Methods, devices, equipment, and storage media for early warning of land subsidence based on synthetic aperture radar

Abstract

This application relates to the technical field of land subsidence early warning, and particularly to a method, apparatus, device, and storage medium for land subsidence early warning based on synthetic aperture radar (SAR). It includes: acquiring SAR images of a target detection area from the same satellite at different time points and generating time-series deformation data; fusing multi-source auxiliary data with the time-series deformation data to obtain a multi-source data feature set; dividing the multi-source data feature set into a training feature set, a validation feature set, and an input feature set; training and validating a land subsidence prediction model using the training and validation feature sets; inputting the input feature set into the land subsidence prediction model to output subsidence prediction data; updating the subsidence early warning threshold based on the subsidence prediction data and real-time hydrological data; and generating land subsidence early warning information for the target detection area based on the subsidence early warning threshold and the subsidence prediction data. This application improves the reliability and practicality of land subsidence early warning systems.

Description

Technical Field

[0001] This application relates to the technical field of land subsidence early warning, and in particular to a land subsidence early warning method, device, equipment and storage medium based on synthetic aperture radar. Background Technology

[0002] As a typical slowly changing geological hazard, land subsidence's dynamic evolution directly threatens the safety of urban buildings, the stability of underground pipelines, and the balance of the ecological environment. Therefore, real-time early warning and risk management of subsidence trends are of significant engineering importance. With the increasing urbanization and underground resource development, the spatiotemporal differences in subsidence rates are becoming more pronounced. Traditional static threshold early warning models are no longer sufficient to meet the dynamic monitoring needs in complex geological environments, necessitating the development of an early warning mechanism that can respond to real-time data and environmental changes.

[0003] In existing technologies, InSAR (Interferometric Synthetic Aperture Radar) technology, especially TS-InSAR (Time Series Interferometric Synthetic Aperture Radar), has achieved millimeter-level deformation monitoring accuracy through multi-period image analysis, providing a data foundation for capturing the dynamic evolution of subsidence. In the early warning stage, current mainstream methods mostly rely on preset fixed thresholds based on historical subsidence data, triggering alarms by comparing real-time monitoring values ​​with these thresholds. For example, an early warning message is automatically issued when the cumulative subsidence in a certain area exceeds 50 mm.

[0004] However, existing early warning mechanisms have significant limitations: fixed threshold settings do not consider the dynamic changes in subsidence rates and the real-time impact of environmental factors. On the one hand, differences in geological conditions and the spatiotemporal fluctuations in human activity intensity can cause subsidence rates to exhibit nonlinear characteristics, meaning fixed thresholds may provide delayed warnings during periods of accelerated subsidence or over-warn during periods of deceleration. On the other hand, a single threshold cannot adapt to the risk tolerance of different areas (such as the difference in warning sensitivity requirements between urban centers and suburbs), resulting in insufficient accuracy and practicality of early warnings. This mismatch between static thresholds and dynamic subsidence processes directly reduces the early warning system's ability to proactively manage disaster risks. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this application provides a method, device, equipment, and storage medium for land subsidence early warning based on synthetic aperture radar. By constructing a land subsidence early warning method that integrates multi-source data, physical mechanisms, and dynamic decision-making, it effectively overcomes the limitations of static threshold early warning mechanisms in the prior art, thereby improving the reliability and practicality of land subsidence early warning.

[0006] The technical solution adopted by this application to solve its technical problem is:

[0007] In a first aspect, this application provides a method for early warning of land subsidence based on synthetic aperture radar, the method comprising:

[0008] Acquire synthetic aperture radar images of the target detection area from the same satellite at different time points, and generate time-series deformation data based on all the synthetic aperture radar images;

[0009] Acquire multi-source auxiliary data, fuse the multi-source auxiliary data with the time-series deformation data, perform feature construction on the fused data, and obtain a multi-source data feature set;

[0010] The multi-source data feature set is divided into a training feature set, a validation feature set, and an input feature set. The training feature set and the validation feature set are input into the model to be trained, and the surface subsidence prediction model is trained and validated. The model to be trained is based on Terzaghi's effective stress principle as a physical constraint.

[0011] The input feature set is input into the surface subsidence prediction model, and the surface subsidence prediction model outputs subsidence prediction data.

[0012] Acquire real-time hydrological data, and update the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data;

[0013] Based on the subsidence warning threshold and the subsidence prediction data, surface subsidence warning information for the target detection area is generated.

[0014] Optionally, the step of acquiring synthetic aperture radar images of the target detection area from the same satellite at different time points, and generating time-series deformation data based on all the synthetic aperture radar images, includes:

[0015] All the synthetic aperture radar images are sorted by time to obtain an image time series;

[0016] The image time series is preprocessed, and based on any two images in the preprocessed image time series that are separated by a preset time interval, a deformation profile of the target detection area is obtained.

[0017] Local deformation regions are extracted from the deformation profile, and all synthetic aperture radar images are cropped based on the local deformation regions.

[0018] Based on all the cropped synthetic aperture radar images, a time-series detail map of the local deformation region is generated, and the time-series deformation data is extracted from the time-series detail map.

[0019] Optionally, the multi-source auxiliary data includes environmental auxiliary data, remote sensing auxiliary data, and historical auxiliary data;

[0020] The step of fusing the multi-source auxiliary data with the time-series deformation data, and performing feature construction on the fused data to obtain a multi-source data feature set includes:

[0021] The multi-source auxiliary data and the temporal deformation data are spatiotemporally aligned to a preset spatiotemporal resolution.

[0022] Deformation features are extracted based on the aligned temporal deformation data, environmental features and geological category features are extracted based on the aligned environmental auxiliary data, surface ecological features are extracted based on the aligned remote sensing auxiliary data, and historical case features are extracted based on the aligned historical auxiliary data.

