CNN and SBAS-InSAR technology fused earth surface deformation detection and classification method and system

Abstract

The invention relates to an earth surface deformation detection and classification method and system fusing CNN and SBAS-InSAR technologies, and the method comprises the steps: obtaining high-precision deformation data through a small baseline subset and an interference synthesis radar technology (SBAS-InSAR), constructing a CNN classification frame through combining PCA-enhanced optical satellite images, elevation, soil physical parameters, geological landform maps and other multi-source geographic information, and carrying out the detection and classification of the earth surface deformation through the CNN classification frame and the SBAS-InSAR technology. And identification of deformation types such as landslide, settlement and lifting is realized. The process comprises multi-source satellite data acquisition and preprocessing (including InSAR sight velocity decomposition and optical / topographic feature extraction), generation of a preliminary deformation label based on a classification algorithm of a gradient and a deformation velocity threshold, and recognition of a deformation type through a multilayer CNN model. According to the method, the deformation mode is automatically clustered and analyzed in a large-range area, the accuracy and detail performance of deformation type recognition are remarkably improved, the accuracy rate reaches 93%, the method is superior to a traditional method, a stable and extensible tool platform is provided for geological disaster monitoring, and positive risk management of landslide or land subsidence prone areas is assisted.

Description

(I) Field of the Art

[0001] The application relates to a surface deformation detection and classification method and system combining a CNN and an SBAS-InSAR technology, and belongs to the field of satellite remote sensing applications and radar image processing and AI application in surface deformation detection. (II) BACKGROUND

[0002] Landslides, ground subsidence, and surface uplift are ground deformation phenomena that pose a significant threat to infrastructure and communities, especially in areas with complex geology or frequent human activities. It is challenging to monitor and distinguish these deformation types over a large area. Traditional methods mainly rely on two approaches: one is to use interferometric synthetic aperture radar (InSAR) analysis alone to detect surface displacement, and the other is to use machine learning combined with static factors (such as slope, geological conditions) to assess disaster susceptibility. Although interferometric SAR technology (such as the SBAS method) can monitor millimeter-level ground movement in real time, the single-dimensional range velocity data it provides need further analysis to determine the displacement properties (for example, to determine whether the observed displacement is caused by a landslide or regional subsidence). On the other hand, existing research attempts to integrate multi-source data (such as optical images, digital elevation models, and geological maps) to perform machine learning classification of geological disasters to predict disaster susceptibility. For example, a random forest (RF) classifier is used in combination with multi-source remote sensing data to identify ground displacement types. Although this method helps to understand the influencing factors, it may lack the time depth of InSAR or the predictive ability of deep learning, resulting in limited accuracy.

[0003] Therefore, there is an urgent need for an innovative method that combines the advantages of InSAR technology in deformation measurement with the pattern recognition capabilities of modern deep learning. The present application proposes a new method that innovatively combines satellite radar interferometric measurement technology with deep learning algorithms to develop a new solution for ground deformation monitoring. By integrating InSAR technology for surface displacement measurement and using a convolutional neural network (CNN) classification algorithm, combined with multi-source geographic information data (including satellite radar data, optical images, and terrain and soil data auxiliary spatial layers), the system achieves automatic detection of ground deformation types and accurate identification of surface deformation types (such as landslides, surface uplift, and subsidence) in a large area. By using a convolutional neural network, the system can learn the complex non-linear relationship between input features and deformation types, and its accuracy is expected to exceed that of traditional techniques, thereby generating more reliable disaster maps. The new surface deformation classification technology based on CNN+SBAS proposed by the present application effectively meets the application requirements of automatic and dynamic monitoring of underlying disasters. (III) SUMMARY

[0004] 1. A surface deformation detection and classification method and system combining a CNN and an SBAS-InSAR technology, characterized by comprising the following steps:

[0005] (1) Acquire multi-temporal SAR images of the study area and use Small Baseline Set Interferometric Synthetic Aperture Radar (SBAS-InSAR) technology to process the images to obtain ground deformation measurements and line-of-sight (LOS) displacement velocity.

[0006] (2) Obtain auxiliary geospatial data for the region, including optical satellite imagery, digital elevation data, and thematic layers describing soil and geological conditions;

[0007] (3) The optical image was processed by the principal component analysis (PCA) method to extract the principal component feature map and extract the terrain attributes, including slope and aspect, from the elevation data.

[0008] (4) Register and integrate the deformation measurement values ​​with the auxiliary data to form a multi-dimensional feature set for each location in the region;

[0009] (5) A rule-based classification algorithm is applied to classify deformation measurements and terrain attributes, wherein the algorithm uses a predefined deformation rate threshold V. min and slope standard S max Assign an initial deformation category to each location selected from the following categories: stable, subsidence, uplift, landslide, and collapse;

[0010] (6) Using the multidimensional feature set as input and the initial deformation category as the target output, train a convolutional neural network (CNN) classifier; learn to classify ground deformation types based on multi-source data combination through the network, and deploy the trained CNN to the entire area for deformation type recognition, thereby generating a classification deformation map that can distinguish different deformation types.

