Intelligent interpretation method and system for slope deformation based on insar time-series images

Abstract

The application discloses a slope deformation intelligent interpretation method and system based on InSAR time sequence images, and belongs to the technical field of geological disaster monitoring. The method comprises the following steps: acquiring multi-temporal SAR images and generating a deformation rate field and a time sequence; adopting a space-time dual-domain self-adaptive enhancement algorithm to improve the deformation signal quality of a low-coherence region; extracting the space-time features of deformation based on a multi-scale space-time graph attention-Transformer coupling network; fusing deformation features, geological and topographic factors and environmental trigger factors to perform multi-modal risk intelligent evaluation; and reversely optimizing the preposed module parameters according to the evaluation confidence, so as to realize closed-loop collaborative optimization. Through the deep coupling and closed-loop feedback among the modules, the method realizes the collaborative improvement of the deformation signal quality, the feature extraction effect and the risk evaluation accuracy, and provides a high-precision and high-reliability technical means for large-range slope monitoring and geological disaster prevention.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and prevention technology, and in particular to an intelligent interpretation method and system for slope deformation based on time-series images of synthetic aperture radar interferometry, belonging to the interdisciplinary field of remote sensing monitoring, deep learning and geological disaster early warning. Background Technology

[0002] Slope instability is a common type of geological hazard in mountainous and reservoir areas, posing a serious threat to infrastructure safety, people's lives and property, and the ecological environment. Traditional slope monitoring mainly relies on ground-based surveying equipment such as total stations, GPS, and inclinometers. Although these methods are highly accurate, they suffer from problems such as difficulty in deployment, limited coverage, and high labor costs, making it difficult to meet the real-time monitoring needs of large-scale slope groups.

[0003] Synthetic Aperture Radar (InSAR) interferometry can extract surface deformation information from satellite remote sensing imagery, offering advantages such as wide coverage, all-weather observation, and centimeter-level accuracy, making it an important tool for geological disaster monitoring. In particular, permanent scatterer interferometry (PSI) and distributed scatterer interferometry (DSI), through joint processing of long-term SAR imagery, can acquire high-density deformation monitoring points and continuous deformation time series. However, slope areas are typically located in vegetated mountainous regions, where severe spatiotemporal decoherence problems exist, leading to decreased InSAR deformation measurement accuracy. Deformation signals are submerged in noise, making it difficult to accurately identify potential landslide hazards.

[0004] Existing InSAR deformation interpretation methods mainly suffer from the following shortcomings:

[0005] Chinese invention CN114119643A discloses an automatic method for extracting deformation zone boundaries based on the Alphashape algorithm. This method uses statistical analysis to screen PS points with abnormal deformation rates, constructs a Delaunay triangulation, and then uses the Alphashape algorithm to extract the deformation zone boundaries. While this method can automatically extract deformation zone boundaries, it has the following limitations: First, it only uses deformation rate as a single indicator for anomaly point screening, without considering the evolutionary characteristics and spatial distribution patterns of deformation time series, resulting in insufficient ability to identify complex deformation types such as accelerated deformation and periodic deformation. Second, it uses a fixed statistical threshold for anomaly point discrimination, which cannot adapt to the differences in deformation characteristics under different geological conditions and environmental factors, easily leading to a large number of false positives or missed detections in low-coherence mountainous areas. Third, it only extracts the spatial boundaries of the deformation zone, lacking in-depth interpretation of deformation mechanisms, risk levels, and development trends, making it difficult to support graded early warning and risk management decisions for landslide hazards. Finally, this method is a one-way processing flow, lacking a feedback optimization mechanism, and cannot dynamically adjust processing parameters based on the reliability of the identification results, resulting in insufficient stability and accuracy in complex scenarios.

[0006] Existing technologies have employed deep learning methods to interpret InSAR deformation, but these methods mostly treat deformation data as a single input feature, failing to fully explore the spatiotemporal structural features of deformation signals. In low-coherence mountainous areas, InSAR deformation data contains significant noise and errors, making direct application of deep learning models prone to overfitting and poor generalization. Furthermore, slope deformation is influenced by a combination of factors, including geological structure, topography, rainfall, and reservoir water level changes. Relying solely on deformation data is insufficient for accurately assessing slope instability risk, necessitating the fusion of multi-source heterogeneous data for comprehensive analysis. Current methods still have significant shortcomings in multimodal data fusion, long-term deformation pattern recognition, and spatial dependency modeling, limiting the effectiveness of InSAR technology in large-scale slope monitoring.

[0007] Therefore, there is an urgent need to develop an intelligent interpretation method for slope deformation that can adaptively enhance deformation signals in low-coherence regions, deeply mine spatiotemporal features, fuse multimodal information, and achieve closed-loop optimization, so as to improve the applicability of InSAR deformation monitoring in complex mountainous areas and the accuracy of slope hazard identification. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent interpretation method and system for slope deformation based on InSAR time-series imagery. This method improves the quality of deformation signals in low-coherence areas through a spatial-temporal dual-domain adaptive enhancement algorithm, employs a multi-scale spatiotemporal map attention-Transformer coupled network to deeply mine the spatiotemporal characteristics of deformation, integrates geological and topographical factors and environmental triggering factors for multimodal intelligent risk assessment, and constructs a closed-loop parameter optimization mechanism to achieve synergistic optimization of deformation signal enhancement and risk assessment. This enables high-precision deformation monitoring of large-scale slope groups, intelligent identification of multiple types of hidden dangers, and graded risk early warning.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] The intelligent interpretation method for slope deformation based on InSAR time-series images includes five steps: InSAR image data acquisition and preprocessing, adaptive enhancement of deformation signals, multi-scale spatiotemporal feature extraction, multi-modal risk intelligent assessment, and closed-loop parameter optimization.

[0011] The InSAR image data acquisition and preprocessing steps acquire multi-temporal SAR images and generate deformation rate fields and cumulative deformation time series using temporal interferometry. The deformation signal adaptive enhancement step employs a spatial-temporal dual-domain adaptive enhancement algorithm, calculating adaptive enhancement weights in both the spatial and temporal domains. Weighted filtering of the deformation data is then performed through weight fusion to generate the enhanced deformation field. The multi-scale spatiotemporal feature extraction step constructs a spatial adjacency graph of slope units. A multi-scale spatiotemporal graph attention-Transformer coupled network is used to extract deformation features. The graph attention network captures spatial dependencies, while the multi-head temporal Transformer captures temporal dependencies. A multi-scale feature pyramid fuses features from different scales. The multi-modal risk intelligent assessment step constructs a multi-modal risk assessment network, fusing deformation features, geological and topographical factors, and environmental triggering factors to identify deformation pattern types and predict risk levels and confidence levels. The closed-loop parameter optimization step triggers parameter optimization based on the risk assessment confidence level, adjusting key parameters of the preceding modules in reverse. Iterative optimization improves the assessment confidence level, achieving closed-loop collaborative optimization.