[0023] The geological category features are encoded, and the deformation features, environmental features, surface ecological features, historical case features, and encoded geological category features are integrated into the multi-source data feature set.

[0024] Optionally, the step of dividing the multi-source data feature set into a training feature set, a validation feature set, and an input feature set, and inputting the training feature set and the validation feature set into the model to be trained, to train and validate the land subsidence prediction model includes:

[0025] The multi-source data feature set is divided into three consecutive time periods in chronological order: a first time period, a second time period, and a third time period, which are respectively used as the training feature set, the verification feature set, and the input feature set.

[0026] The training feature set and the validation feature set are input into the model to be trained. The model is trained using the training feature set, and the validation feature set is used to determine whether the iteration conditions are met. The iteration conditions are met until the iteration conditions are met, and the land subsidence prediction model is obtained.

[0027] Optionally, the step of inputting the input feature set into the land subsidence prediction model and outputting subsidence prediction data through the land subsidence prediction model includes:

[0028] The input feature set is input into the land subsidence prediction model, and the land subsidence prediction model generates the subsidence prediction value and the corresponding confidence interval for a preset future time period.

[0029] The predicted settlement values ​​and their corresponding confidence intervals are integrated into the predicted settlement data output.

[0030] Optionally, the step of acquiring real-time hydrological data and updating the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data includes:

[0031] The settlement prediction data and the real-time hydrological data are input into a preset reinforcement learning agent, and the reinforcement learning agent is used to match the optimal threshold adjustment strategy under the current data background.

[0032] The original early warning threshold is updated according to the threshold adjustment strategy to generate the settlement early warning threshold.

[0033] Optionally, after the step of generating surface subsidence early warning information for the target detection area based on the subsidence early warning threshold and the subsidence prediction data, the method further includes:

[0034] Obtain the actual settlement monitoring data of the target detection area;

[0035] The actual settlement monitoring data is compared with the corresponding surface settlement early warning information to obtain the prediction deviation value;

[0036] The prediction bias value is input into the reinforcement learning agent to continuously perform optimization updates.

[0037] Secondly, this application provides a land subsidence early warning device based on synthetic aperture radar, comprising:

[0038] The deformation data acquisition module is used to acquire synthetic aperture radar images of the target detection area from the same satellite at different time points, and generate time-series deformation data based on all the synthetic aperture radar images.

[0039] A multi-source feature construction module is used to acquire multi-source auxiliary data, fuse the multi-source auxiliary data with the time-series deformation data, perform feature construction on the fused data, and obtain a multi-source data feature set.

[0040] The prediction model training module is used to divide the multi-source data feature set into a training feature set, a validation feature set, and an input feature set, and input the training feature set and the validation feature set into the model to be trained, thereby training and validating the land subsidence prediction model; the model to be trained is based on Terzaghi's effective stress principle as a physical constraint.

[0041] The settlement prediction generation module is used to input the input feature set into the surface settlement prediction model and output settlement prediction data through the surface settlement prediction model.

[0042] The early warning threshold update module is used to acquire real-time hydrological data and update the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data.

[0043] The settlement early warning output module is used to generate surface settlement early warning information for the target detection area based on the settlement early warning threshold and the settlement prediction data.

[0044] Thirdly, this application provides an electronic device, comprising:

[0045] One or more processors;

[0046] One or more memory units;

[0047] And one or more computer programs, wherein the one or more computer programs are stored in the one or more memories, and the one or more computer programs include instructions that, when executed by the one or more processors, cause the electronic device to perform the methods described above.

[0048] Fourthly, this application provides a computer-readable storage medium storing a program or instructions that, when executed, implement the above-described method.

[0049] The beneficial effects of this application are as follows: First, high-precision temporal deformation data is acquired through synthetic aperture radar imagery, thus providing a data foundation for dynamic analysis. Subsequently, multi-source auxiliary data is introduced and feature fusion is performed, overcoming the early warning bias caused by traditional methods neglecting dynamic environmental changes. In the model building stage, the Terzaghi effective stress principle is introduced as a physical constraint, making the prediction model not only data-driven but also more consistent with geomechanical laws, significantly improving the scientific validity and reliability of the prediction results.

[0050] Subsequently, a reinforcement learning agent mechanism is adopted to dynamically adjust the early warning threshold based on real-time hydrological data and forecast results. This mechanism replaces the original static threshold setting method, enabling the system to respond to differences in risk sensitivity across different regions and the non-stationary characteristics of the settlement process, thereby avoiding delayed and over-warning issues. By continuously learning from forecast deviations and optimizing decision-making strategies, the system ultimately achieves proactive risk management capabilities, significantly improving the accuracy and practicality of early warnings. Attached Figure Description

[0051] Figure 1 This is a schematic flowchart of the land subsidence early warning method based on synthetic aperture radar provided in the embodiments of this application;

[0052] Figure 2 This is a virtual structural diagram of the land subsidence early warning device based on synthetic aperture radar provided in this application;

[0053] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0054] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0055] The following will clearly and completely describe the concept, specific structure, and resulting technical effects of this application in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of this application. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are all within the scope of protection of this application. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this application can be combined interactively without contradicting each other.

[0056] Reference Figure 1 , Figure 1 This is a flowchart illustrating the land subsidence early warning method based on synthetic aperture radar provided in this application embodiment. The flowchart shows several key steps of the dynamic land subsidence early warning scheme based on synthetic aperture radar imagery acquired by synthetic aperture radar, as provided in this application embodiment. These steps are described in detail below:

[0057] In step S1, synthetic aperture radar images of the same satellite at different time points of the target detection area are acquired, and time-series deformation data is generated based on all the synthetic aperture radar images.

[0058] Among them, synthetic aperture radar imagery refers to remote sensing images of the target detection area taken by synthetic aperture radar satellites in repeating orbit wide swath mode. It has all-weather and all-time imaging capabilities, and its observation cycle is fixed and positively correlated with the time span. Repeating orbit wide swath mode refers to the wide coverage imaging mode adopted when the satellite passes over the target detection area multiple times on the same orbital plane, ensuring that the images have overlapping spatial coverage and consistent observation geometry, providing a data basis for interferometry.