[0011] 6. In step (1) above, when processing synthetic aperture radar (SAR) images, deformation velocity maps in two line-of-sight directions are obtained by simultaneously using ascending and descending orbit datasets. Using the known geometric relationship between the line-of-sight direction and ground displacement components, the line-of-sight velocity at each location is decomposed into vertical and horizontal deformation components, thereby obtaining three-dimensional ground deformation measurement data; the three-dimensional velocity V in the line-of-sight direction... LOS East-west deformation velocity component v East-West North-south deformation velocity component v North-South and vertical direction v Vertical The deformation velocity components are composed of:

[0012]

[0013] Where satellite represents the serial number of the satellites passing through the radar orbit in different orbits, θ is the satellite incident angle, and α is the heading angle of the radar orbit; the line-of-sight velocity of the above matrix equation is projected onto the vertical plane, and the north-south component is deliberately excluded. Assuming that the incident angle is consistent throughout the observation area, the calculation is optimized by simplifying the influence of the horizontal component:

[0014]

[0015] Where V asc V represents the velocity component of the ascending orbit. desc The velocity components of the descending orbit are represented; the velocities of the ascending and descending orbits are integrated using multi-geometric data fusion technology to construct the deformation velocity field as follows:

[0016]

[0017] Among them, V v and V h Vertical and horizontal velocity components, respectively.

[0018] 2. The specific method of the rule-based initial classification algorithm in step (5) above is as follows: Based on the acquired deformation measurement data and terrain information, a rule-based ground deformation clustering algorithm is used to generate an initial deformation type classification; this algorithm introduces a new threshold standard based on domain knowledge: for example, setting a minimum deformation rate threshold V. min This is used to distinguish significant motion from noise interference; the system calculates the total three-dimensional velocity amplitude at each location. and the benchmark value V min A comparison is performed; simultaneously, the local slope angle α extracted based on the digital elevation model (DEM) is compared with the characteristic threshold α. min and higher threshold S max A comparison will be made; based on these criteria, each location will be initially classified into one of several deformation categories:

[0019] o If the total speed V t Below the minimum threshold V min V t <V min If the condition is met, the region is classified as stable with no significant deformation.

[0020] oIf V t ≥V min Furthermore, the slope is relatively gentle, α < α min Therefore, it can be inferred that the motion is mainly vertical: the positive vertical component v U A value greater than 0 indicates ground uplift, while the negative vertical component v U <0 indicates ground subsidence;

[0021] oIf Vt ≥V min , with a moderate to steep slope, a min ≤a max This area is classified as prone to landslides because of the sufficient slope and movement indicating a landslide process;

[0022] o If the slope is very steep, a ≥ S max , the area is marked as a rockfall or cliff collapse risk, even small movements on near-vertical slopes can indicate a potential collapse;

[0023] o These rules can be adjusted according to local geological conditions, but their role is to cluster deformation data into meaningful classes: landslides, subsidence, uplift, rockfall, stable, for training the CNN model.

[0024] 3. The CNN classifier of step (6) above comprises a multi-layer feedforward neural network characterized by:

[0025] • CNN classification architecture: the deep feedforward network adopted aims to learn the mapping from multi-source input features to deformation class labels determined by the algorithm above; the input layer receives a multi-dimensional feature vector for each geographic location, specifically including at least the InSAR-derived deformation indicators, one or more principal components of optical imagery, and the feature vectors of terrain and soil parameters; the input feature vector has 10-20 dimensions; the CNN is composed of multiple fully connected dense layers, using a nonlinear activation function ReLU and a dropout rate regularization layer to prevent overfitting; in the preferred design, the network contains an input layer followed by four progressively scaled hidden layers with nonlinear activation functions for learning intermediate representations of features; the output layer uses a softmax activation function to output probability values for each deformation type class; the CNN is trained by a supervised learning algorithm to minimize the classification error of the initial deformation class; wherein the CNN hidden layer contains at least one dense layer with about 200-250 neurons, followed by a dropout layer with a dropout rate of 10% to 20%, and this pattern is repeated in successive layers to reduce overfitting and improve generalization ability, thus forming a network architecture that can achieve high classification accuracy for multiple deformation types; finally, a 6-unit fully connected output layer classifies the results into one of the six categories: no deformation, subsidence, uplift, landslide, rockfall, and an optional additional secondary class, through the Softmax activation function;

[0026] • Training and model optimization: the core process of this method includes training a convolutional neural network (CNN) on a representative dataset of the target study area and validating its performance; the training data is obtained by sampling multi-source input layers at the target location and using the preliminary classification results as the target labels for each sample; following the standard machine learning process, including data standardization, division of training set / validation set, and multiple rounds of iterative optimization, which can use Adam or SGD optimizer; ensure model robustness through Dropout layers that randomly disable neurons, with an early layer dropout rate of about 0.2 and a later layer dropout rate of 0.1; use the early stopping strategy to terminate training when the validation loss stops declining and retain the best model weights; through this process, the CNN can break through the limitations of the initial rule labels and re-analyze the velocity grid and auxiliary data to optimize the classification effect;