[0012] This invention also provides an intelligent slope deformation interpretation system based on InSAR time-series imagery, comprising an InSAR image data acquisition and preprocessing module, a deformation signal adaptive enhancement module, a multi-scale spatiotemporal feature extraction module, a multi-modal risk intelligent assessment module, and a closed-loop parameter optimization module. These modules form a deeply coupled closed-loop collaborative system through data flow and parameter feedback. The output of the pre-module serves as the key input to the post-module, and the evaluation results of the post-module inversely optimize the parameters of the pre-module, achieving a synergistic improvement in deformation signal quality, feature extraction effectiveness, and risk assessment accuracy.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] First, the spatial-temporal dual-domain adaptive enhancement algorithm proposed in this invention can adaptively adjust the filtering intensity in the spatial domain based on neighborhood coherence and deformation gradient, and in the temporal domain based on the smoothness of the time series curve and abrupt change characteristics, thus significantly improving the signal-to-noise ratio of the deformation signal through dual-domain fusion enhancement. Compared with existing fixed threshold filtering methods, the adaptive enhancement algorithm of this invention can preserve small deformation signals while suppressing noise, and improves the deformation signal enhancement effect by more than 40% in low-coherence regions, providing high-quality input data for subsequent deformation feature extraction and risk assessment.

[0015] Second, this invention innovatively combines graph attention networks with multi-head temporal Transformers to construct a multi-scale spatiotemporal graph attention-Transformer coupled network. The graph attention network captures the spatial dependencies between adjacent slope units through a dynamic attention mechanism, enabling the identification of deformation propagation and diffusion patterns in space. The multi-head temporal Transformer captures the long-range dependencies of deformation time series through a self-attention mechanism, enabling the identification of deformation patterns at different time scales. The multi-scale feature pyramid achieves the fusion of fine-grained local deformation and coarse-grained overall trends. Compared with existing single CNN or RNN methods, this spatiotemporally coupled network architecture improves the accuracy of complex deformation pattern identification by 35%, effectively distinguishing different types such as linear creep, accelerated deformation, periodic deformation, and abrupt deformation.

[0016] Third, the multimodal risk intelligent assessment network constructed in this invention integrates three types of heterogeneous data: InSAR deformation features, geological and topographical factors, and environmental triggering factors. Through deep semantic understanding, the network mines the correlation patterns between different data sources, identifying deformation triggering mechanisms and instability modes. Compared to existing methods that only base risk assessment on deformation rate, the multimodal fusion assessment of this invention can more accurately predict slope instability risk, achieving a risk level prediction accuracy of 92% and reducing the false alarm rate by more than 50%. Simultaneously, this invention outputs a confidence index for risk assessment, providing decision-makers with a reliability reference and enhancing the credibility of the early warning system.

[0017] Fourth, this invention creatively constructs a closed-loop parameter optimization mechanism. Based on the confidence level of the risk assessment, it adjusts the key parameters of the deformation signal enhancement and feature extraction modules in reverse, achieving synergistic optimization of deformation signal quality and risk assessment accuracy. This closed-loop feedback mechanism enables the system to adaptively adjust its processing strategy according to the reliability of the assessment results, significantly improving stability in complex scenarios. Compared to existing unidirectional processing flows, the closed-loop optimization system of this invention maintains high recognition accuracy under various geological conditions and environmental scenarios, improving system robustness by 60%.

[0018] Fifth, this invention achieves deep coupling and synergy among its modules. Improved deformation signal enhancement improves the accuracy of feature extraction, which in turn enhances the reliability of risk assessment. Feedback on risk assessment confidence optimizes the signal enhancement strategy, creating a positive feedback loop where modules mutually reinforce each other. This deep coupling and synergistic effect results in a non-linear growth characteristic where 1+1 is greater than 2 in the overall system performance. Compared to the simple process superposition of existing technologies, the collaborative optimization system of this invention improves the overall performance of slope hazard identification by more than 70%.

[0019] In summary, this invention utilizes four innovative technologies—spatial-temporal dual-domain adaptive enhancement, multi-scale spatiotemporal graph attention-Transformer coupling, multimodal fusion evaluation, and closed-loop parameter optimization—to construct a high-precision, highly reliable, and robust intelligent interpretation system for slope deformation, providing advanced technical means for large-scale slope group monitoring and geological disaster prevention. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the intelligent slope deformation interpretation method of the present invention;

[0021] Figure 2 This is a module architecture diagram of the system of the present invention;

[0022] Figure 3 This is a flowchart of the deformation signal adaptive enhancement module.

[0023] Figure 4 Network architecture diagram of the multi-scale spatiotemporal feature extraction module;

[0024] Figure 5 This is a structural diagram of the multimodal risk intelligent assessment module. Detailed Implementation

[0025] Please refer to the attached document. Figures 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0026] like Figure 1 As shown, the intelligent interpretation method for slope deformation based on InSAR time-series images provided by this invention includes the following steps:

[0027] Step 1: InSAR image data acquisition and preprocessing:

[0028] Sentinel-1 C-band SAR imagery covering the target monitoring area was acquired, spanning from January 2019 to December 2024, totaling 150 images with a temporal resolution of 12 days. Permanent scatterer interferometry was employed for joint processing of the long-term SAR imagery. First, image registration was performed, aligning all SAR images to a unified master image coordinate system with a registration accuracy better than 0.1 pixels. Then, high-coherence pixels were selected as candidate permanent scatterer points, with a coherence coefficient threshold set to 0.65.

[0029] Phase unwrapping was performed on each interferometric image pair, and a minimum cost flow algorithm was used to handle phase entanglement. Considering the large topographic relief in mountainous areas, a 30m resolution SRTM digital elevation model was used for terrain phase removal. Atmospheric delay correction employed a spatiotemporal filtering method, using high-pass filtering in the time domain to separate atmospheric and deformation phases, and polynomial fitting in the spatial domain to remove spatially correlated atmospheric delay. Orbital error correction was achieved by estimating orbital polynomial parameters.

[0030] The average deformation rate and cumulative deformation time series of each PS point were obtained through time-series network inversion. The deformation rate field, in mm / year, reflects the average deformation rate during the monitoring period. The deformation time series, in mm, records the cumulative deformation at each moment relative to the reference moment. The ascending-descending trajectory fusion technology decomposes the line-of-sight deformation into vertical and east-west deformation, improving the three-dimensional accuracy of deformation monitoring. After processing, approximately 500,000 high-quality PS points were obtained in the monitoring area, achieving a spatial density of 200 points / km. 2 .

[0031] Step 2: Adaptive enhancement of deformation signal:

[0032] like Figure 3 As shown, the deformation signal adaptive enhancement module 1 adopts a spatial-temporal dual-domain adaptive enhancement algorithm to improve the deformation signal quality, taking into account the characteristics of InSAR deformation data in low-coherence mountainous areas.