[0059] Among them, the temporal deformation data is the time series data of surface deformation obtained by performing temporal interferometry processing on multi-period synthetic aperture radar images. It contains dynamic change information such as subsidence amount, subsidence rate and acceleration with millimeter-level precision, which can reflect the long-term evolution law of surface subsidence.

[0060] Specifically, by analyzing the time series of repeated orbit images from the same satellite, the consistency of image observation conditions is ensured, reducing systematic errors caused by differences in sensors from different satellites, and providing a stable data foundation for deformation monitoring. In this embodiment, the step of acquiring synthetic aperture radar images of the target detection area from the same satellite at different time points and generating time series deformation data based on all the synthetic aperture radar images includes:

[0061] All the synthetic aperture radar images are sorted by time to obtain an image time series;

[0062] Among them, the image time series refers to the sequence formed by arranging synthetic aperture radar images acquired by the same satellite at different time points in chronological order. It must meet the following requirements: fixed observation period, time span greater than 1 year, total number of images not less than 18, and avoid interference periods such as snow cover.

[0063] The image time series is preprocessed, and based on any two images in the preprocessed image time series that are separated by a preset time interval, a deformation profile of the target detection area is obtained.

[0064] Among them, the deformation profile is a schematic diagram of the deformation distribution of the target detection area obtained by processing any two images with a time interval of 1 year using differential synthetic aperture radar interferometry technology, which is used to quickly identify the range of potential deformation areas.

[0065] Local deformation regions are extracted from the deformation profile, and all synthetic aperture radar images are cropped based on the local deformation regions.

[0066] Among them, the local deformation region is a rectangular area that completely covers the deformation area extracted from the deformation overview map, with an area of ​​not less than 1km×1km, which is the spatial range for subsequent fine processing.

[0067] Based on all the cropped synthetic aperture radar images, a time-series detail map of the local deformation region is generated, and the time-series deformation data is extracted from the time-series detail map.

[0068] Among them, the time series detail map is a high-precision deformation data obtained by processing the cropped local deformation area image with time series synthetic aperture radar interferometry technology (such as permanent scatterer or small baseline set interferometry), which contains millimeter-level precision information on the change of settlement over time.

[0069] Specifically, firstly, wide-swath synthetic aperture radar images of the same satellite in repeating orbits are collected and sorted chronologically to form an image time series, ensuring that the image time span is positively correlated with the observation period to cover the complete deformation cycle. Preprocessing is then performed on the image series, including orbital fine correction (eliminating orbital errors using precise orbital information), atmospheric delay correction (subtracting atmospheric phase delay based on meteorological models or permanent scatterer points), and phase unwrapping (restoring the continuous phase field) to improve the quality of the original data.

[0070] Subsequently, two images with a one-year time interval were selected, and centimeter-level deformation maps were generated using differential synthetic aperture radar interferometry to quickly locate the approximate range of surface deformation. Local deformation areas were extracted from the map, and the entire image was cropped according to rectangular boundaries to reduce data redundancy in non-deformation areas. Finally, time-series synthetic aperture radar interferometry was used to perform precise calculations on the cropped images. By selecting high-coherence points (such as permanent scatterers), a deformation time-series model was constructed to generate detailed time-series maps of local deformation areas, from which time-series deformation data containing dynamic indicators such as subsidence rate and cumulative subsidence were extracted.

[0071] More specifically, by sequencing and preprocessing the same satellite image sequence over time, consistency of observation conditions is ensured, reducing systematic errors caused by differences between different sensors and providing a stable data foundation for deformation monitoring. Rapid generation of deformation profile maps enables deformation screening over large areas, avoiding the computational waste of performing high-precision processing on the entire region directly; cropping of local deformation areas focuses computation on key regions, significantly improving computational efficiency. The millimeter-level accuracy and time-series analysis capabilities of the detailed time-series maps can capture the dynamic evolution of land subsidence, providing high-precision basic data support for subsequent dynamic early warning.

[0072] For example, in monitoring surface subsidence in a certain area, 24 wide-swath synthetic aperture radar (SAR) images from the Sentinel-1A satellite were collected, with an observation period of 12 days and a time span of 3 years. The images were sorted chronologically to avoid winter snow accumulation periods. The image sequence underwent preprocessing: precise orbital data was used to control orbital errors to the centimeter level; atmospheric delay correction was performed based on meteorological model data; and continuous phase field was recovered using least-squares phase unwrapping. Deformation maps were generated from images from 2020 and 2021, identifying local deformation areas of 2km × 2km, from which all 24 images were cropped. The cropped images were then processed using permanent scatterer SAR interferometry (PSA), selecting 1000 permanent scatterer points with coherence greater than 0.9 to generate a detailed time-series map. From this map, temporal deformation data for the area over the past 3 years was extracted, including dynamic changes in cumulative subsidence from 5mm to 32mm and the average annual subsidence rate from 1.7mm / month to 2.8mm / month.

[0073] In step S2, multi-source auxiliary data is acquired, the multi-source auxiliary data is fused with the time-series deformation data, and feature construction is performed on the fused data to obtain a multi-source data feature set.

[0074] Among them, multi-source auxiliary data refers to non-image data that drives and constrains land subsidence, including environmental auxiliary data (groundwater level, daily rainfall, geological parameters such as clay layer thickness and compression modulus), satellite auxiliary data (NDVI vegetation cover, surface temperature, etc.) and historical auxiliary data (location, scale and triggering factors of historical subsidence events in the region), which are key inputs to supplement the physical mechanisms of time-series deformation data.

