connected networks to process the renormalized features. A nonlinear transformation is performed, and the corresponding prediction result is output. For example, a probabilistic decoding head for permafrost can provide: in, For the first The probability that a grid cell belongs to the "permafrost exists" category. and These are the weights and biases of the decoding head for this task. Similarly, the salting level decoding head can output a multi-class probability vector: in, This represents the probability distribution of the grid belonging to different salinization levels (e.g., mild, moderate, severe). and This corresponds to the weight and bias matrix. Similarly, the engineering risk assessment decoding head outputs: in, Give the probability distribution of the engineering risk level (e.g., general, relatively high, high, extremely high) to which the grid belongs.

[0053] During model training, a multi-task joint loss function can be constructed, incorporating the classification losses of the three tasks along with the physical consistency constraint into the optimization objective. For example: in, The cross-entropy loss is a binary classification based on the presence or absence of permafrost. For the loss of multiple categories due to salinization grade, For engineering risk levels, multiple categories of losses are classified. The physical consistency constraint loss is used to penalize situations where there are physical discrepancies between the prediction results and ground temperature, deformation, and comprehensive indices. The weighting coefficients are used. Through this multi-task joint training and physical constraint mechanism, the multi-task deep learning model constructed in this embodiment can not only accurately predict the probability of permafrost existence, salinization level, and engineering risk level in a statistical sense, but also maintain self-consistency in its internal feature flow and output results on the permafrost thermal-salt-deformation mechanism, thereby providing a more reliable intelligent judgment basis for risk identification and engineering decision-making along railway lines.

[0054] In one feasible implementation, such as Figure 6 As shown, step S160 above generates a risk profile of saline permafrost along the railway mileage direction based on the aforementioned engineering risk level, and identifies risk zones along the railway line, including: S1601. Combining the above-mentioned engineering risk level with the above-mentioned digital elevation model and the geometric information of the railway line centerline, the grid cells corresponding to the above-mentioned engineering risk level are subjected to track reprojection processing to obtain the risk mileage sequence along the above-mentioned railway line mileage direction. S1602. Based on the above risk mileage sequence, a risk profile in the mileage direction is constructed according to a preset mileage step size. In the process of construction, the above InSAR linear deformation rate and the above seasonal deformation amplitude factor are integrated to enhance the ability of the risk profile to characterize the sensitivity of frost heave and thaw settlement. S1603. Perform dynamic classification processing on the above risk profile to obtain dynamic classification results; S1604. Generate a risk zoning identification map along the railway line based on the above dynamic grading results, so as to realize the fine zoning of the above-mentioned saline permafrost risk in the direction of railway mileage.

[0055] For example, the process of generating a risk profile of saline permafrost based on the above-mentioned engineering risk level along the railway mileage and identifying risk zones along the railway line aims to transform the risk distribution results of the two-dimensional plane into a one-dimensional "mileage-risk" expression that directly serves railway engineering decision-making, and on this basis, to achieve refined division and spatial positioning of high-risk sections.

[0056] Specifically, the engineering risk level results output by the aforementioned multi-task deep learning model are first combined with the geometric information of the digital elevation model and the railway line centerline: by using the grid data within the railway line centerline and its buffer zone, the spatial relationship between each engineering risk level grid and the line centerline can be determined. Combined with the slope, aspect and other information derived from the digital elevation model, these grids are subjected to track reprojection processing, that is, the risk levels originally distributed on the two-dimensional plane are projected onto the one-dimensional mileage coordinates along the railway line centerline, thereby obtaining a risk mileage sequence arranged along the railway line mileage direction. In this sequence, each mileage position corresponds to the comprehensive risk measure of one or more grids.

[0057] After obtaining the risk mileage sequence, this embodiment further constructs a risk profile along the mileage direction according to a preset mileage step size. The preset mileage step size can be selected according to engineering application requirements, for example, every 50 meters or every 100 meters can be used as a sampling point. The risk levels within this interval are statistically summarized or weighted averaged, and InSAR linear deformation rate and seasonal deformation amplitude factor are introduced for fusion in this process. Specifically, within the segment corresponding to each mileage step size, not only the engineering risk level value of the area is considered, but also the long-term settlement or uplift rate of the segment and the frost heave-thaw settlement amplitude within a one-year cycle are superimposed. By adjusting the weights or constructing enhanced risk indicators, the final risk profile is more numerically sensitive to frost heave and thaw settlement, and can highlight those potentially dangerous segments that are actively deformed but may not have been fully manifested in the original risk level.