[0033] In the spatial domain, for each pixel within the monitoring area The spatial domain adaptive augmentation weights within the neighborhood window are calculated. The neighborhood window size is set to 5×5 pixels, approximately a practical range of 250m×250m. The spatial domain augmentation weights are calculated using the following innovative algorithm:

[0034] ,

[0035] in, For pixels Spatial domain enhancement weights, The average coherence coefficient of all pixels within the neighborhood window reflects the reliability of InSAR measurements in that area. The variance of the deformation gradient within the neighborhood window reflects the degree of spatial variation in deformation. and As a weighting coefficient, in the preferred embodiment The value is 0.6. A value of 0.4 balances the contributions of coherence and gradient variance to the enhancement weights. The scale parameter controls the influence of gradient variance on the weights, with a preferred value of 5.0 mm / m. This formula achieves a nonlinear mapping where a larger deformation gradient variance results in a smaller weight enhancement through an exponential decay function. It reduces the filtering intensity in boundary regions with drastic deformation changes to preserve deformation details, and increases the filtering intensity in regions with gentle deformation to suppress noise.

[0036] In the time domain, for each cell Deformation time series The time-domain adaptive augmentation weights are calculated using the following innovative algorithm:

[0037] ,

[0038] in, For pixels Time-domain augmentation weights, The coefficient of determination is the fit between the time-series deformation curve and the linear regression, ranging from 0 to 1. It reflects the smoothness of the deformation time series and the significance of the linear trend. The deformation acceleration index is calculated using the second-order finite difference method, with units of mm / year. 2 This reflects the abrupt changes in deformation. The maximum deformation acceleration within the study area is used for normalization. and As a weighting coefficient, in the preferred embodiment The value is 0.7. The value is set to 0.3, which makes the contribution of temporal smoothness to the enhancement weights slightly greater than that of abrupt changes. This formula realizes a positive correlation between a smoother deformation curve and a larger enhancement weight, and a negative correlation between a more obvious deformation abrupt change and a smaller enhancement weight. It increases the filtering strength in the deformation stability region and decreases the filtering strength in the deformation abrupt change region to preserve the abrupt change signal.

[0039] The overall enhancement weights are generated using an adaptive fusion algorithm:

[0040] ,

[0041] in, For pixels The overall enhancement weight, The fusion coefficient for the spatial domain weights is calculated using the following formula:

[0042] ,

[0043] in, The threshold value for the coherence coefficient is preferably 0.65. This is the kurtosis parameter of the sigmoid function, preferably set to 10, which controls the smoothness of the weight transformation. This fusion strategy achieves coherence-driven adaptive weight allocation: when the regional coherence is higher than the coherence coefficient threshold, When the coherence approaches 0, the fusion weights rely more on time-domain enhancement; when regional coherence is below the coherence coefficient threshold, Approaching 1, the fusion weights rely more on spatial domain enhancement. This adaptive mechanism allows the enhancement strategy to be dynamically adjusted based on local data quality.

[0044] The original deformation data is weighted and filtered using a comprehensive enhancement weighting method. For the deformation rate field, the enhanced deformation rate is calculated as follows:

[0045] ,

[0046] in, For pixels Enhanced deformation rate, For neighboring pixels The original deformation rate, For pixels neighborhood window, Let be the comprehensive enhancement weight for neighboring pixel j. For the deformation time series, the enhanced deformation is calculated as follows:

[0047] ,

[0048] in, For pixels At any moment Enhanced deformation, For time windows, including There are 11 time points in total, with 5 time points before and after. The time scale parameter is preferably set to 3 time units, approximately 36 days, to control the range of time filtering. The Gaussian kernel function in the formula implements time weighting centered on the current time, with data closer to the current time having a greater weight. Increase the weights in the time domain at time t'; This provides an index for each moment within the time window.

[0049] The signal-to-noise ratio (SNR) of the deformation signal was significantly improved through a spatial-temporal dual-domain adaptive enhancement algorithm. In low-coherence mountainous areas, the standard deviation of the enhanced deformation rate field decreased from the original 5.2 mm / year to 2.8 mm / year, with an SNR improvement of approximately 85%. The enhanced deformation time series maintained the true trend of deformation change while effectively suppressing random noise and systematic errors, providing high-quality input for subsequent feature extraction.

[0050] Step 3: Multi-scale spatiotemporal feature extraction:

[0051] like Figure 4 The multi-scale spatiotemporal feature extraction module 2 constructs a spatial adjacency graph of slope units based on the enhanced deformation data, and uses a multi-scale spatiotemporal graph attention-Transformer coupled network to extract deformation features.

[0052] First, the monitoring area was divided into slope units based on a terrain segmentation algorithm. Topographic features, including slope, aspect, and curvature, were extracted using a digital elevation model. Hydrological analysis methods were employed to identify catchment lines and watersheds, using the watersheds as boundaries to divide the area into several slope units. Each slope unit represents a slope surface with relatively homogeneous terrain and geological conditions, with a unit area controlled between 0.5 and 2.0 km². 2 Between. The entire area is divided into approximately 2,000 slope units.

[0053] For each slope unit, the deformation information of all PS points within its range is statistically analyzed, and characteristic indicators such as average deformation rate, maximum deformation rate, standard deviation of deformation rate, and average cumulative deformation are calculated. A deformation time series of the slope unit is constructed by using a weighted average of the deformation of all PS points, with the weights determined by the coherence coefficient of the PS points and the distance to the unit center.

[0054] Constructing the spatial adjacency diagram of slope units ,in This is a set of nodes, where each node represents a slope element. Let be the set of edges, representing the connections between nodes. Edge construction follows these rules: first, establish topologically adjacent edges, connecting slope elements sharing a boundary; then establish distance-adjacent edges, connecting slope elements with a centroid distance of less than 1500m. Edge weights... Determined by both Euclidean distance and deformation correlation coefficient:

[0055] ,

[0056] in, For slope units and The distance between the centroids, For distance scale parameters, a preferred value of 1000m is preferred. The Pearson correlation coefficient is the time series of deformation of two slope units. and The weighting coefficients are preferably 0.4 and 0.6. This weighting calculation formula takes into account both spatial proximity and deformation correlation, resulting in greater connection weights between slope units that are spatially close and have similar deformation characteristics.

[0057] The graph attention network layer is used to capture the spatial dependencies between slope units. A multi-head attention mechanism is employed, with eight attention heads set to capture spatial relationships in different dimensions. For slope units... It comes from its neighboring nodes The attention coefficient of the received data is calculated as follows:

[0058] ,

[0059] in, For slope units The input feature vector contains features such as deformation rate, cumulative deformation, slope, and aspect. It is a learnable linear transformation matrix with dimensions of 128×64. This is the attention weight vector, with a dimension of 128. This represents a vector concatenation operation. Attention weights are obtained through softmax normalization.

[0060] ,

[0061] in, For slope units The set of neighbors. Slope unit. The updated features are:

[0062] ,

[0063] in, The ELU activation function is used. By stacking three layers of graph attention networks, each containing eight attention heads, the feature dimensions increased from 64 to 128 and then to 256, enabling deep learning of spatial dependencies. The graph attention network can adaptively learn the influence intensity between different slope units and identify the spatial propagation and diffusion patterns of deformation, such as the transmission of deformation from the top of the slope to the bottom and the linkage deformation of adjacent slopes.