[0075] Specifically, firstly, based on the acquired temporal deformation data, multi-source auxiliary data are collected simultaneously; then, data fusion and spatiotemporal unification are performed, followed by feature construction to extract key indicators from the fused data, ultimately forming a multi-source data feature set containing multi-dimensional indicators, providing comprehensive input for subsequent model training. In this step, multi-source auxiliary data, by integrating multi-dimensional auxiliary data, lays the foundation for high-precision settlement prediction.

[0076] More specifically, the step of fusing the multi-source auxiliary data with the temporal deformation data, and performing feature construction on the fused data to obtain a multi-source data feature set includes:

[0077] The multi-source auxiliary data and the temporal deformation data are spatiotemporally aligned to a preset spatiotemporal resolution.

[0078] The multi-source auxiliary data includes environmental auxiliary data, remote sensing auxiliary data, and historical auxiliary data. Environmental auxiliary data are key physical factors driving land subsidence, including daily monitored groundwater levels, daily rainfall, and geological parameters (such as clay layer thickness and compression modulus), directly reflecting changes in effective soil stress and the impact of hydrogeological conditions on subsidence. Remote sensing auxiliary data are surface environmental characteristics acquired through satellite remote sensing technology, including NDVI vegetation cover and surface temperature, indirectly reflecting the moderating effect of surface ecological changes on subsidence. Historical auxiliary data are records of historical subsidence, landslides, and other geological disasters in the region, including event location, scale, and triggering factors (such as extreme rainfall and groundwater over-extraction), used for risk pattern identification and validation in model training.

[0079] Among them, the spatiotemporal alignment process unifies multi-source auxiliary data and time-series deformation data to the same spatial grid and time scale. Spatially, it resamples to a unified spatial resolution, and temporally, it aggregates to monthly data (high-frequency data such as daily rainfall are statistically analyzed monthly), ensuring the consistency of data in the spatiotemporal dimensions. The preset spatiotemporal resolution is a standardized scale for the fused data, with a spatial resolution of 1km×1km and a temporal resolution of monthly, balancing monitoring accuracy and computational efficiency, and adapting to the subsidence prediction needs of large-scale areas.

[0080] Deformation features are extracted based on the aligned temporal deformation data, environmental features and geological category features are extracted based on the aligned environmental auxiliary data, surface ecological features are extracted based on the aligned remote sensing auxiliary data, and historical case features are extracted based on the aligned historical auxiliary data.

[0081] Among them, environmental features are physical driving indicators constructed from environmental auxiliary data, such as groundwater level fluctuations and rainfall concentration, which quantify the intensity of the impact of hydrogeological conditions on subsidence; surface ecological features are environmental regulation indicators extracted from remote sensing auxiliary data, such as NDVI vegetation coverage and surface temperature inversion values, which reflect the indirect effect of surface ecosystems on subsidence; historical case features are risk pattern indicators extracted from historical auxiliary data, such as the frequency of historical subsidence events and triggering rainfall thresholds, which are used for model training to identify similar scenarios.

[0082] The geological category features are encoded, and the deformation features, environmental features, surface ecological features, historical case features, and encoded geological category features are integrated into the multi-source data feature set.

[0083] Among them, geological category characteristics are classification data that characterize the geological attributes of the region, such as lithology (clay layer, sand layer) and soil compression modulus, which need to be converted into numerical features that the model can recognize through coding.

[0084] Specifically, the temporal deformation data and the three types of auxiliary data are first aligned in time and space. Spatially, they are unified to 1km×1km grid cells to ensure spatial matching of data from different sources. Temporally, they are unified to monthly data. High-frequency environmental data (such as daily rainfall) are aggregated monthly, and low-frequency geological data (such as clay layer thickness) are filled with time dimension to eliminate data misalignment caused by differences in temporal and spatial scales.

[0085] Subsequently, the system calculates dynamic indicators such as cumulative subsidence and subsidence rate from time-series deformation data to capture the temporal evolution trend of subsidence; it constructs physical driving characteristics such as groundwater level fluctuation and rainfall concentration from environmental auxiliary data to quantify the impact of hydrogeological conditions; it extracts ecological characteristics such as NDVI and surface temperature from remote sensing auxiliary data to reflect the regulatory role of the surface environment on subsidence; it extracts case features such as event frequency and trigger threshold from historical auxiliary data to help the model identify risk patterns; and it performs One-Hot encoding on geological category features (such as lithology) to convert the categorical data into numerical vectors.

[0086] Finally, all features are integrated into a multi-source data feature set, forming a multi-dimensional input covering deformation dynamics, physical drivers, ecological regulation, and historical risks, providing comprehensive data support for subsequent training of the physical constraint model. Through the classification and fusion of multi-source auxiliary data, environmental auxiliary data enables the model to learn geomechanical laws; remote sensing auxiliary data introduces the correlation between surface ecology and subsidence (e.g., drought leading to increased surface temperature and exacerbated soil shrinkage), enhancing the model's response to indirect influencing factors; historical auxiliary data improves the model's generalization ability through risk pattern recognition, avoiding misjudgments of new scenarios. Furthermore, spatiotemporal alignment ensures data consistency, balances local details with computational efficiency, and reduces data redundancy.

[0087] For example, in the monitoring of surface subsidence in a certain city, the time-series deformation data are detailed time-series maps of subsidence at the millimeter level obtained based on PS-InSAR technology from 2020 to 2023, with a spatial resolution of 10m×10m and a temporal resolution of 12 days; the multi-source auxiliary data include: environmental auxiliary data (daily groundwater level monitoring data, daily rainfall, clay layer thickness of 20-50m, compression modulus of 15-30MPa), remote sensing auxiliary data (Landsat 8 monthly NDVI, surface temperature products), and historical auxiliary data (records of 3 subsidence events from 2015 to 2019, triggered by groundwater over-extraction and extreme rainfall).