[0058] After constructing a continuous risk profile, this embodiment performs dynamic grading processing on the profile to obtain dynamic grading results reflecting the risk changes along the route. Dynamic grading can be based on the statistical distribution, gradient changes, or set multi-level thresholds of the risk profile. For example, risk value intervals can be divided into different levels such as general risk, relatively high risk, and high risk. Combined with the continuous trend of risk value changes along the mileage direction, mileage sections with rapidly rising risks, consistently high values, or frequent fluctuations can be identified. This dynamic grading process avoids the randomness of judging based solely on single-point values ​​and better captures the spatial clustering characteristics and change boundaries of risks along the route.

[0059] Finally, based on the aforementioned hierarchical dynamic results, the risk level of each mileage section is mapped back to its spatial location, generating a risk zoning identification map along the railway line. This risk zoning identification map divides the railway line into multiple risk areas with spatial continuity and hierarchical differences. For example, continuous high-risk sections are marked as key remediation or key monitoring sections, while medium-risk sections are designated as candidate areas for reserve preventive engineering measures.

[0060] Through the above steps, this embodiment realizes the complete transformation from a two-dimensional risk grid to a one-dimensional mileage profile and then to a zoning map along the railway line. This enables the risk of saline permafrost to be expressed and zoned in a refined manner along the railway mileage in an intuitive and easy-to-use way for engineers. It provides targeted decision-making basis for route selection optimization, subgrade structure type selection, and the formulation of monitoring and maintenance strategies during operation.

[0061] In one feasible implementation, such as Figure 7 As shown, the above method also includes: S210. When the risk level of multiple adjacent mileage sections continuously crosses the preset risk change threshold, the above sections are divided into continuous high-risk sections. S220. Based on the above-corrected comprehensive permafrost index, the risk level of the above-mentioned high-risk sections is corrected a second time.

[0062] For example, the above method may further include the step of identifying and secondary correcting continuous high-risk sections based on the risk results in the direction of railway mileage, so as to enhance the sensitivity and reliability of risk zoning to actual engineering hazards.

[0063] Specifically, after constructing the risk profile and performing preliminary dynamic classification along the railway mileage direction, a risk level sequence discretized according to a preset mileage step size can be obtained, with each mileage segment corresponding to a risk level label. To identify dangerous areas that are spatially concentrated in segments rather than scattered points, this embodiment sets a preset risk change threshold and performs a continuity analysis on the risk levels of multiple adjacent mileage segments: when it is detected that the risk level of several adjacent mileage segments is continuously at a high risk or rises from medium-high risk to a high risk level by gradually crossing the preset threshold, it is considered that the risk along the line in that segment shows a significant clustering and accumulation effect in the mileage direction, and this entire segment is divided into a continuous high-risk segment. Through this judgment method based on "continuous crossing threshold", the interference of occasional anomalies at a single mileage point on the overall judgment can be effectively avoided, which is more in line with the railway engineering's focus on "segmental defects" and "segmental hidden dangers".

[0064] After identifying continuous high-risk sections, this embodiment does not directly use them as the final risk level. Instead, it introduces the aforementioned corrected permafrost comprehensive index to perform a secondary correction on these sections. On the one hand, the corrected permafrost comprehensive index has integrated salinity, ice content, surface thermal state, and deformation characteristics, and has undergone physical consistency correction through the permafrost heat-salt coupling relationship. Therefore, it can reflect the stability and degradation trend of permafrost in the section at a deeper level. In specific applications, for mileage units that are classified as continuous high-risk sections but whose comprehensive index is at a relatively moderate level and do not show significant high salinity-high frost or strong thaw settlement risk characteristics, their final risk level can be appropriately lowered to avoid overestimation caused by local noise or short-term deformation disturbances. Conversely, for mileage units in continuous high-risk sections with a significantly high comprehensive index that also meet the high salinity-high frost or strong thaw settlement risk criteria, their risk level can be further raised, and they can be marked as priority treatment or key monitoring sections. Through the above-mentioned secondary correction, this embodiment realizes the refined classification and re-screening of continuous high-risk sections, so that the final risk zoning results not only consider the spatial continuity of risk in the mileage direction, but also take into account the comprehensive judgment of multi-source physical indicators, thereby providing a more robust and reliable quantitative basis for the prevention and control of saline permafrost disasters along railway lines.