[0027] • Automated classification and output: after completing the training, the CNN model will be deployed to classify the ground deformation types in the entire target area; for each pixel or unit area, the model generates probability values or scores for each category by integrating multi-source features, and finally determines the class with the highest probability as the predicted deformation type; the thematic deformation map generated in this way can clearly label landslide, subsidence, uplift and other areas at high spatial resolution;

[0028] • Model verification and accuracy evaluation: further includes comparing the classification output of the trained CNN classifier with known true values or results of independent methods to verify, comparing the accuracy or prediction performance with baseline machine learning methods that do not use deep neural networks, and evaluating the parameters using overall classification accuracy and area under the ROC curve (AUC) to quantify.

[0029] 4. A method and system for land surface deformation detection and classification by fusing CNN and SBAS-InSAR technology, characterized in that it comprises the following modules: a data acquisition module for receiving and storing SAR images and auxiliary geospatial data of a region; an InSAR processing module for performing SBAS interferometric analysis on the SAR images and outputting deformation velocity maps; a data integration module for extracting features from optical principal component analysis (PCA) components and auxiliary data including terrain attributes, and integrating all features into a unified geospatial grid; a classification module comprising a processor for executing instructions to complete two tasks: (1) applying threshold rules to deformation velocity and slope data to generate preliminary classification labels; and (2) training and applying a convolutional neural network (CNN) for final deformation prediction at each grid location; and an output module for generating visual deformation maps and reports marking locations determined by the CNN as landslides, subsidence, uplift, rockfalls, or stable ground; the system supports real-time updating of classification results and automatically updates classification information when new SAR or satellite data is obtained, thereby constructing an automated and continuously learning land deformation monitoring tool; the CNN in the classification module of the system is pre-trained on historical data from the region and is incrementally updated with new data, enabling real-time or near-real-time risk classification without the need to retrain from scratch for each update; the output module of the system further calculates and displays confidence levels for each classified region and allows user interaction to query underlying multi-source data values that lead to a particular classification, thereby facilitating interpretable results for end users.

[0030] 5. Compared with the prior art, the advantages of the present application are:

[0031] (1) Improved accuracy and reliability of land surface deformation detection: SBAS-InSAR technology can obtain large-scale, high-precision land surface deformation time series data, but is susceptible to noise (such as atmospheric interference, terrain residual error); CNN can automatically extract non-linear features in the data through deep learning, effectively filter out noise and enhance weak deformation signals, and improve the stability of the detection results.

[0032] (2) Intelligent classification of land surface deformation types: traditional SBAS-InSAR is mainly used for deformation quantification and cannot directly distinguish the causes of deformation (such as geological structure movement, groundwater exploitation, and engineering construction); after fusing CNN, its powerful image classification ability can be utilized in combination with the spatial distribution characteristics (such as range, rate, and morphology) of deformation to achieve automatic identification and classification of different types of deformation, improving analysis efficiency.

[0033] (3) Extend the application scenarios of the technology: single SBAS-InSAR is limited in complex terrain (such as urban dense area, mountainous area) or low coherence area; CNN can enhance the adaptability of the technology in low-quality data scenarios through data enhancement and feature migration, and expand the application range of ground deformation monitoring (such as urban subsidence, landslide warning, etc.).

[0034] The innovation of the present application mainly lies in:

[0035] (1) Methodological innovation of technology fusion: convert the time series deformation data (SBAS-InSAR output) into "feature image" (such as deformation rate map, cumulative deformation map) suitable for CNN processing, and build an end-to-end model of "time series deformation-spatial feature-intelligent classification", which breaks through the limitation of traditional InSAR technology only for quantitative analysis, and realizes the leap from "detection" to "understanding".

[0036] (2) Intelligent upgrade of feature extraction: compared with traditional artificial design features (such as deformation gradient, spatial correlation), CNN can learn deep features (such as nonlinear deformation pattern, multi-scale spatial correlation) independently, which can more accurately capture the essential differences of different types of deformation and improve the generalization ability of classification.

[0037] (3) Practical innovation of system integration: build an automatic system integrating "data preprocessing (SBAS-InSAR)-feature learning (CNN)-deformation detection and classification-result visualization", reduce manual intervention, improve the efficiency and engineering practicability of large-scale ground deformation monitoring, and provide more direct decision support for geological disaster warning, urban planning, etc. Through the fusion of CNN and SBAS-InSAR, this method not only retains the high-precision deformation measurement capability of InSAR, but also endows it with new functions of intelligent analysis and classification, realizing technology cooperation and performance leap in ground deformation monitoring field.