[0088] During the spatiotemporal alignment phase, all data were resampled to a 1km×1km grid, and the time was unified to monthly. The groundwater level was calculated using the monthly average value, the daily rainfall was aggregated into the total monthly rainfall and the rainfall concentration was calculated, and the NDVI and surface temperature were calculated using the monthly maximum values. In the geological category characteristics, the clay layer and sand layer were coded as [1,0] and [0,1], respectively.

[0089] In the feature extraction stage, cumulative settlement, annual settlement rate, and acceleration are obtained from time-series deformation data; groundwater level fluctuations (monthly range 2-8m) and rainfall concentration are obtained from environmental data; NDVI and surface temperature are obtained from remote sensing data; and rainfall thresholds at the time of events are obtained from historical data. These are then integrated to form a multi-source data feature set containing multi-dimensional features, providing comprehensive input dimensions for subsequent physical constraint model training. This enables the model to simultaneously capture the combined impact of groundwater level decline (driving factor), vegetation degradation (moderating factor), and historical extreme events (risk patterns) on settlement.

[0090] In step S3, the multi-source data feature set is divided into a training feature set, a validation feature set, and an input feature set. The training feature set and the validation feature set are input into the model to be trained, and the land subsidence prediction model is trained and validated.

[0091] The model to be trained uses Terzaghi's effective stress principle as a physical constraint. The core of Terzaghi's effective stress principle is to strongly correlate the groundwater level (u) and geological parameters (H, E_s) with the settlement (S) through the formula S=(σ -u)*H / E_s, so as to ensure that the prediction results of the model conform to the laws of geomechanics and avoid the occurrence of physically impossible prediction values.

[0092] Specifically, Terzaghi's effective stress principle serves as a physical constraint, forcing the model to learn mapping relationships that conform to geological laws, thus avoiding overfitting of purely data-driven models to noisy data or distortion of physical meaning due to sample bias. During the model training phase, the strict division of training feature sets, validation feature sets, and input feature sets achieves independence between model training and validation. Specifically, the steps of dividing the multi-source data feature set into training feature sets, validation feature sets, and input feature sets, inputting the training feature sets and validation feature sets into the model to be trained, and training and validating the surface subsidence prediction model include:

[0093] The multi-source data feature set is divided into three consecutive time periods in chronological order: a first time period, a second time period, and a third time period, which are respectively used as the training feature set, the verification feature set, and the input feature set.

[0094] The first time period is the earliest continuous time interval in the multi-source data feature set, used to build the training feature set. It needs to cover at least 3 years of monthly data to include complete settlement cycle patterns and provide basic learning samples for the model. The second time period is a continuous time interval following the first time period, used to build the validation feature set. The time span is no less than 6 months and it is independent of the training set, used to monitor the model's generalization ability. The third time period is the latest continuous time interval, used to build the input feature set, which includes multi-source data features from the 12 months prior to the prediction, as input data for the model to predict future settlement.

[0095] The iteration condition is the termination criterion in the model training process. It is usually the condition that the prediction error (such as RMSE) on the verification feature set no longer decreases for 5 consecutive rounds, or the preset number of training rounds (such as 100 rounds) is reached to ensure that the model converges and avoids overfitting.

[0096] The training feature set and the validation feature set are input into the model to be trained. The model is trained using the training feature set, and the validation feature set is used to determine whether the iteration conditions are met. The iteration conditions are met until the iteration conditions are met, and the land subsidence prediction model is obtained.

[0097] Specifically, the multi-source data feature set is divided into three consecutive time periods in chronological order. The first time period serves as the training feature set, containing long-term dynamic patterns of historical subsidence; the second time period serves as the validation feature set, used to evaluate the model's suitability for recent data; and the third time period serves as the input feature set, providing the latest environmental and deformation data. The division process strictly follows the chronological order to prevent future data from leaking into the training phase, ensuring that the model learns the mapping relationship from history to the future.

[0098] Subsequently, the training feature set is input into the model to be trained (e.g., a physical information LSTM), and the predicted values ​​are calculated through forward propagation. A loss function is constructed by combining the physical constraint terms of Terzaghi's effective stress principle, and the model parameters are optimized through backpropagation. After each training round, the prediction error (e.g., MAE, RMSE) is calculated using the validation feature set. If the error does not meet the iteration conditions, the learning rate, physical constraint weights, and other parameters are adjusted, and training continues. If the error is consistently below the threshold for 5 consecutive rounds, the iteration stops, and a converged land subsidence prediction model is obtained.

[0099] More specifically, in the embodiments of this application, the expression for the loss function is:

[0100] ;

[0101] in, The mean square error between the predicted and actual values. For physical constraint terms ( Permeability coefficient, For the specific gravity of water, (Water level).

[0102] In step S4, the input feature set is input into the surface subsidence prediction model, and the surface subsidence prediction model outputs subsidence prediction data.

[0103] Among them, the subsidence prediction data is the quantitative result of the future period output by the surface subsidence prediction model, including the monthly subsidence prediction value, the subsidence rate change trend and the preset confidence interval (e.g., 95%), so as to realize the quantification of uncertainty and provide a basis for adjusting the early warning threshold.

[0104] Specifically, the input feature set is organized into a three-dimensional tensor according to a time window (e.g., 12 consecutive months), and input into the trained land subsidence prediction model. The model extracts key features through a spatiotemporal attention mechanism, and finally generates subsidence prediction data through multiple samplings by the Dropout layer in the model.

[0105] More specifically, in this embodiment of the application, the step of inputting the input feature set into the land subsidence prediction model and outputting subsidence prediction data through the land subsidence prediction model includes:

[0106] The input feature set is input into the land subsidence prediction model, and the land subsidence prediction model generates the subsidence prediction value and the corresponding confidence interval for a preset future time period.

[0107] The predicted settlement values ​​and their corresponding confidence intervals are integrated into the predicted settlement data output.