[0064] The multi-head temporal Transformer layer is used to capture the long-range temporal dependencies of the deformation time series. For the deformation time series of each slope unit, location encoding is first performed to preserve temporal order information. The location encoding adopts a sine-cosine encoding method:

[0065] ,

[0066] ,

[0067] in, For time location index, Indexed by feature dimensions, The model dimension is set to 256. The multi-head self-attention mechanism is implemented by calculating the query, key, and value matrix:

[0068] ,

[0069] in, , , , The input deformation time series matrix has dimensions of , The length of the time series. , , Let be a learnable linear transformation matrix, with dimensions of 1. , Set as the dimension for query and key. , The number of attention heads is set to 16. Multi-head attention works by computing 16 attention heads in parallel and then concatenating their outputs.

[0070] ,

[0071] in, To output a linear transformation matrix with dimension 1 The Transformer layer also includes a feedforward neural network and residual connections. The feedforward network consists of two fully connected layers with an intermediate layer dimension of 1024 and the activation function is ReLU. Six Transformer encoder layers are stacked, each containing a multi-head self-attention sublayer and a feedforward network sublayer, with residual connections and layer normalization used between layers.

[0072] The temporal Transformer can capture multi-scale temporal patterns in deformation time series. Short-term fluctuation characteristics reflect the periodic impact of environmental factors such as seasonal rainfall and reservoir water level changes, with a time scale of several months; medium-term trend characteristics reflect the continuous creep or accelerated deformation process of slopes, with a time scale of 1 to 2 years; long-term evolution characteristics reflect the cumulative effects of geological tectonic activity and long-term environmental changes, with a time scale of many years. Through a self-attention mechanism, the Transformer can automatically learn the correlation between characteristics at different time scales and identify the evolutionary patterns and development trends of deformation.

[0073] The multi-scale feature pyramid module achieves the fusion of features at different spatiotemporal scales. Spatially, a three-level feature pyramid is constructed: Level 1 corresponds to the fine-grained deformation features of a single slope unit, with a spatial resolution of approximately 50m; Level 2 corresponds to the medium-grained deformation features of a group of slope units, aggregating 4 to 8 adjacent slope units into a supernode through graph pooling, with a spatial resolution of approximately 200m; Level 3 corresponds to the coarse-grained deformation features of the entire slope zone, aggregating them through further graph pooling, with a spatial resolution of approximately 1000m. Temporally, a three-level feature pyramid is also constructed: Level 1 corresponds to the original 12-day resolution deformation time series; Level 2 obtains the monthly-scale deformation time series through temporal downsampling; and Level 3 obtains the quarterly-scale deformation time series through further downsampling.

[0074] The feature pyramid fusion employs a bidirectional feature transfer approach, both bottom-up and top-down. The bottom-up path aggregates features level by level through graph pooling to learn coarse-grained features; the top-down path, through upsampling and residual connections, transfers coarse-grained semantic information to fine-grained features, enhancing the semantic expressive power of the fine-grained features. Finally, the features from the three scales are concatenated and fused to form a feature tensor containing multi-scale spatiotemporal information, with a dimension of [missing value]. ,in 768 represents the number of slope units, and 768 represents the feature dimensions (256 dimensions per scale).

[0075] The multi-scale spatiotemporal graph attention-Transformer coupled network automatically extracts spatiotemporal features from deformation data through deep learning, exhibiting stronger expressive and generalization capabilities compared to manually designed features. This network achieves an accuracy of 89% in recognizing different deformation patterns on a validation dataset, significantly outperforming traditional methods.

[0076] Step 4: Multimodal Risk Intelligent Assessment:

[0077] like Figure 5 As shown, the multimodal risk intelligent assessment module 3 constructs a multimodal risk assessment network, integrating deformation characteristics, geological and topographical factors, and environmental triggering factors to perform deformation pattern recognition and risk level prediction.

[0078] The deformation pattern recognition branch employs a deep convolutional neural network to classify multi-scale spatiotemporal feature tensors. The network structure consists of four convolutional blocks, each containing two 1D convolutional layers, a batch normalization layer, and a ReLU activation function. The kernel size is 3, the stride is 1, and the padding is 1. Downsampling is performed between convolutional blocks using max pooling layers with a pooling window size of 2. Features extracted by the convolutional layers are input to fully connected layers for classification. These fully connected layers consist of two layers: a hidden layer dimension of 512 and an output layer dimension of 4, corresponding to four deformation patterns: linear creep, accelerated deformation, periodic deformation, and abrupt deformation. A softmax activation function is used to output the probability distribution for each category, and the category with the highest probability is selected as the predicted deformation pattern.

[0079] Linear creep patterns are characterized by a relatively stable deformation rate, a linear or near-linear deformation time series, and near-zero deformation acceleration. This pattern typically occurs on slopes with relatively stable geological conditions and minimal environmental changes, where deformation is primarily caused by slow creep under gravity. Accelerated deformation patterns are characterized by an increasing deformation rate over time, a convex deformation time series, and positive, gradually increasing deformation acceleration. This pattern is a precursor to landslide instability, indicating that the slope is in an accelerated instability phase, requiring close monitoring and timely intervention. Periodic deformation patterns are characterized by periodic fluctuations in the deformation rate and a clear seasonal or interannual periodic component in the deformation time series. This pattern is usually associated with periodic environmental factors such as rainfall, reservoir water level changes, and freeze-thaw cycles, causing reversible or partially reversible deformation of the slope under the influence of these environmental factors. Abrupt deformation patterns are characterized by a rapid increase in the deformation rate over a short period, resulting in steep jumps in the deformation time series. This pattern may correspond to sudden destructive events within the slope, such as sudden expansion of cracks or localized landslides, indicating a sharp increase in the risk of slope instability.

[0080] The geological and topographic factor fusion branch is used to integrate the inherent susceptibility information of slopes. The extracted geological and topographic factors include: slope, reflecting the steepness of the slope; a steeper slope results in a greater gravity component and greater slope instability; aspect, affecting conditions such as sunshine, rainfall, and weathering; elevation, related to geological structure, lithology distribution, and rainfall; topographic curvature, including profile curvature and planar curvature, reflecting the concavity and convexity of the terrain; stratigraphy and lithology, obtained from geological maps, showing significant differences in mechanical strength among different lithologies; geological structure, including the spatial distribution of structural lines such as faults, folds, and joints; and slope structure, including types such as bedding slopes, reverse-dip slopes, and oblique slopes.

[0081] Geological and topographic factors are encoded into high-dimensional feature vectors through a feature embedding layer. For categorical factors such as lithology and slope structure type, an embedding lookup table is used to map them into 64-dimensional vectors. For numerical factors such as slope and elevation, normalization is first performed, and then they are mapped into 64-dimensional vectors through a fully connected layer. The embedded vectors of all geological and topographic factors are concatenated to form a geological and topographic feature vector with a dimension of 512. This vector is fused with a multi-scale spatiotemporal feature tensor through a feature fusion layer. The fusion layer adopts a multilayer perceptron structure, containing three fully connected layers with hidden layer dimensions of 1024, 512, and 256, respectively, and the activation function is ReLU. The last layer outputs the fused feature vector.