[0038] The present application fuses advanced satellite data processing technology and deep learning algorithm to create a powerful tool for ground deformation mapping. It provides integrated processing flow to effectively overcome the shortcomings of existing technologies and accurately classify deformation types. The system can be run as a software tool or platform, which can generate the latest risk map in real time by integrating conventional satellite data, and help relevant departments effectively prevent and control landslide and ground subsidence risks in prone areas.

[0039] (I) Drawing description

[0040] Figure 1 is the technical process of the present application.

[0041] Figure 2is a research example area (Bolivia Rochariver basin) classification deformation map. Different colors represent different deformation types identified by the invention (such as landslide area, subsidence area, uplift area, stable area, etc.).

[0042] (II) Specific implementation method

[0043] In order to better illustrate the surface deformation detection and classification method and system of the present application, the applicability test of the present application in the Rochariver basin of Cochabamba, Bolivia is carried out, and good results are obtained. The specific implementation method is as follows:

[0044] 1. Data collection and arrangement

[0045] Multi-source geospatial data collection: This method first needs to collect a complete set of geospatial data for the target area. By obtaining satellite synthetic aperture radar (SAR) data (such as Sentinel-1 radar images) of the target area within a certain time period, the foundation for subsequent InSAR processing is laid. Taking a typical case as an example, time series data of Sentinel-1 satellite images covering the Rochariver basin in the Cochabamba region of Bolivia over the years are obtained. At the same time, optical satellite image data (such as Sentinel-2 multispectral satellite and / or high-resolution PlanetScope satellite images) of the region are also collected. In order to fully capture relevant environmental background information, the following supplementary data are also obtained:

[0046] · Digital Elevation Model (DEM): Provides elevation and can calculate the slope at each location.

[0047] · Soil and geology maps: Include data on soil types, soil physical properties (such as density, cohesion, etc.), lithology, fault lines, and geomorphic units (such as alluvial plains, slopes).

[0048] · Land cover type images: Multispectral optical data that can show vegetation cover, moisture, and other surface conditions related to ground stability.

[0049] All data sets are projected into a common geographic coordinate system and spatially aligned. If necessary, satellite images will be resampled to a uniform resolution (such as 10-meter or 30-meter grid) to balance detail and computational efficiency. Table 1 lists the representative input data sources in this embodiment. All these layers are aligned on a per-pixel basis, forming a multi-dimensional feature vector for each pixel.

[0050] Optical image preprocessing (principal component analysis enhancement): To address the high dimensionality of multi-band optical images and potential noise, this method uses principal component analysis (PCA) to extract key features. For example, multi-band images taken by Sentinel-2 satellite and PlanetScope satellite in a relevant time period can be overlaid and PCA is used to extract several principal component images that capture the main variability features (e.g., vegetation contrast, moisture, or soil exposure). These PCA-extracted bands serve as enhanced input features that highlight important spatial patterns while effectively reducing redundant information. In a specific implementation, the first three principal components of the optical images are used as input.

[0051] PC i = a i1 X1+a i2 X2+…+a in X n

[0052] SBAS-InSAR processing: Using SAR images, a SBAS interferometric processing workflow is performed to calculate ground deformation, including:

[0053] (1) Interferogram formation: SAR image pairs with small temporal and spatial (orbital) baselines are selected and difference interferograms are computed, which represent phase changes due to ground motion.

[0054] (2) Time series analysis: Displacement time series for each pixel are estimated by inverting the network of interferograms, typically using singular value decomposition or least squares known from SBAS techniques.

[0055] (3) Velocity map generation: Average LOS velocity maps are derived from the time series. This results in a raster image where each pixel value represents the estimated deformation rate (mm / year) in the satellite line-of-sight direction.

[0056] (4) Error correction and filtering: Atmospheric corrections are applied (if external atmospheric data or filters are available) and low-coherence areas are removed to improve reliability. A threshold of temporal coherence can be used to mask pixels with unstable phase behavior.

[0057] When ascending and descending track data are acquired simultaneously, the system generates two line-of-sight velocity maps. By solving the system of equations defined by the line-of-sight geometry, these velocity maps can be decomposed into vertical and horizontal components of motion. The output of this step includes a vertical deformation velocity map (v U ) and a horizontal (east-west) deformation velocity map (v E ) for the area. In regions dominated by north-south motion, although the north-south component can also be derived, the focus of the analysis is still primarily on the vertical and east-west components of motion.

[0058] The final processed dataset for each pixel contains the following features: e.g. [Principal Component 1 (PC1), Principal Component 2 (PC2), Principal Component 3 (PC3), Slope, Elevation, Soil Type, Soil v U v E Density, Lithology, etc.]. These numerical or categorical values represent the characteristic attributes of that location and will be used as input parameters for the classification step.