[0065] In one feasible implementation, such as Figure 8As shown, step S1602 above constructs a risk profile along the mileage direction based on the risk mileage sequence according to a preset mileage step size, and integrates the InSAR linear deformation rate and the seasonal deformation amplitude factor during the construction process to enhance the characterization ability of the risk profile on the sensitivity of frost heave and thaw settlement, including: S16021. Perform orbital interpolation on the above-mentioned InSAR linear deformation rate and the above-mentioned seasonal deformation amplitude factor to align the above-mentioned deformation factor with the above-mentioned risk mileage sequence in the same mileage coordinate system. S16022. The aligned deformation factor is fused with the above-mentioned risk milestone sequence through a weighted superposition method. The weighted superposition method adaptively adjusts the weights based on the magnitude of the above-mentioned seasonal deformation amplitude factor to highlight the frost heave or thaw settlement response characteristics of the high deformation section. S16023. Based on the fused risk enhancement sequence, a risk profile is constructed according to the preset mileage step size, so that the risk profile can reflect the deformation sensitivity distribution characteristics along the mileage direction of the railway line.

[0066] For example, in this embodiment, when constructing a mileage-direction risk profile based on the aforementioned risk mileage sequence according to a preset mileage step size, the aforementioned InSAR linear deformation rate and the aforementioned seasonal deformation amplitude factor are introduced for fusion. The purpose is to ensure that the risk profile not only reflects the spatial distribution of static engineering risk levels but also sensitively characterizes the dynamic response features of frost heave and thaw settlement along the railway line. Specifically, firstly, for the aforementioned InSAR linear deformation rate and the aforementioned seasonal deformation amplitude factor, track-direction interpolation is performed according to the trajectory of the railway line centerline: on the one hand, based on the geometric orientation of the railway line centerline on the plane, the deformation factor originally given by the satellite line of sight or regular grid coordinates is projected onto a one-dimensional track-direction coordinate with railway mileage as the independent variable; on the other hand, discrete deformation sampling points are interpolated in the track direction so that each preset mileage position can obtain the corresponding linear deformation rate and seasonal deformation amplitude value, thereby aligning the aforementioned deformation factor with the aforementioned risk mileage sequence in the same mileage coordinate system, realizing "one mileage position corresponds to a set of risk-deformation joint features".

[0067] After aligning the deformation factor with the risk mileage sequence, this embodiment fuses the deformation information with the original risk level using a weighted superposition method to construct a risk enhancement sequence. Specifically, for each mileage location, the original risk value is combined with the corresponding InSAR linear deformation rate and seasonal deformation amplitude according to certain weights. These weights are not fixed constants but are adaptively adjusted based on the magnitude of the seasonal deformation amplitude factor. When the seasonal deformation amplitude of a certain mileage segment is small, indicating insignificant frost heave and thaw settlement responses, the weight of the deformation factor can be appropriately reduced during the weighted superposition process to maintain the risk level at that location primarily determined by the original engineering risk level. Conversely, when the seasonal deformation amplitude is large, indicating significant frost heave or thaw settlement behavior in that section within a year, the weight of the deformation factor is increased, numerically "amplifying" the risk enhancement value at that mileage location, thereby highlighting the expression of high-deformation sections on the risk profile. Through this adaptive weight adjustment based on seasonal deformation amplitude, this embodiment can explicitly emphasize potentially hazardous road sections that are actively deformed but may not yet fully exhibit high engineering risk levels.

[0068] Finally, based on the aforementioned fused risk enhancement sequence, a risk profile is constructed according to a preset mileage step size. This involves continuously depicting the enhanced risk values ​​within each step size interval along the railway mileage direction in the form of curves or strips, thereby forming a mileage-direction risk profile that includes both the original risk level information and the characteristics of frost heave and thaw settlement response.