[0082] The environmental triggering factor fusion branch was used to capture the correlation between slope deformation and environmental factors. The acquired environmental triggering factors included: daily and cumulative rainfall, obtained from meteorological station observation data; rainfall is one of the main factors inducing slope instability; reservoir water level changes, for reservoir slopes, daily water level records were obtained from reservoir scheduling data; rises and falls in reservoir water level cause changes in slope infiltration and seepage conditions; seismic activity, obtained from earthquake catalogs of earthquake events near the study area during the monitoring period, including magnitude and epicentral distance; and human engineering activities, including the time and location information of excavation, blasting, and construction activities.

[0083] The time-lag correlation between environmental factors and deformation acceleration was analyzed. For each slope unit, the cross-correlation coefficients between its deformation time series and environmental factor time series at different time lags were calculated. The impact of rainfall on deformation typically has a lag of several days to several weeks, the impact of reservoir water level changes has a lag of several weeks to several months, and the impact of earthquakes may be instantaneous or have aftershock effects. By identifying the lag time corresponding to the maximum cross-correlation coefficient, the triggering response characteristics of environmental factors on the deformation of this slope unit were determined.

[0084] Environmental factors and their time-lag correlation features are encoded into an environmental trigger feature vector with a dimension of 256. This vector is further fused with the feature vector previously incorporated with geological and topographic factors to form a comprehensive multimodal feature representation with a dimension of 512. The multimodal feature vector is then input into the risk level prediction module for final risk assessment.

[0085] The risk level prediction module employs an ensemble learning algorithm for risk classification. Specifically, it uses the gradient boosting decision tree algorithm, which sequentially trains multiple weak classifiers (decision trees). Each classifier fits the residual of the previous classifier, and the prediction results of all classifiers are then integrated. A total of 500 decision trees are used, each with a maximum depth of 6, a learning rate of 0.05, and a subsampling rate of 0.8. Risk levels are divided into three categories: high risk, medium risk, and low risk.

[0086] The risk level assessment follows these criteria: when the deformation rate of the slope unit is greater than 20 mm / year and the deformation acceleration is greater than 5 mm / year. 2 When the deformation pattern is identified as accelerated deformation or abrupt deformation, it is judged as high risk; when the deformation rate is between 10 and 20 mm / year, the deformation exhibits linear creep or periodic deformation characteristics, and there is a significant correlation between environmental triggering factors such as rainfall and reservoir water level and the deformation, it is judged as medium risk; when the deformation rate is less than 10 mm / year, the deformation curve is stable, the deformation pattern is linear creep, and there is no obvious effect of environmental triggering factors or the correlation between environmental factors and deformation is weak, it is judged as low risk.

[0087] Confidence level is calculated based on the consistency of prediction results and historical validation accuracy. For each slope element, the confidence level calculation formula is:

[0088] ,

[0089] in, The confidence score ranges from 0 to 1. The maximum value among the predicted probabilities of risk level reflects the model's certainty regarding the prediction result. The historical misclassification rate represents the proportion of slope units that were misclassified in historical predictions. A higher historical misclassification rate indicates greater difficulty in predicting the area. A score for the completeness of multimodal data sources, when all data sources are complete. When some data sources are missing Reduced according to the missing percentage, , , The weighting coefficients are preferably 0.5, 0.3, and 0.2. A confidence score higher than 0.8 is considered a high-confidence prediction, while a score lower than 0.5 is considered a low-confidence prediction, requiring further optimization or manual review.

[0090] The multimodal risk intelligent assessment network achieves a comprehensive assessment of slope instability risk by integrating InSAR deformation features, geological and topographical factors, and environmental triggering factors. Validation results in the test area show that the accuracy of risk level prediction reaches 92%. Compared with traditional methods based solely on deformation rate thresholds, the false alarm rate is reduced by 53%, and the missed detection rate is reduced by 47%, significantly improving the reliability of early warnings.

[0091] Step 5: Closed-loop parameter optimization:

[0092] The closed-loop parameter optimization module 4 triggers the parameter optimization mechanism based on the confidence result of the risk assessment, and adjusts the key parameters of the front-end module in reverse to achieve closed-loop collaborative optimization of deformation signal enhancement and risk assessment.

[0093] The confidence level monitoring mechanism monitors the risk assessment confidence level of each slope unit in real time. A lower confidence level threshold is set. When the confidence score of a certain slope element is lower than If the assessment results for a given unit are deemed unreliable, the parameter optimization process is triggered. The number and spatial distribution of low-confidence slope units across the entire region are statistically analyzed to identify areas where assessment is difficult; these areas typically correspond to low-coherence mountainous regions, densely vegetated areas, or areas with complex geological conditions.

[0094] Gradient backpropagation optimization employs an end-to-end differentiable optimization framework. Deformation signal enhancement, feature extraction, and risk assessment are constructed into a unified deep learning model, defining a confidence loss function:

[0095] ,

[0096] in, For loss function, This represents the total number of slope units. For slope units The confidence score is calculated by considering the first term of the loss function, which is the negative sum of the confidence scores of all slope elements. Minimizing the loss is equivalent to maximizing the overall confidence score. This is the regularization coefficient, preferably 0.001. and These are the weighting parameters for the spatial and temporal domains in the deformation signal enhancement module. For the network parameters in the multi-scale spatiotemporal feature extraction module, the regularization term prevents overfitting caused by excessive parameter optimization.

[0097] The Adam optimization algorithm is used to calculate the gradient of the loss function with respect to the parameters, with the learning rate set to 0.0001 and the momentum parameter... , The backpropagation gradient is passed from the risk assessment module to the feature extraction module and the signal enhancement module, calculating the gradient of the parameters for each module. For the signal enhancement module, the optimized parameters include the spatial domain weight coefficients. , and scale parameters Time domain weighting coefficients , And the threshold in the calculation of the fusion coefficient and steepness parameter For the feature extraction module, the optimized parameters include the number of attention heads and feature dimensions of the graph attention network, the number of encoder layers and hidden layer dimensions of the temporal Transformer, and the number of scales of the feature pyramid.

[0098] Parameter updates are performed in the direction of gradient descent:

[0099] ,

[0100] in, Represents model parameters, For learning rate, This represents the gradient of the loss function with respect to the parameters. After the parameters are updated, the deformation signal enhancement, feature extraction, and risk assessment processes are re-executed, and the updated confidence score is calculated.

[0101] Iterative convergence is determined by monitoring the improvement in the confidence score. A confidence threshold is set. When the proportion of low-confidence slope units in the entire region drops below 5%, and more than 90% of the slope units have a confidence score higher than [missing information], When convergence is reached, the iteration is terminated. If convergence is not achieved after 10 iterations, the optimization is also terminated to avoid excessive iteration and wasted computational resources.

[0102] In practical applications, after 3 to 5 iterations, the average confidence score for the entire region increased from the initial 0.72 to 0.86, and the number of low-confidence slope units decreased from approximately 400 to approximately 80, a reduction of 80%. Parameter optimization made the signal enhancement strategy more adaptable to the data characteristics of different regions, enhanced the expressive power of the feature extraction network, and significantly improved the accuracy and reliability of risk assessment.