[0108] The preset future period is the time range for model prediction, usually 1-12 months from now. It can be dynamically adjusted according to engineering needs to ensure coverage of short-term early warning (1-3 months) and medium- to long-term trend (6-12 months) analysis. The subsidence prediction value is a point estimate of the monthly subsidence within the preset future period by the surface subsidence prediction model. The unit is millimeters, reflecting the average subsidence level within that period, such as "subsidence of 2.5 mm in January 2024". The confidence interval is a statistical range that characterizes the uncertainty of the prediction value. Usually, a 95% confidence interval is used to quantify the reliability of the model prediction.

[0109] Specifically, the input feature set (including deformation features, environmental features, remote sensing features, and geocoding features for the 12 months prior to prediction) is organized into a three-dimensional tensor according to a time window and input into the trained surface subsidence prediction model (Physical Information LSTM). The model focuses on the spatial features of subsidence hotspots through 3D convolutional layers, captures the temporal evolution patterns through LSTM layers, and forces the model to output predicted values ​​that conform to geological laws through physical constraints based on Terzaghi's effective stress principle.

[0110] More specifically, to generate confidence intervals, the model uses a Dropout layer and performs multiple forward propagations during the inference phase. Each propagation produces slightly different prediction results due to randomly deactivated neurons, forming a probability distribution of the predicted values. The 95% confidence interval is calculated using statistical distribution to quantify the prediction uncertainty. Finally, the monthly settlement prediction values ​​for a preset future time period are integrated with the corresponding 95% confidence intervals in chronological order to form the settlement prediction data.

[0111] In step S5, real-time hydrological data is acquired, and the settlement early warning threshold is updated based on the settlement prediction data and the real-time hydrological data.

[0112] Real-time hydrological data refers to hydrogeological parameters monitored in real time by sensors, including groundwater level, daily rainfall, and soil moisture content. It directly reflects the dynamic impact of the current hydrological environment on land subsidence and is a key real-time input for adjusting early warning thresholds.

[0113] Specifically, firstly, real-time hydrological data of the target detection area is continuously collected through a sensor network such as the Internet of Things (IoT). This real-time hydrological data is then spatiotemporally matched with the subsidence prediction data output by the surface subsidence prediction model to analyze the potential impact of current hydrological conditions on the subsidence trend. Based on this analysis, a reinforcement learning agent (RL agent) is used to dynamically update the subsidence early warning threshold. The RL agent aims to minimize the false alarm rate and maximize the timeliness of early warnings. It is trained using historical early warning cases to develop a decision-making strategy, thereby adjusting the subsidence early warning threshold.

[0114] More specifically, in this embodiment of the application, the step of acquiring real-time hydrological data and updating the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data includes:

[0115] The settlement prediction data and the real-time hydrological data are input into a preset reinforcement learning agent, and the reinforcement learning agent is used to match the optimal threshold adjustment strategy under the current data background.

[0116] The original early warning threshold is updated according to the threshold adjustment strategy to generate the settlement early warning threshold.

[0117] Among them, the threshold adjustment strategy is a specific decision scheme output by the reinforcement learning agent, which includes the direction and magnitude of threshold adjustment, and is generated based on the matching results of the current settlement prediction data and real-time hydrological data; the original warning threshold is the settlement warning critical value set before the model update, which is determined based on historical statistical data or engineering specifications, and serves as the benchmark value for dynamic adjustment.

[0118] Specifically, the settlement prediction data and real-time hydrological data are first input into a preset reinforcement learning agent. The agent uses a deep reinforcement learning algorithm (such as PPO) to match the optimal threshold adjustment strategy. Through continuous interaction with the environment, the agent gradually learns the optimal strategy under different data backgrounds. For example, when the real-time groundwater level drop rate is >0.3m / h and the settlement amount in the settlement prediction data for the next 3 months is >90% of the original threshold, the strategy of "lowering the threshold by 5mm" is output; when the water level is stable and the predicted settlement rate is <50% of the original threshold, the strategy of "raising the threshold by 3mm" is output.

[0119] Subsequently, the original early warning threshold is updated according to the threshold adjustment strategy to generate a settlement early warning threshold adapted to the current environment. During the adjustment process, if the 95% confidence interval of the settlement prediction data is wide, the strategy will automatically reduce the adjustment range to ensure the robustness of the threshold update.

[0120] For example, suppose a city's initial warning threshold is set at 50mm. In July 2024, the data input into the reinforcement learning agent was: subsidence prediction data showed a subsidence of 48mm over the next three months (95% confidence interval ±3mm), and real-time hydrological data showed the groundwater level dropping from 10m to 8m within 24 hours (a rate of decrease of 0.083m / h). The agent, by matching historical cases (a similar water level drop event in 2022 leading to a 20% acceleration in subsidence), output a strategy to "reduce the threshold to 45mm." After the update, the predicted subsidence of 48mm exceeded the 45mm threshold, and the system issued a warning five days in advance. Actual monitoring showed that the actual subsidence in the area reached 47mm within three months. Due to timely groundwater recharge measures implemented after the warning, the final subsidence was controlled at 44mm, preventing a disaster. During this process, the agent improved the warning accuracy in similar scenarios from 60% to 95% by learning from historical false alarm cases (three missed alarms caused by a fixed threshold of 50mm in 2021), validating the effectiveness of the dynamic strategy.

[0121] In step S6, based on the subsidence warning threshold and the subsidence prediction data, surface subsidence warning information for the target detection area is generated.

[0122] Among them, the surface subsidence early warning information is a multi-dimensional output that comprehensively reflects the subsidence risk of the target detection area, including the early warning level, the spatial distribution of risk (subsidence risk heat map) and targeted prevention and control suggestions. The early warning level is a risk level divided according to the comparison results of subsidence prediction data and early warning threshold.