[0059] Table 1. Multi-source data input for ground deformation classification

[0060]

[0061] 2. Ground deformation clustering algorithm (preliminary classification)

[0062] After the data preparation is completed, the method first applies a rule-based algorithm for the preliminary classification of ground deformation. This step transforms the raw deformation measurement data and topographic data into a classification label to identify the possible type of deformation at each location. The specific algorithm can be summarized in the following steps:

[0063] • Compute total deformation: For each pixel, compute the total 3D deformation velocity V t (total) from the resolved components (if not fully resolved, use the LOS velocity magnitude). For example, if only the east and vertical components are considered, then represents the velocity of the ground movement, independent of the direction.

[0064] • Apply velocity threshold (V min ): Define a minimum significant velocity threshold V min (5 mm / year, for example, based on known InSAR sensitivity and noise levels). If V t < V min , the pixel is labeled as class 0: stable / no significant deformation. Any movement detected in these areas is negligible or within the error range.

[0065] • Classify by slope and vertical motion: If V t ≥ V min , significant deformation is present, then classify using the slope angle a and the vertical velocity v U :

[0066] o When the slope is set a < a min , a min 10°, the terrain is relatively flat. In this case, significant deformation is less likely to form landslides (which require steeper slopes). Instead, this deformation is more likely to be caused by ground vertical displacement. Therefore, if v U > 0 (ground uplift), classify as class 1: ground uplift; if v U<0 (surface subsidence), it is classified as the second type: ground subsidence. For example, at the bottom of a flat valley, ground subsidence can indicate the presence of groundwater exploitation or soil consolidation phenomena.

[0067] o When the slope reaches a = a min , such as 10° or above, the slope of the region is sufficient to trigger landslide activity. At this time, it should be marked as landslide-related deformation. If the slope is extremely steep a = a max (such as 62°), the failure mode can be different (such as cliff collapse or rockfall), which can be classified as the third type: rockfall / steep slope collapse. Other cases (moderate steepness) are marked as the fourth type: landslide. The difference lies in: the fourth type (landslide) usually involves the movement of soil or weathered rock along the slope surface, while the third type (collapse) is more of a free-falling or toppling phenomenon on a nearly vertical rock wall.

[0068] o In practical applications, the classification of such landslides needs to consider both the slope and the horizontal displacement direction: landslide movement usually has a horizontal component (v E or v N consistent with the direction of gravity), while pure subsidence / upsurge hardly produces horizontal displacement. This algorithm confirms the landslide classification by detecting whether the horizontal velocity is significant and roughly consistent with the direction of the steepest slope.

[0069] • Output preliminary labels: according to the above rules, all pixels with significant deformation are assigned preliminary labels (uplift, subsidence, landslide, collapse), and stable areas are marked as the corresponding type. These labels, based on simple physical principles, collectively form the initial deformation map. Figure 1 The execution flowchart of this decision logic in the complete process is shown.

[0070] This rule-based classification method, although with some approximation, as an innovative component, ensures that the next stage of Convolutional Neural Network (CNN) has a clear target category to learn. The threshold parameters a min , v min can be adjusted and calibrated according to different environmental conditions (for example, when local experience shows that a slope of ~8° may cause landslides, the threshold v min should be appropriately lowered). In the case of the study area, a min = 10°, rainfall v min = 5 mm / year, the range of these parameters has been proven to be appropriate. This algorithm effectively realizes the transformation of continuous geophysical data into classification categories by embedding expert knowledge (such as the typical range of landslide slopes) as parameters into the system. To approximately estimate the velocity direction, the surface gradient analysis method is used, which exhibits three-dimensional gradient decomposition characteristics. By performing point-by-point multiplication of the velocity value with the data points—based on their approximate alignment characteristics—the estimation accuracy is further optimized, as follows:

[0071]

[0072] 3. CNN model architecture and training

[0073] CNN input and output: The convolutional neural network (CNN) used in this method is configured as a classifier, which takes a multi-source feature vector at a certain location as input and outputs a probability distribution over deformation classes. In the described embodiment, the CNN input for each pixel contains the following features:

[0074] • Principal component analysis enhanced optical reflectance value (can capture environmental context);

[0075] • Terrain parameters (slope angle, elevation);

[0076] • Soil / geology indicators (may be encoded as numerical values or categories);

[0077] • InSAR-derived velocity (v U ,v E or possibly line-of-sight velocity and incidence angle);

[0078] • Any other relevant attributes.

[0079] The output result contains N classes. According to the previous steps, we define N = 5 classes in this case (stable, uplift, subsidence, landslide, collapse). Although an "unknown" or "unclassified" class can also be added, but the algorithm based on rules has covered all possibilities. In some cases, the sixth class may need to be introduced (for example, to distinguish different types of landslides or to distinguish shallow and deep displacements), but the main types are as described above. The final layer of the convolutional neural network uses a softmax activation function to generate confidence values for each class, thus obtaining a clear classification prediction result.