[0069] Compared to profiles constructed solely based on engineering risk levels, the risk profile obtained in this embodiment is more sensitive to deformation anomalies in terms of longitudinal numerical changes. It can clearly reflect which sections along the aforementioned railway line mileage direction are in sensitive areas with high deformation response and significant heat-water-salt coupling effects, thus providing a more refined and reliable basis for subsequent identification of continuous high-risk sections, secondary risk correction, and delineation of key remediation mileage sections.

[0070] In one feasible implementation, such as Figure 9 As shown, step S1604 above generates a risk zoning identification map along the railway line according to the above dynamic classification results, including: S16041. Based on the above risk profile, construct a mileage-based risk grid sequence along the above railway line, and use the slope factor derived from the above digital elevation model to perform slope aspect correction on the above mileage-based risk grid sequence in order to suppress risk jumps caused by abrupt changes in terrain. S16042. Execute a boundary detection algorithm based on a variable threshold on the risk grid sequence after slope aspect correction to extract the partition boundaries between different risk levels. The variable threshold is dynamically adjusted according to the local variation of the seasonal deformation amplitude factor to enhance the response of boundary detection to melting and subsidence sensitive sections. S16043. Generate a spatially continuous risk zone identification map based on the extracted risk zone boundaries to support the selection of structural types and the formulation of engineering treatment strategies for high-risk sections along railway lines.

[0071] For example, a risk zoning identification map along the railway line is generated based on the above dynamic grading results. This aims to further transform the previously obtained mileage-direction risk profile into a more engineering-friendly spatial zoning result, clearly marking continuous sections with different risk levels in geographic space. Specifically, the first step is to construct a mileage-based risk grid sequence along the railway line based on the aforementioned risk profile. The previous step has already formed risk values ​​or risk levels discrete along the mileage direction at preset mileage steps. In this embodiment, these discrete points are "gridized," that is, corresponding risk grid units are constructed at each mileage location, forming a one-dimensional risk grid sequence extending along the line. Based on this, in order to avoid false signals caused by abrupt changes in terrain, this embodiment introduces a slope factor derived from the digital elevation model to perform slope aspect correction on the above-mentioned mileage risk grid sequence: by calculating the slope and aspect information along the route and its adjacent areas, for pseudo-high-risk points that are caused by abrupt changes in slope and local hydrothermal conditions but do not necessarily correspond to actual salinization and permafrost degradation, the values ​​are appropriately smoothed or weakened, thereby suppressing risk jumps caused by abrupt changes in terrain such as extremely steep slopes, cliffs or deep gullies, and making the risk distribution more consistent with the actual understanding of segmental defects in engineering.

[0072] After completing the aspect correction, this embodiment executes a variable threshold-based boundary detection algorithm on the aspect-corrected risk raster sequence to extract the partition boundaries between different risk levels.

[0073] Unlike traditional methods that use a fixed threshold for simple segmentation, the variable threshold in this embodiment is not a uniform constant, but rather varies with the railway mileage. Its magnitude is dynamically adjusted based on the local variation of the seasonal deformation amplitude factor: in mileage sections with large seasonal deformation amplitudes, indicating strong frost heave and thaw subsidence activity, the threshold can be appropriately lowered, making it easier to identify independent high-risk zone boundaries even with slight increases in risk within these sections; while in sections with relatively stable deformation, the threshold is relatively higher to avoid frequently "slicing" risk zones due to minor fluctuations. Through this dynamic threshold mechanism coupled with InSAR seasonal deformation amplitude, the boundary detection algorithm exhibits stronger responsiveness to thaw-sensitive sections, thereby more accurately identifying mileage intervals that are truly under strong frost-thaw activity and potential structural damage risks in space.