[0123] Specifically, the subsidence prediction data output by the surface subsidence prediction model is first compared with the dynamically updated subsidence warning threshold on a grid-by-grid basis. For each cell in the target detection area grid, it is determined whether its predicted subsidence exceeds the corresponding warning threshold. At the same time, the confidence interval is referenced. If the lower limit of the interval exceeds the threshold, it is directly determined to be of that level; if the interval contains the threshold, it is marked as "potential risk".

[0124] Subsequently, warning levels were determined based on the comparison results. Grid cells were marked based on predicted settlement, and the corresponding level information was overlaid onto the geospatial data using a GIS system to generate a settlement risk heatmap. Red indicated dangerous areas, yellow indicated areas of concern, and green indicated general areas, visually displaying the spatial distribution of risk. Finally, prevention and control recommendations were automatically pushed out based on the risk level classification.

[0125] Furthermore, after the step of generating surface subsidence early warning information for the target detection area based on the subsidence early warning threshold and the subsidence prediction data, the method further includes:

[0126] Obtain the actual settlement monitoring data of the target detection area;

[0127] Among them, the actual settlement monitoring data are the real settlement data of the target detection area obtained by means of synthetic aperture radar time-series interferometry (TS-InSAR), leveling or GNSS after the surface settlement early warning information is generated. The time resolution is matched with the predicted data (such as monthly) to verify the accuracy of the early warning information.

[0128] The actual settlement monitoring data is compared with the corresponding surface settlement early warning information to obtain the prediction deviation value;

[0129] The prediction deviation value is the difference between the actual settlement monitoring data and the predicted data in the corresponding surface settlement early warning information. It is usually expressed as root mean square error (RMSE) or mean absolute error (MAE) to quantify the degree of deviation between the predicted value and the actual value.

[0130] The prediction bias value is input into the reinforcement learning agent to continuously perform optimization updates.

[0131] Specifically, after generating surface subsidence early warning information, the system continuously collects actual subsidence monitoring data of the target detection area, and needs to ensure that the monitoring accuracy is consistent with the spatial resolution in order to avoid bias and misjudgment caused by data heterogeneity; the actual monitoring data and the predicted data in the corresponding early warning information are aligned in time and space dimensions, the prediction deviation value of each grid is calculated, and the overall deviation of the region is statistically analyzed.

[0132] Subsequently, the predicted bias values ​​are input into the reinforcement learning agent as new training samples to optimize its threshold adjustment strategy. If the reinforcement learning agent detects that the predicted value is higher than the actual value, it determines that the threshold is set too low and the adjustment sensitivity needs to be increased; otherwise, it indicates that the warning threshold is set too high and the adjustment magnitude needs to be reduced. Finally, the agent updates the policy network parameters using the gradient descent algorithm, enabling the new strategy to reduce similar biases in future warnings while retaining historical effective experience. The optimized agent will be applied to the next round of warning threshold adjustments.

[0133] Reference Figure 2 , Figure 2 This is a virtual structural diagram of the land subsidence early warning device based on synthetic aperture radar provided in this application. A second aspect of this application provides a land subsidence early warning device based on synthetic aperture radar, comprising:

[0134] The deformation data acquisition module 100 is used to acquire synthetic aperture radar images of the target detection area from the same satellite at different time points, and generate time-series deformation data based on all the synthetic aperture radar images.

[0135] The multi-source feature construction module 200 is used to acquire multi-source auxiliary data, fuse the multi-source auxiliary data with the time-series deformation data, perform feature construction on the fused data, and obtain a multi-source data feature set.

[0136] The prediction model training module 300 is used to divide the multi-source data feature set into a training feature set, a validation feature set, and an input feature set, and input the training feature set and the validation feature set into the model to be trained, thereby training and validating the land subsidence prediction model; the model to be trained is based on Terzaghi's effective stress principle as a physical constraint.

[0137] The settlement prediction generation module 400 is used to input the input feature set into the surface settlement prediction model and output settlement prediction data through the surface settlement prediction model.

[0138] The early warning threshold update module 500 is used to acquire real-time hydrological data and update the settlement early warning threshold according to the settlement prediction data and the real-time hydrological data.

[0139] The settlement early warning output module 600 is used to generate surface settlement early warning information for the target detection area based on the settlement early warning threshold and the settlement prediction data.

[0140] The surface subsidence early warning device based on synthetic aperture radar described in this application embodiment can execute the surface subsidence early warning method based on synthetic aperture radar provided in the above embodiments. The surface subsidence early warning device based on synthetic aperture radar has the corresponding functional steps and beneficial effects of the surface subsidence early warning method based on synthetic aperture radar described in the above embodiments. For details, please refer to the embodiments of the surface subsidence early warning method based on synthetic aperture radar described above. The embodiments of this application will not be repeated here.

[0141] This application also provides an electronic device, please refer to... Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include a processor and a memory, which can be connected via a bus or other means. The processor may be a Central Processing Unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the synthetic aperture radar-based land subsidence early warning method in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby realizing the synthetic aperture radar-based land subsidence early warning method in the above method embodiments.

[0142] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. The one or more modules are stored in the memory and, when executed by the processor, perform the land subsidence early warning method based on synthetic aperture radar as described in the above method embodiments. Specific details of the above electronic device can be understood by referring to the corresponding descriptions and effects in the above method embodiments, and will not be repeated here. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it may include the processes of the embodiments of the above methods. The storage medium may be a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD), etc.; the storage medium may also include a combination of the above types of memory.

[0143] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0144] Similarly, it should be understood that, in order to streamline this disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this application, various features of this application are sometimes grouped together in a single embodiment, figure, or description thereof. However, this approach to disclosure should not be construed as reflecting an intention that the claimed application requires more features than expressly recited in each claim. Rather, as reflected in the claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0145] It should be noted that the above embodiments are illustrative of this application and not restrictive of this application, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims.