[0103] The closed-loop parameter optimization mechanism achieves deep coupling and collaborative optimization between deformation signal enhancement, feature extraction, and risk assessment. The output quality of the pre-modules (signal enhancement and feature extraction) directly affects the accuracy of the post-module (risk assessment), while the evaluation results of the post-modules, in turn, guide the parameter adjustment of the pre-modules. This closed-loop feedback mechanism enables the system to adaptively optimize the processing parameters throughout the entire process according to the ultimate goal (improving the reliability of risk assessment). Compared with unidirectional processing with fixed parameters, the system's robustness and adaptability are significantly enhanced. Under different geological conditions, seasons, and data quality scenarios, the closed-loop optimization system maintains a high recognition accuracy, and the system performance stability is improved by approximately 60%.

[0104] like Figure 2 As shown, the present invention also provides an intelligent interpretation system for slope deformation based on InSAR time-series images, including an InSAR image data acquisition and preprocessing module, a deformation signal adaptive enhancement module 1, a multi-scale spatiotemporal feature extraction module 2, a multi-modal risk intelligent assessment module 3, and a closed-loop parameter optimization module 4.

[0105] The InSAR image data acquisition and preprocessing module is responsible for acquiring SAR images from satellite data sources and performing preprocessing such as registration, phase unwrapping, atmospheric correction, and orbit correction to generate deformation rate fields and deformation time series. This module integrates multiple time-series InSAR processing algorithms, including permanent scatterer (PS) technology, distributed scatterer (DSS) technology, and small baseline set (SLS) technology. Users can select the appropriate algorithm based on the characteristics of the study area. The deformation data output by the module is stored in a standard format, including the spatial coordinates, deformation rate, coherence coefficient, and deformation time series for each PS point.

[0106] The deformation signal adaptive enhancement module 1 receives the deformation data output from the preprocessing module and enhances the deformation signal using a spatial-temporal dual-domain adaptive enhancement algorithm. Internally, the module implements a spatial domain enhancement weight calculation unit, a temporal domain enhancement weight calculation unit, a weight adaptive fusion unit, and a weighted filtering unit. Users can set key parameters such as weight coefficients and threshold parameters through the configuration interface, or choose to enable the automatic parameter optimization function, which allows the closed-loop optimization module to automatically adjust the parameters. The enhanced deformation data is then output to the feature extraction module.

[0107] The multi-scale spatiotemporal feature extraction module 2 includes a slope unit partitioning unit, a spatial adjacency graph construction unit, a graph attention network unit, a temporal Transformer unit, and a multi-scale feature pyramid unit. The slope unit partitioning unit automatically partitions slope units based on the digital elevation model and hydrological analysis results; the user can adjust the unit size range. The spatial adjacency graph construction unit establishes connections between slope units based on topological and distance adjacency relationships and calculates edge weights. The graph attention network unit and the temporal Transformer unit constitute the core of the deep learning model; model parameters are learned through training data and can also be dynamically adjusted by the closed-loop optimization module. The multi-scale feature pyramid unit fuses features at different scales, outputting a multi-scale spatiotemporal feature tensor.

[0108] The multimodal risk intelligent assessment module 3 includes a deformation pattern recognition unit, a geological and topographical factor fusion unit, an environmental triggering factor fusion unit, and a risk level prediction unit. The deformation pattern recognition unit uses a deep convolutional neural network to classify deformation features and identify deformation types. The geological and topographical factor fusion unit extracts geological and topographical factors from a geographic information system database and encodes them into vectors through feature embedding. The environmental triggering factor fusion unit obtains environmental factors from data sources such as meteorology, hydrology, and earthquakes, and analyzes their time-lag correlation with deformation. The risk level prediction unit integrates multimodal features and uses an ensemble learning algorithm to output the risk level and confidence score. The module provides a visual interface, displaying the risk distribution in map form, and supports interactive querying and analysis.

[0109] The closed-loop parameter optimization module 4 monitors the confidence level of the risk assessment. When the confidence level falls below a threshold, it triggers the parameter optimization process. The module employs an end-to-end differentiable optimization framework, constructs a unified loss function, calculates parameter gradients through gradient backpropagation, and automatically adjusts key parameters of the signal enhancement and feature extraction modules. The optimization process is iterative until the confidence level reaches the target threshold or the maximum number of iterations is reached. The module provides visual monitoring of the optimization process, displaying real-time confidence level change curves and parameter adjustment history.

[0110] The system's modules form a deeply coupled closed-loop collaborative architecture through data flow and parameter feedback. The output of the deformation signal enhancement module serves as the input to the feature extraction module, with signal quality directly affecting the feature extraction effect. The output of the feature extraction module serves as the input to the risk assessment module, with feature representation capability determining assessment accuracy. The confidence output of the risk assessment module triggers the parameter optimization module, guiding parameter adjustments based on assessment reliability. The adjustment signals from the parameter optimization module are fed back to the signal enhancement and feature extraction modules, updating parameters to improve signal quality and feature performance. This closed-loop collaborative mechanism enables each module to promote mutual benefit: clearer deformation signals lead to more accurate features; more accurate features lead to more reliable assessments; more reliable assessments lead to more optimized parameters; and more optimized parameters lead to clearer signals, achieving a positive cycle of synergistic efficiency.

[0111] The system supports batch processing and real-time monitoring modes. Batch processing mode is used to process historical SAR image data for regional-scale hazard identification and risk assessment. Real-time monitoring mode is used for rapid processing and early warning dissemination of newly acquired images; when a high-risk slope unit is detected, the system automatically generates an early warning and pushes it to management personnel. The system provides standardized data interfaces, supporting integration with geological disaster monitoring platforms and emergency management systems to achieve data sharing and operational collaboration.

[0112] The system has been applied in slope monitoring projects across multiple watersheds and reservoir areas, covering an area exceeding 10,000 square kilometers. It has identified over 200 high-risk and over 500 medium-risk hazard points. Following ground surveys and expert assessments, over 90% of the high-risk hazard points identified by the system showed obvious signs of deformation, and some have been included in key monitoring and engineering remediation programs. The system provides scientific basis and technical support for slope disaster prevention and control, achieving significant social and economic benefits.

[0113] Taking the slope monitoring of a large hydropower station reservoir area as an example, the application effect of the method of this invention is illustrated. This reservoir area is located in the southwestern mountainous region, with dramatic topographic relief and a relative elevation difference exceeding 1000m. The rock mass is mainly composed of argillaceous and clastic rocks, with well-developed faults and folds, resulting in complex geological conditions. After the reservoir was filled, the reservoir bank slopes were affected by both the rise and fall of the reservoir water level and rainfall, leading to deformation signs on some slopes and a risk of landslides.

[0114] The monitoring area covers approximately 500 square kilometers, and 150 Sentinel-1 SAR images from January 2019 to December 2024 were acquired. After permanent scatterer interferometry processing, approximately 100,000 PS points were obtained, with a spatial density of approximately 200 points / square kilometer. The deformation rate field shows that the deformation rate in most areas is between -10 and 10 mm / year, indicating a relatively stable state. However, several anomaly zones were identified on the north and east banks of the reservoir, with deformation rates exceeding -30 mm / year and exhibiting an accelerating deformation trend.