[0074] Finally, based on the risk zoning boundaries extracted above, this embodiment spatially reorganizes the risk level intervals to generate a spatially continuous risk zoning identification map. Specifically, mileage segments along the railway mileage direction that are classified as having the same risk level and are spatially continuous or nearly continuous are aggregated into risk zoning units, and these units are backfilled and mapped onto the actual geographic coordinate system and railway line alignment to form an intuitive spatial distribution layer. In this layer, high-risk zones, relatively high-risk zones, and general-risk zones are clearly visible, allowing engineers to directly select different structural types. For example, in continuous high-risk zones, bridges can be prioritized for road replacement, combined insulation and desalination measures can be strengthened, or enhanced monitoring sections can be set up. Simultaneously, differentiated engineering treatment strategies and long-term operation and maintenance plans can be formulated accordingly, such as increasing the frequency of settlement and deformation monitoring in certain high-risk zones, and implementing preventative reinforcement or drainage projects in medium-risk zones.

[0075] Through this implementation, the present invention not only improves the accuracy and physical reliability of identifying the risk of saline permafrost, but also enhances the overall accuracy of the invention.

[0076] The second aspect, such as Figure 10 As shown, the present invention also proposes a remote sensing identification and control system for railways in saline-alkali permafrost, used to execute the remote sensing identification method for railways in saline-alkali permafrost as described in any of the first aspects, comprising: The acquisition unit 21 is used to acquire multi-source remote sensing and monitoring data within a preset buffer zone centered on the railway line. The multi-source remote sensing and monitoring data includes multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, InSAR deformation time series data, digital elevation model, ground temperature monitoring and borehole salinity profile data. Construction unit 22 is used to perform unified coordinate system registration, orthorectification and spatial resampling on the above-mentioned multi-source remote sensing and monitoring data in order to construct a multi-source data cube for the railway corridor; Extraction unit 23 is used to extract remote sensing sensitive factors characterizing salinity, ice content, thermal state and surface deformation based on the above-mentioned multi-source data cube of the railway corridor. The remote sensing sensitive factors include spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation amplitude factor, InSAR linear deformation rate and seasonal deformation amplitude factor. The correction unit 24 is used to construct a comprehensive index of salinized permafrost using the above-mentioned remote sensing sensitive factors, and to perform physical constraint correction on the comprehensive index based on the heat-salt coupling relationship of permafrost to obtain the corrected permafrost comprehensive index. Output unit 25 is used to input the above-corrected permafrost comprehensive index and the above-mentioned remote sensing sensitive factors into the multi-task deep learning model, and output the probability of permafrost existence, salinization level and engineering risk level respectively. The identification unit 26 is used to generate a risk profile of saline permafrost along the railway mileage based on the above-mentioned engineering risk level, and to identify risk zones along the railway line.

[0077] The remote sensing identification and control system for railways in saline permafrost proposed in this invention can also perform the method described in any of the first aspects.

[0078] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A remote sensing identification method for railways in saline-alkali permafrost, characterized in that, include: Acquire multi-source remote sensing and monitoring data within a pre-defined buffer zone centered on the railway line. The multi-source remote sensing and monitoring data includes multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, InSAR deformation time series data, digital elevation model, ground temperature monitoring, and borehole salinity profile data. The multi-source remote sensing and monitoring data are registered in a unified coordinate system, orthorectified, and spatially resampled to construct a multi-source data cube for the railway corridor. Based on the multi-source data cube of the railway corridor, remote sensing sensitive factors characterizing salinity, ice content, thermal state and surface deformation are extracted. The remote sensing sensitive factors include spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation amplitude factor, InSAR linear deformation rate and seasonal deformation amplitude factor. A comprehensive index of salinized permafrost was constructed using the remote sensing sensitive factors, and the comprehensive index was physically constrained and corrected based on the heat-salt coupling relationship of permafrost to obtain the corrected permafrost comprehensive index. The corrected permafrost comprehensive index and the remote sensing sensitive factor are input into a multi-task deep learning model, which outputs the probability of permafrost existence, salinization level and engineering risk level, respectively. Based on the aforementioned engineering risk level, a risk profile of saline permafrost is generated along the railway mileage, and risk zones are identified along the railway line.

2. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 1, characterized in that, The process of performing unified coordinate system registration, orthorectification, and spatial resampling on the multi-source remote sensing and monitoring data to construct a multi-source data cube for the railway corridor includes: For the multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data and InSAR deformation time series data in the multi-source remote sensing and monitoring data, unified coordinate system registration is performed based on the ground control points of the railway line, and terrain correction and orthorectification are performed using the digital elevation model. Spatial resampling is performed on the registered and corrected image data according to the preset spatial resolution to obtain a remote sensing image dataset with a unified spatial grid. The geothermal monitoring data and borehole salt profile data are mapped to the corresponding grid cells in the unified spatial grid by spatial interpolation or along-line projection. Using the railway line centerline as a reference, all resampled grid data are indexed according to time sequence and mileage direction to construct a multi-source data cube for railway corridors with a unified spatial grid, unified time step and railway mileage coordinate system.

3. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 1, characterized in that, The remote sensing sensitive factors characterizing salinity, ice content, thermal state, and surface deformation features based on the multi-source data cube of the railway corridor include: The spectral salinity index is calculated based on the multispectral or hyperspectral images to characterize the salt distribution of the surface soil along the railway line. The radar scattering coefficient combination factor is extracted based on the synthetic aperture radar image to characterize the ice content and water content of the active layer along the railway line. The surface temperature sequence is retrieved based on the thermal infrared data and the annual variation factor of surface temperature is calculated to characterize the surface thermal state along the railway line. Based on the InSAR deformation time series data, the InSAR linear deformation rate and seasonal deformation amplitude factor are obtained to characterize the frost heave and thaw settlement deformation characteristics along the railway line.

4. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 1, characterized in that, The process involves constructing a comprehensive index for salinized permafrost using the remote sensing sensitive factors, and then applying physical constraints to the comprehensive index based on the heat-salt coupling relationship of permafrost to obtain the corrected comprehensive permafrost index, including: Based on the spectral salinity index, the radar scattering coefficient combination factor, the annual variation factor of surface temperature, the InSAR linear deformation rate, and the seasonal deformation amplitude factor, a comprehensive index of salinized permafrost is constructed using a multiple regression method. The comprehensive index of salinized permafrost is physically consistent based on the thermal-salt coupling relationship of permafrost, and the corrected comprehensive index of salinized permafrost is output. When the annual variation factor of surface temperature is lower than a preset temperature threshold and the spectral salinity index is higher than a preset salinity threshold, the comprehensive index is corrected to a high-salt-high-freezing state. When the surface temperature is close to zero degrees Celsius and the seasonal deformation amplitude factor exceeds a preset deformation threshold, the comprehensive index is corrected to a state of strong melt subsidence risk.

5. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 1, characterized in that, The multi-task deep learning model includes a shared feature extraction module, a physical consistency gating layer, and three task decoding heads. The process involves inputting the corrected permafrost comprehensive index and the remote sensing sensitive factor into a multi-task deep learning model, which outputs the probability of permafrost existence, salinization level, and engineering risk level, respectively, including: The corrected permafrost comprehensive index, the spectral salinity index, the radar scattering coefficient combination factor, the annual surface temperature variation factor, the InSAR linear deformation rate, and the seasonal deformation amplitude factor are input into the shared feature extraction module. The shared feature extraction module jointly encodes the spatial features and temporal features through a convolutional neural network sub-layer, a long short-term memory network sub-layer, and a cross-modal attention layer to obtain the unified feature representation. The unified feature representation is input into the physical consistency gating layer, which generates consistency control weights based on the corrected permafrost comprehensive index, the annual surface temperature variation amplitude factor, and the seasonal deformation amplitude factor, and performs physical constraint renormalization on the unified feature representation. The features, after being renormalized by the physical consistency gating layer, are respectively input into the permafrost identification decoding head, the salinization level classification decoding head, and the engineering risk assessment decoding head to generate the permafrost existence probability, the salinization level, and the engineering risk level, respectively.

6. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 1, characterized in that, The process of generating a risk profile of saline permafrost along the railway mileage direction based on the engineering risk level, and identifying risk zones along the railway line, includes: By combining the engineering risk level with the digital elevation model and the geometric information of the railway line centerline, the grid cells corresponding to the engineering risk level are subjected to track reprojection processing to obtain a risk mileage sequence along the mileage direction of the railway line. Based on the risk mileage sequence, a risk profile in the mileage direction is constructed according to a preset mileage step size. During the construction process, the InSAR linear deformation rate and the seasonal deformation amplitude factor are integrated to enhance the ability of the risk profile to characterize the sensitivity of frost heave and thaw settlement. Perform dynamic classification processing on the risk profile to obtain dynamic classification results; Based on the dynamic grading results, a risk zoning identification map along the railway line is generated, realizing the refined zoning of the risk of saline permafrost in the direction of railway mileage.

7. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 6, characterized in that, The method further includes: When the risk level of multiple adjacent mileage sections continuously crosses a preset risk change threshold, the section is divided into a continuous high-risk section. The risk level of the high-risk section is then corrected a second time based on the corrected comprehensive permafrost index.

8. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 6, characterized in that, The construction of a risk profile along the mileage direction based on the risk mileage sequence according to a preset mileage step size, and the integration of the InSAR linear deformation rate and the seasonal deformation amplitude factor during the construction process to enhance the characterization ability of the risk profile on the sensitivity of frost heave and thaw settlement, includes: The InSAR linear deformation rate and the seasonal deformation amplitude factor are subjected to orbital interpolation to align the deformation factor with the risk mileage sequence in the same mileage coordinate system. The aligned deformation factor is fused with the risk milestone sequence through a weighted superposition method, wherein the weighted superposition method adaptively adjusts the weight based on the magnitude of the seasonal deformation amplitude factor to highlight the frost heave or thaw settlement response characteristics of the high deformation segment. Based on the fused risk enhancement sequence, a risk profile is constructed according to the preset mileage step size, so that the risk profile can reflect the deformation sensitivity distribution characteristics along the mileage direction of the railway line.

9. The remote sensing identification method for railways in saline-alkali permafrost as described in claim 6, characterized in that, The step of generating a risk zoning identification map along the railway line based on the dynamic grading results includes: Based on the risk profile, a mileage-based risk grid sequence along the railway line is constructed, and the slope factor derived from the digital elevation model is used to perform slope aspect correction on the mileage-based risk grid sequence to suppress risk jumps caused by abrupt terrain changes. A variable threshold-based boundary detection algorithm is executed on the slope-corrected risk grid sequence to extract the partition boundaries between different risk levels. The variable threshold is dynamically adjusted according to the local variation of the seasonal deformation amplitude factor to enhance the response of boundary detection to melt-settlement sensitive sections. Based on the extracted risk zoning boundaries, a spatially continuous risk zoning identification map is generated to support the selection of structural types and the formulation of engineering treatment strategies for high-risk sections along railway lines.

10. A remote sensing identification system for railways in saline-permafrost areas, used to execute the remote sensing identification method for railways in saline-permafrost areas as described in any one of claims 1 to 9, characterized in that, include: The acquisition unit is used to acquire multi-source remote sensing and monitoring data within a preset buffer zone centered on the railway line. The multi-source remote sensing and monitoring data includes multispectral or hyperspectral images, synthetic aperture radar images, thermal infrared data, InSAR deformation time series data, digital elevation models, ground temperature monitoring and borehole salinity profile data. The construction unit is used to perform unified coordinate system registration, orthorectification and spatial resampling on the multi-source remote sensing and monitoring data to construct a multi-source data cube for the railway corridor; The extraction unit is used to extract remote sensing sensitive factors characterizing salinity, ice content, thermal state and surface deformation features based on the multi-source data cube of the railway corridor. The remote sensing sensitive factors include spectral salinity index, radar scattering coefficient combination factor, annual surface temperature variation amplitude factor, InSAR linear deformation rate and seasonal deformation amplitude factor. The correction unit is used to construct a comprehensive index of salinized permafrost using the remote sensing sensitive factors, and to perform physical constraint correction on the comprehensive index based on the heat-salt coupling relationship of permafrost to obtain the corrected permafrost comprehensive index. The output unit is used to input the corrected permafrost comprehensive index and the remote sensing sensitive factor into the multi-task deep learning model, and output the probability of permafrost existence, salinization level and engineering risk level, respectively. The identification unit is used to generate a risk profile of saline permafrost along the railway mileage direction based on the engineering risk level, and to identify risk zones along the railway line.

Images