[0080] Network architecture design: The convolutional neural network (CNN) architecture is powerful enough in depth to capture complex patterns, while remaining lightweight to avoid overfitting, thanks to the structured nature of the input data. It should be noted that although this invention is called CNN, in this implementation the network operates on each pixel as a feature vector (rather than based on two-dimensional image blocks), so the proportion of convolution operations is very low; essentially, this model behaves like a deep multi-layer perceptron. (In other embodiments, spatial context can be captured by considering pixel neighborhoods to introduce convolution operations, but the focus of this invention is on pixel-level classification tasks.)

[0081] The structure of the network can be decomposed into several key components, each of which has a specific role in transforming the input data through successive layers of computation:

[0082] Table 2 shows an example architecture. The architecture contains an input layer equal to the number of input features (e.g., 15 features), followed by multiple fully connected (Dense) layers connected in sequence. A typical configuration is as follows:

[0083] • Dense Layer 1: contains 250 neurons with ReLU activation. This layer is used to expand the feature space and initiate learning of first-order combinations of features. (When the number of input features is M, this layer contains trainable parameters; if M = 15, the number of parameters is M x 250 + 250 = 4,000)

[0084] • Dropout rate for the first layer: dropout rate is 0.2. During training, 20% of the neurons in Dense Layer 1 are randomly deleted from each batch, which helps prevent overfitting and forces the network to not rely on any single feature or combination.

[0085] • Dense Layer 2: contains 200 neurons with ReLU activation. This layer (and subsequent layers) continues to learn more abstract representations. When the input comes from the 250 data from the previous layer, Dense Layer 2 updates its parameters (in practice, this method uses similar parameter sizes).

[0086] • Dropout rate for the second layer: 0.2;

[0087] • Dense Layer 3: 200 neurons, ReLU;

[0088] • Dropout rate for the third layer: 0.2;

[0089] • Dense Layer 4: 200 neurons, ReLU;

[0090] • Dropout rate for the fourth layer: dropout rate is 0.1 (in subsequent layers, the dropout rate can be slightly reduced to preserve more information as the model converges);

[0091] • Output layer: fully connected layer containing 5 neurons (or increased to 6 depending on the number of classes) with softmax activation. This final layer outputs a vector of length 5, where each element represents the probability that the input data belongs to each class. The system will select the class with the highest probability value as the prediction result.

[0092] Table 2. Example CNN architecture and parameters

[0093]

[0094] (Note: The above architecture can be adjusted (e.g., adjusting the number of layers or units) according to the scale and complexity of the dataset. For example, convolutional layers can be applied to gridded input data, or recurrent layers can be introduced when dealing with time series. The fully connected network structure described herein is particularly suitable for pixel-level classification tasks, which is the core advantage of this method.)

[0095] Training Process: The convolutional neural network (CNN) is trained using supervised learning. Each input feature vector X corresponds to a target class Y according to the preliminary labels provided in Section 2. The data is typically divided into a training set and a validation set. The model parameters are optimized by minimizing the categorical cross-entropy loss between the predicted class probabilities and the one-hot encoded true class. An optimizer such as Adam is used with an appropriate learning rate (e.g., 0.001) to ensure stable convergence. The training process is conducted for multiple epochs (e.g., 100-500 epochs), and if the validation set loss does not improve after a certain number of epochs, an early stopping strategy is employed to prevent overfitting. A batch training mode is used (batch size such as 32 or 64). If certain classes are underrepresented, class balancing techniques can be applied (e.g., when the number of uplift points is much less than the number of stable points, up-sampling or adjusting the weight of that class in the loss function can be used).

[0096] During training, the invention monitors the overall accuracy of the validation set as well as the precision / recall of each class to ensure that the model can effectively identify different types of deformation. Based on these performance data, hyperparameters such as the number of layers, units, dropout rates, etc. can be adjusted. In this example implementation, the selected architecture (Table 2) achieves a good balance, ensuring high accuracy while avoiding overfitting problems.

[0097] 4. Model Deployment and Results

[0098] After training is complete, the final convolutional neural network model will be used to infer the deformation class for all locations within the region (or new input data). By streaming the data through the input feature layers (Table 1) through the network, a classification map is generated.

[0099] Accuracy and Validation: In the case of the Roça River Basin, the CNN model was validated using known true data and compared to a baseline random forest model. The CNN model achieved an overall classification accuracy of approximately 93% in identifying the correct deformation type for test samples. In contrast, the random forest method using the same input features only achieved an AUC value of ~0.86 in a multi-class ROC analysis, and had lower classification accuracy. The CNN model also generated more spatially coherent terrain maps, while the random forest output had more noise, especially in complex mountainous regions. These results confirm that combining SBAS-InSAR measurement data with deep learning techniques can significantly improve the detection accuracy of landslides and subsidence areas.