Claims

1. A ground subsidence early warning method based on synthetic aperture radar, characterized by, The method includes: Acquire synthetic aperture radar images of the target detection area from the same satellite at different time points, and generate time-series deformation data based on all the synthetic aperture radar images; Acquire multi-source auxiliary data, fuse the multi-source auxiliary data with the time-series deformation data, perform feature construction on the fused data, and obtain a multi-source data feature set; The multi-source data feature set is divided into a training feature set, a validation feature set, and an input feature set. The training feature set and the validation feature set are input into the model to be trained, and the surface subsidence prediction model is trained and validated. The model to be trained is based on Terzaghi's effective stress principle as a physical constraint. The input feature set is input into the surface subsidence prediction model, and the surface subsidence prediction model outputs subsidence prediction data. Acquire real-time hydrological data, and update the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data; Based on the subsidence warning threshold and the subsidence prediction data, surface subsidence warning information for the target detection area is generated.

2. The synthetic aperture radar-based land subsidence early warning method according to claim 1, characterized in that, The step of acquiring synthetic aperture radar images of the same satellite at different time points for the target detection area, and generating time-series deformation data based on all the synthetic aperture radar images, includes: All the synthetic aperture radar images are sorted by time to obtain an image time series; The image time series is preprocessed, and based on any two images in the preprocessed image time series that are separated by a preset time interval, a deformation profile of the target detection area is obtained. Local deformation regions are extracted from the deformation profile, and all synthetic aperture radar images are cropped based on the local deformation regions. Based on all the cropped synthetic aperture radar images, a time-series detail map of the local deformation region is generated, and the time-series deformation data is extracted from the time-series detail map.

3. The surface subsidence early warning method based on synthetic aperture radar according to claim 1, characterized in that, The multi-source auxiliary data includes environmental auxiliary data, remote sensing auxiliary data, and historical auxiliary data; The step of fusing the multi-source auxiliary data with the time-series deformation data, and performing feature construction on the fused data to obtain a multi-source data feature set includes: The multi-source auxiliary data and the temporal deformation data are spatiotemporally aligned to a preset spatiotemporal resolution. Deformation features are extracted based on the aligned temporal deformation data, environmental features and geological category features are extracted based on the aligned environmental auxiliary data, surface ecological features are extracted based on the aligned remote sensing auxiliary data, and historical case features are extracted based on the aligned historical auxiliary data. The geological category features are encoded, and the deformation features, environmental features, surface ecological features, historical case features, and encoded geological category features are integrated into the multi-source data feature set.

4. The surface subsidence early warning method based on synthetic aperture radar according to claim 1, characterized in that, The steps of dividing the multi-source data feature set into a training feature set, a validation feature set, and an input feature set, and inputting the training feature set and the validation feature set into the model to be trained, and training and validating the land subsidence prediction model include: The multi-source data feature set is divided into three consecutive time periods in chronological order: a first time period, a second time period, and a third time period, which are respectively used as the training feature set, the verification feature set, and the input feature set. The training feature set and the validation feature set are input into the model to be trained. The model is trained using the training feature set, and the validation feature set is used to determine whether the iteration conditions are met. The iteration conditions are met until the iteration conditions are met, and the land subsidence prediction model is obtained.

5. The surface subsidence early warning method based on synthetic aperture radar according to claim 1, characterized in that, The step of inputting the input feature set into the land subsidence prediction model and outputting subsidence prediction data through the land subsidence prediction model includes: The input feature set is input into the land subsidence prediction model, and the land subsidence prediction model generates the subsidence prediction value and the corresponding confidence interval for a preset future time period. The predicted settlement values ​​and their corresponding confidence intervals are integrated into the predicted settlement data output.

6. The surface subsidence early warning method based on synthetic aperture radar according to claim 1, characterized in that, The step of acquiring real-time hydrological data and updating the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data includes: The settlement prediction data and the real-time hydrological data are input into a preset reinforcement learning agent, and the reinforcement learning agent is used to match the optimal threshold adjustment strategy under the current data background. The original early warning threshold is updated according to the threshold adjustment strategy to generate the settlement early warning threshold.

7. The surface subsidence early warning method based on synthetic aperture radar according to claim 6, characterized in that, After the step of generating surface subsidence early warning information for the target detection area based on the subsidence early warning threshold and the subsidence prediction data, the method further includes: Obtain the actual settlement monitoring data of the target detection area; The actual settlement monitoring data is compared with the corresponding surface settlement early warning information to obtain the prediction deviation value; The prediction bias value is input into the reinforcement learning agent to continuously perform optimization updates.

8. A surface subsidence early warning device based on synthetic aperture radar, characterized in that, include: The deformation data acquisition module is used to acquire synthetic aperture radar images of the target detection area from the same satellite at different time points, and generate time-series deformation data based on all the synthetic aperture radar images. A multi-source feature construction module is used to acquire multi-source auxiliary data, fuse the multi-source auxiliary data with the time-series deformation data, perform feature construction on the fused data, and obtain a multi-source data feature set. The prediction model training module is used to divide the multi-source data feature set into a training feature set, a validation feature set, and an input feature set, and input the training feature set and the validation feature set into the model to be trained, thereby training and validating the land subsidence prediction model; the model to be trained is based on Terzaghi's effective stress principle as a physical constraint. The settlement prediction generation module is used to input the input feature set into the surface settlement prediction model and output settlement prediction data through the surface settlement prediction model. The early warning threshold update module is used to acquire real-time hydrological data and update the settlement early warning threshold based on the settlement prediction data and the real-time hydrological data. The settlement early warning output module is used to generate surface settlement early warning information for the target detection area based on the settlement early warning threshold and the settlement prediction data.

9. An electronic device, characterized in that, include: One or more processors; One or more memory units; And one or more computer programs, wherein the one or more computer programs are stored in the one or more memories, the one or more computer programs including instructions that, when executed by the one or more processors, cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program or instructions that, when executed, implement the method as described in any one of claims 1 to 7.

Images