[0115] Due to the high vegetation coverage in the monitoring area and the low coherence coefficient in some mountainous areas, the original deformation data was quite noisy. The spatial-temporal dual-domain adaptive enhancement algorithm of this invention was used to process the deformation data. After enhancement, the signal-to-noise ratio of the deformation signal was significantly improved, the spatial continuity of the deformation rate field was significantly improved, and the random fluctuations of the deformation time series were reduced. The enhancement effect was particularly significant in the low coherence region, allowing the subtle deformation signals that were originally submerged in noise to become apparent.

[0116] The monitoring area was divided into 1500 slope units based on a terrain segmentation algorithm, and a spatial adjacency graph of the slope units was constructed. A multi-scale spatiotemporal graph attention-Transformer coupled network was used to extract deformation features. The network was trained using 30 known landslides with historical records in the area as samples. After 50 rounds of iterative training, the network achieved an accuracy of 88% on the validation set.

[0117] A multimodal risk assessment network was constructed by integrating geological and topographical factors and environmental triggering factors. Geological and topographical factors were extracted from a 1:50,000 geological map and a 30m resolution DEM. Environmental triggering factors included daily water level data from five hydrological stations and daily rainfall data from ten meteorological stations in the reservoir area. The risk assessment results showed that 45 high-risk slope units were identified across the entire area, mainly distributed on the steep slopes of the north bank and the fault zone of the east bank; 120 medium-risk slope units were identified, with a relatively scattered distribution; and the remaining slope units were classified as low-risk.

[0118] On-site verification was conducted on the identified high-risk slope units. Of these, 38 units showed obvious signs of deformation, including slope cracks, bulging, and localized landslides, validating the accuracy of the system's identification. Two units, while not showing obvious deformation, had poor slope structure and geological conditions, and experts assessed that they indeed posed potential risks. Five units showed no obvious anomalies on-site; analysis suggested these might be false positives due to InSAR measurement errors or changes in surface vegetation. The system assigned relatively low confidence scores to these units, ranging from 0.55 to 0.65, indicating the need for further monitoring and verification.

[0119] For areas with low confidence, a closed-loop parameter optimization mechanism was triggered. After three iterations of optimization, the weight coefficients of the signal enhancement module and the network parameters of the feature extraction module were adjusted. The confidence scores of these areas improved to above 0.75, and some units that were originally identified as high-risk were adjusted to medium-risk after optimization, which better matched the actual situation on site. Closed-loop optimization significantly improved the stability and accuracy of the system in complex scenarios.

[0120] The system also predicted the deformation trends of high-risk slope units. Based on historical deformation time series and environmental factor prediction data, the deformation development trend for the next six months was predicted. The prediction results indicate that eight high-risk slope units may enter an accelerated deformation phase in the future, requiring enhanced monitoring and preventive measures. The prediction information provides important basis for risk management and emergency preparedness.

[0121] Based on the monitoring and assessment results, relevant departments took targeted measures for high-risk slope units, including installing ground monitoring equipment, restricting personnel activities, evacuating residents within the threat area, and implementing engineering remediation. During the flood season of 2024, small-scale landslides occurred in three high-risk slope units, but due to early warnings and appropriate measures, no casualties or significant property damage occurred. This application case fully demonstrates the application value of the method of this invention in actual slope monitoring and disaster prevention.

[0122] The above embodiments of the present invention are merely preferred technical solutions and do not constitute a limitation on the scope of protection of the present invention. Those skilled in the art, inspired by the technical solutions of the present invention, can make various modifications, substitutions, or changes, all of which should be included within the scope of protection of the present invention. For example, other time-series InSAR processing techniques such as small baseline set techniques can be used; other deep learning network architectures such as recurrent neural networks and generative adversarial networks can be used; more data sources such as optical remote sensing images and ground monitoring data can be fused; and it can be extended to monitoring other types of geological hazards such as ground subsidence and ground fissures. Any implementation that conforms to the technical concept and innovation of the present invention should be considered an equivalent embodiment of the present invention.

Claims

1. A method for intelligent interpretation of slope deformation based on InSAR time-series images, characterized in that, include: InSAR image data acquisition and preprocessing steps: acquire multi-temporal SAR images covering the target monitoring area, and process them using temporal interferometry to generate a deformation rate field and a cumulative deformation time series. The deformation rate field contains the average deformation rate value and spatial coordinate information of each pixel, and the deformation time series contains the deformation at each time step. Deformation signal adaptive enhancement steps: For InSAR deformation data in low-coherence mountainous areas, a spatial-temporal dual-domain adaptive enhancement algorithm is used to enhance the deformation signal. In the spatial domain, the spatial domain adaptive enhancement weight is calculated based on the coherence of neighboring pixels and the deformation gradient. In the temporal domain, the temporal domain adaptive enhancement weight is calculated based on the smoothness and trend of the time-series deformation curve. The spatial domain enhancement weight and the temporal domain enhancement weight are fused to generate a comprehensive enhancement weight. The original deformation data is weighted and filtered using the comprehensive enhancement weight to generate an enhanced deformation rate field and deformation time series. Multi-scale spatiotemporal feature extraction steps: Based on the enhanced deformation data, a spatial adjacency graph of slope units is constructed. A multi-scale spatiotemporal graph attention-Transformer coupled network is used to extract deformation features. The graph attention network layer captures the spatial dependencies between adjacent slope units, and the multi-head temporal Transformer layer captures the long-range temporal dependencies of the deformation time series. Deformation features at different spatiotemporal scales are fused through a multi-scale feature pyramid to output a multi-scale spatiotemporal feature tensor. Multimodal risk intelligent assessment steps: Construct a multimodal risk assessment network, integrate the multi-scale spatiotemporal feature tensors, geological and topographic factors and environmental triggering factors, identify deformation pattern types and predict deformation development trends through a deep semantic understanding network, calculate the slope hazard risk level based on deformation rate, deformation acceleration, spatial clustering characteristics and deformation pattern type, and output the confidence index of risk assessment.

2. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 1, characterized in that, The adaptive enhancement step for the deformation signal includes the following spatial-temporal dual-domain adaptive enhancement algorithm: Calculate the spatial domain enhancement weight for each pixel, which is positively correlated with the average coherence coefficient of neighboring pixels and negatively correlated with the variance of the neighborhood deformation gradient. The temporal enhancement weight of each pixel is calculated. This weight is positively correlated with the goodness of linear fit of the temporal deformation curve and negatively correlated with the deformation abrupt change detection index. The spatial domain enhancement weights and temporal domain enhancement weights are used to generate a comprehensive enhancement weight through an adaptive fusion algorithm. The fusion coefficient is dynamically adjusted according to the coherence level of the local region. When the local average coherence coefficient is lower than the coherence coefficient threshold, the proportion of spatial domain weight is increased. When the local average coherence coefficient is higher than or equal to the coherence coefficient threshold, the proportion of temporal domain weight is increased.

3. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 1, characterized in that, The multi-scale spatiotemporal feature extraction step includes a multi-scale spatiotemporal graph attention-Transformer coupled network comprising: The graph attention network layer constructs a dynamic attention graph based on the spatial adjacency relationship and deformation correlation between slope units. It calculates the spatial dependency strength between slope units in different neighborhoods through a multi-head attention mechanism and outputs a spatial feature matrix. The multi-head time-series Transformer layer encodes the deformation time series of each slope unit and captures deformation patterns at different time scales through a multi-head self-attention mechanism, including short-term fluctuations, medium-term trends and long-term evolution characteristics, and outputs a time-series feature matrix. The multi-scale feature pyramid module extracts deformation features at different spatiotemporal scales and achieves multi-scale feature fusion through bidirectional feature transfer from bottom to top and top to bottom, generating a multi-scale spatiotemporal feature tensor that includes fine-grained local deformation and coarse-grained overall trend.

4. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 1, characterized in that, In the aforementioned multimodal risk intelligent assessment step, the multimodal risk assessment network includes: The deformation pattern recognition branch classifies deformation patterns based on multi-scale spatiotemporal feature tensors using deep convolutional neural networks, identifying linear creep, accelerated deformation, periodic deformation, and abrupt deformation. The geological and topographic factor fusion branch extracts geological and topographic factors such as slope, aspect, elevation, lithology and geological structure. Through the feature embedding layer, the geological and topographic factors are encoded into high-dimensional feature vectors and deeply fused with deformation features. The environmental triggering factor fusion branch acquires environmental triggering factors such as rainfall, reservoir water level changes, seismic activity, and human engineering activities, analyzes the time-lag correlation between environmental factors and deformation acceleration, and identifies deformation triggering mechanisms. The risk level prediction module integrates deformation pattern type, geological and topographic susceptibility, and environmental trigger sensitivity, and uses an ensemble learning algorithm to output the risk level. At the same time, it calculates the confidence index based on the consistency of the prediction results and the historical verification accuracy.

5. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 1, characterized in that, It also includes a closed-loop parameter optimization step: based on the risk assessment results and confidence index, the parameter optimization mechanism is triggered. When the confidence level is lower than the lower confidence level threshold, the enhancement weight calculation parameters in the deformation signal adaptive enhancement step and the attention weight in the multi-scale spatiotemporal feature extraction step are adjusted in reverse. Through iterative optimization, the confidence level of the risk assessment is increased to above the upper confidence level threshold. The automatic extraction step for slope hazard areas involves extracting the spatial boundaries of high-risk and medium-risk areas based on risk assessment results, and using morphological operations to connect discrete risk pixels to form continuous hazard areas. The deformation prediction step involves using a temporal Transformer network to predict the deformation of the identified hazard areas in the future, and combining the predicted data of environmental factors to give the deformation development trend. The early warning threshold is dynamically adjusted based on historical monitoring data and prediction accuracy statistics to achieve adaptive early warning.

6. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 1, characterized in that, In the InSAR image data acquisition and preprocessing steps: Long-term SAR images are jointly processed using permanent scatterer interferometry or distributed scatterer interferometry to obtain high-density deformation monitoring points. Atmospheric delay correction and orbital error correction are performed on the interferometric phase, and a spatiotemporal filtering method is used to separate the atmospheric phase and the deformation phase. The multi-temporal SAR images include ascending orbit images and descending orbit images. The line-of-sight deformation of ascending and descending orbits is decomposed into vertical deformation and horizontal deformation. Ascending and descending orbit fusion technology is used to improve the three-dimensional accuracy of deformation monitoring.

7. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 3, characterized in that, The construction of the slope unit spatial adjacency graph includes: The target monitoring area is divided into several slope units based on the terrain segmentation algorithm. Each slope unit has relatively homogeneous terrain and geological conditions. The connection edges between nodes are established based on the spatial adjacency relationship of the slope elements. The weight of the edge is determined by the Euclidean distance and the deformation correlation coefficient between the slope elements. A dynamic graph structure is constructed, and the edge weights are dynamically updated according to the deformation evolution process to strengthen the connection between slope units with high deformation correlation.

8. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 4, characterized in that, The risk level prediction module uses the following judgment criteria: When the deformation rate is greater than the first rate threshold, which is determined based on the slope geological conditions and historical deformation data, the deformation acceleration is greater than the acceleration threshold, which is determined based on the rate of change of the deformation time series, and the deformation pattern is identified as an accelerated deformation type, it is judged as a high-risk level. When the deformation rate is between the first rate threshold and the second rate threshold, and the deformation exhibits linear creep or periodic deformation characteristics, and there are environmental triggering factors, it is judged to be of medium risk level. When the deformation rate is less than the second rate threshold, the deformation curve is stable, and there are no obvious environmental triggering factors, it is judged as a low-risk level.

9. The intelligent interpretation method for slope deformation based on InSAR time-series images according to claim 5, characterized in that, The closed-loop parameter optimization steps include: The confidence level monitoring mechanism monitors the confidence level indicators of risk assessment in real time, and triggers the parameter optimization process when the confidence level falls below the lower confidence level threshold. Gradient backpropagation optimization calculates the parameter gradient based on the confidence loss function, and adjusts the spatial domain weight coefficients and temporal domain weight coefficients in the deformation signal adaptive enhancement step, as well as the number of attention heads and the number of feature pyramid layers in the multi-scale spatiotemporal feature extraction step. The iteration convergence judgment is performed. After parameter adjustment, the deformation signal enhancement, feature extraction and risk assessment process is re-executed. If the confidence level is increased to above the upper confidence level threshold, the optimization is terminated. Otherwise, the iteration continues until the maximum number of iterations is reached.

10. A slope deformation intelligent interpretation system based on InSAR time-series imagery, used to implement the slope deformation intelligent interpretation method based on InSAR time-series imagery as described in claim 9, characterized in that, include: The InSAR image data acquisition and preprocessing module is used to acquire multi-temporal SAR images and generate deformation rate fields and deformation time series. The deformation signal adaptive enhancement module is used to perform spatial-temporal dual-domain adaptive enhancement processing on deformation data in low-coherence regions. The multi-scale spatiotemporal feature extraction module is used to extract the spatiotemporal features of deformation through a multi-scale spatiotemporal graph attention-Transformer coupled network; The multimodal risk intelligent assessment module is used to integrate deformation characteristics, geological and topographical factors, and environmental triggering factors to predict risk levels and calculate confidence levels. The closed-loop parameter optimization module is used to reverse-optimize the key parameters of the preceding module based on the risk assessment confidence level, so as to achieve closed-loop collaborative optimization. The output of the deformation signal adaptive enhancement module serves as the input of the multi-scale spatiotemporal feature extraction module, and the output of the multi-scale spatiotemporal feature extraction module serves as the input of the multi-modal risk intelligent assessment module. The confidence output of the multi-modal risk intelligent assessment module triggers the closed-loop parameter optimization module, and the parameter adjustment signal of the closed-loop parameter optimization module is fed back to the deformation signal adaptive enhancement module and the multi-scale spatiotemporal feature extraction module.

Images