[0100] Case Study: Output Map ( Figure 2 The landslide-prone areas in the Tunali Mountains region have been accurately marked. This region covers an area of ​​approximately 110 km². 2 Known for its frequent landslides, the model classifies most of its slopes as Level 4 (landslide risk). The Valle Alto region, located in the southeastern part of the basin, is characterized by intensive agriculture and heavy groundwater use; its valley floors are primarily classified as Level 2 (land subsidence), covering an area of ​​approximately 400 km². 2 This is consistent with surface subsidence caused by aquifer depletion. Meanwhile, some areas in the north are classified as Level 1 (crustal uplift), possibly related to groundwater recharge or tectonic uplift. The central area of ​​Cochabamba and other areas with minimal movement are classified as Level 0 (stable zone). Scattered Level 3 (rockfall) areas are found on the steepest cliffs of the Tunari Mountains, and these locations coincide with known rockfall sites.

[0101] In a demonstration application in the Cochabamba Barocha River basin of Bolivia, the method successfully identified known disaster areas and provided quantitative classifications. For example, the Tunari mountain area (~110km) was included. 2 The area is mainly designated as a landslide-prone zone, while the Alto Valley region (~400km) is also considered a landslide-prone area. 2 This area was classified as a subsidence-prone zone, which highly aligns with field observations. Accuracy assessments show that the CNN model's classification accuracy is significantly higher than the benchmark random forest method. This demonstrates the improved reliability of the proposed method.

[0102] Finally, it should be noted that the above are merely preferred embodiments and technical principles of the present invention. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made to the present invention without departing from its scope of protection. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, the present invention may also include many other equivalent embodiments, the scope of which is determined by the scope of the appended claims.

Claims

1. A method and system for detecting and classifying surface deformation by integrating CNN and SBAS-InSAR technologies, characterized in that... Includes the following steps: (1) Acquire multi-temporal SAR images of the study area and use Small Baseline Set Interferometric Synthetic Aperture Radar (SBAS-InSAR) technology to process the images to obtain ground deformation measurements and line-of-sight (LOS) displacement velocity. (2) Obtain auxiliary geospatial data for the region, including optical satellite imagery, digital elevation data, and thematic layers describing soil and geological conditions; (3) The optical image was processed by the principal component analysis (PCA) method to extract the principal component feature map and extract the terrain attributes, including slope and aspect, from the elevation data. (4) Register and integrate the deformation measurement values ​​with the auxiliary data to form a multi-dimensional feature set for each location in the region; (5) A rule-based classification algorithm is applied to classify deformation measurements and terrain attributes, wherein the algorithm uses a predefined deformation rate threshold V. min and slope standard S max Assign an initial deformation category to each location selected from the following categories: stable, subsidence, uplift, landslide, and collapse; (6) Using the multidimensional feature set as input and the initial deformation category as the target output, train a convolutional neural network (CNN) classifier; learn to classify ground deformation types based on multi-source data combination through the network, and deploy the trained CNN to the entire area for deformation type recognition, thereby generating a classification deformation map that can distinguish different deformation types.

2. The method according to claim 1, wherein in step (1), when processing synthetic aperture radar (SAR) images, deformation velocity maps of two line-of-sight directions are obtained by simultaneously using ascending and descending orbit datasets, and the line-of-sight velocity at each location is decomposed into vertical and horizontal deformation components using the known geometric relationship between the line-of-sight direction and the ground displacement components, thereby obtaining three-dimensional ground deformation measurement data; the three-dimensional velocity V in the line-of-sight direction LOS East-west deformation velocity component V East-West North-south deformation velocity component V North-South and vertical direction V Vertical The deformation velocity components are composed of: Among them, satellite N The sequence number represents the satellites passing through the radar orbit, θ is the satellite incident angle, and α is the heading angle of the radar orbit. The line-of-sight velocity from the above matrix equation is projected onto the vertical plane, with the north-south component intentionally excluded. Assuming a consistent incident angle across the entire observation area, the calculation is optimized by simplifying the influence of the horizontal component. Among them, V asc V represents the velocity component of the ascending orbit. desc The velocity components of the descending orbit are represented; the velocities of the ascending and descending orbits are integrated using multi-geometric data fusion technology to construct the deformation velocity field as follows: Among them, V v and V h Vertical and horizontal velocity components, respectively.

3. The method of claim 1, wherein the rule-based initial classification algorithm in step (5) is specifically implemented as follows: based on the acquired deformation measurement data and terrain information, a rule-based ground deformation clustering algorithm is used to generate an initial deformation type classification; this algorithm introduces a novel threshold standard based on domain knowledge: for example, setting a minimum deformation rate threshold V. min This is used to distinguish significant motion from noise interference; the system calculates the total three-dimensional velocity amplitude at each location. and the benchmark value V min A comparison is performed; simultaneously, the local slope angle α extracted based on the digital elevation model (DEM) is compared with the characteristic threshold α. min and higher threshold S max A comparison will be made; based on these criteria, each location will be initially classified into one of several deformation categories: If the total velocity V t Below the minimum threshold V min V t <V min If the condition is met, the region is classified as stable with no significant deformation. If V t ≥V min Furthermore, the slope is relatively gentle, α < α min Therefore, it can be inferred that the motion is mainly vertical: the positive vertical component v U A value greater than 0 indicates ground uplift, while the negative vertical component v U <0 indicates ground subsidence; If V t ≥V min The slope is moderate to steep, α min ≤α max The area was classified as a landslide-prone zone because sufficient slope and movement indicated the presence of a landslide process.​ If the slope is very steep, α ≥ S max If the area is at risk of rockfall or cliff collapse, even small movements on a near-vertical slope can indicate a potential collapse. These rules can be adjusted according to local geological conditions, but their purpose is to cluster deformation data into meaningful categories: landslides, subsidence, uplift, rockfalls, and stability, which are then used to train CNN models.

4. The method of claim 1, wherein the CNN classifier in step (6) comprises a multi-layer feedforward neural network, characterized by: • CNN classification architecture: The deep feedforward network used aims to learn the mapping relationship between multi-source input features and deformation category labels, which are determined by the algorithm described above; the input layer receives a multi-dimensional feature vector for each geographic location, specifically including at least the InSAR derived deformation index, one or more principal components of optical imagery, and feature vectors of terrain and soil parameters. The input feature vector has 10-20 dimensions; the CNN consists of multiple fully connected dense layers, employing the non-linear activation function ReLU and dropout rate regularization layers to prevent overfitting; in the preferred design, the network includes an input layer followed by four progressively smaller hidden layers with non-linear activation functions to learn intermediate representations of the features; the output layer uses the softmax activation function to output probability values ​​for each deformation type category; the CNN is trained using a supervised learning algorithm to minimize the classification error of the initial deformation category; the CNN hidden layers contain at least one dense layer with approximately 200-250 neurons, followed by a dropout layer with a dropout rate of 10% to 20%, and this pattern is repeated in successive layers to reduce overfitting and improve generalization ability, thus forming a network architecture capable of achieving high classification accuracy for multiple deformation types; finally, a 6-unit fully connected output layer uses the softmax activation function to classify the results into one of six categories: no deformation, subsidence, uplift, landslide, rockfall, and an optional secondary category; • Training and Model Optimization: The core process of this method includes training a convolutional neural network (CNN) on a representative dataset of the target research region and validating its performance. Training data is obtained by sampling multi-source input layers at target locations, with the initial classification results used as the target label for each sample. Standard machine learning procedures are followed, including data normalization, splitting the training / validation sets, and multiple rounds of iterative optimization using Adam or SGD optimizers. Robustness is ensured by randomly disabling Dropout layers of neurons, with an early dropout rate of approximately 0.2 and later dropout rates of 0.

1. An early stopping strategy is employed to terminate training when the validation loss stops decreasing, preserving the optimal model weights. Through this process, the CNN can overcome the limitations of the initial rule labels and re-analyze the velocity grid and auxiliary data to optimize classification performance. • Automated classification and output: After training, the CNN model will be deployed to classify the types of surface deformation in the entire target area. For each pixel or unit area, the model generates probability values ​​or scores for each category by integrating multi-source features, and finally determines the category with the highest probability as the predicted deformation type. The resulting theme deformation map can clearly mark landslides, subsidence, uplift and other areas with high spatial resolution. • Model validation and accuracy evaluation: This further includes validation by comparing the classification output of the trained CNN classifier with known ground truth or the results of independent methods, and comparing the accuracy or prediction performance with baseline machine learning methods that do not use deep neural networks. The evaluation parameters are quantified using the overall classification accuracy and the area under the ROC curve (AUC) metric.

5. A method and system for detecting and classifying surface deformation by integrating CNN and SBAS-InSAR technologies, characterized in that... It includes the following modules: a data acquisition module, which receives and stores SAR images and auxiliary geospatial data of the area; an InSAR processing module, which performs SBAS interferometric analysis to process SAR images and output deformation velocity maps; and a data integration module, which extracts features from optical principal component analysis (PCA) components and auxiliary data, including terrain attributes, and integrates all features into a unified geospatial grid. The classification module contains a processor that executes instructions to complete two tasks: (1) applying threshold rules to deformation rate and slope data to generate preliminary classification labels; (2) training and applying a convolutional neural network to perform final deformation prediction for each grid location. The output module generates a visual deformation map and report, marking the locations that the CNN determines to be landslides, subsidence, uplift, rockfalls, or stable ground. The system supports real-time updates to classification results, automatically updating classification information when new SAR or satellite data is acquired, thereby building an automated and continuously learning ground deformation monitoring tool. The convolutional neural network in the system's classification module is pre-trained on historical data from the region and is incrementally updated with new data, enabling real-time or near-real-time risk classification without having to retrain from scratch for each update. The system's output module further calculates and displays the confidence level for each classification region and allows users to interact to query the underlying multi-source data values ​​that led to a specific classification, thus facilitating interpretable results for end users.

Images