Deep learning-based unstable slope identification and detection method

Abstract

The invention provides an unstable slope feature recognition and detection method based on deep learning. The method sequentially comprises the steps of S1, collecting image data of a slope area from multi-source data, and preprocessing the image data to construct a standard data set; s2, constructing and training a SlopeNet model based on a convolutional neural network; s3, identifying the range and boundary of the unstable slope area by using the SlopeNet model trained in the step S2; s4, based on the range and the boundary of the unstable slope area obtained in the step S3, establishing an intelligent analysis model for the evolution trend of the unstable slope in combination with historical meteorological, hydrological and geological monitoring data; and S5, predicting the evolution trend, namely the instability probability, of the unstable slope in a period of time in the future by using the intelligent analysis model for the evolution trend of the unstable slope trained in the step S4. According to the method, through fusion of multi-source data and combination of the convolutional neural network and the long-short-term memory network, automatic identification and detection of the unstable slope are realized, and the method has relatively high precision and efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of slope engineering, and specifically is a method for identifying and detecting unstable slope features based on deep learning SlopeNet. Background Art

[0002] Slope stability is a critical issue in geotechnical engineering. The formation and evolution of unstable slopes not only impacts project safety but can also pose significant threats to the environment and human life and property. Traditional slope stability assessment methods rely on manual surveys, geological investigations, and physical model tests. While these methods can provide some guidance, they suffer from low efficiency and the inability to monitor in real time.

[0003] The rapid development of data acquisition technologies such as remote sensing, drones, and lidar has made it possible to analyze slope characteristics by combining large-scale datasets with advanced computational methods. In recent years, deep learning, as a powerful data analysis tool, has achieved remarkable results in image processing, speech recognition, and natural language processing. However, the automated identification and detection of unstable slopes, particularly in complex terrain, has yet to fully exploit the potential of deep learning.

[0004] Therefore, the slope instability feature recognition and detection method based on deep learning can automatically extract the unstable characteristics of the slope from multi-source data such as remote sensing images and lidar point cloud data, and realize fast and accurate recognition and detection through deep neural network models, and automatically generate a complete survey plan. It is a technical problem that needs to be solved urgently in the field of geotechnical engineering survey. Summary of the Invention

[0005] In response to the problems existing in the prior art, the present invention provides a method for identifying and detecting unstable slope features based on deep learning. The method aims to automatically extract slope instability features from multi-source data through deep learning technology, thereby achieving efficient, accurate and intelligent slope stability assessment.

[0006] In order to achieve the above technical objectives, the present invention provides a method for identifying and detecting unstable slope features based on deep learning, which specifically includes the following steps:

[0007] S1. Collect image data of slope areas from multiple sources and preprocess the collected image data to construct a standard dataset for deep learning model training, validation, and testing;

[0008] S2. Build and train a convolutional neural network-based SlopeNet model. The SlopeNet model integrates multi-scale feature extraction, deep residual connections, an adaptive attention mechanism, and a slope feature enhancement module. It also enhances the learning of unstable slope region features by designing an adaptive loss function. The SlopeNet model is trained end-to-end using the standard dataset preprocessed in S1 to automatically extract unstable slope features and perform image segmentation, thereby completing the pixel-level classification model.

[0009] S3. Use the SlopeNet model trained in step S2 to perform inference on new or unlabeled remote sensing image data to achieve pixel-level classification and image segmentation of slope areas, thereby accurately identifying the scope and boundaries of unstable slope areas.

[0010] S4. Based on the scope and boundaries of the unstable slope area obtained in step S3, combined with historical meteorological, hydrological, and geological monitoring data, a time series dataset is constructed. A long short-term memory (LSTM) time series deep learning model is used to train this time series dataset and establish an intelligent analysis model for the evolution trend of unstable slopes.

[0011] S5. Using the intelligent analysis model of unstable slope evolution trends trained in step S4, input the current and recent spatial characteristics of unstable slopes and historical monitoring data series to predict the evolution trend of unstable slopes in the future, that is, the probability of instability.

[0012] A further technical solution of the present invention: the method further includes the following steps: outputting the recognition result in step S3 and the prediction result in step S5 in a visual manner to provide intuitive evaluation and warning information.

[0013] A further technical solution of the present invention is as follows: the multi-source data in the S1 step includes drone image data, satellite remote sensing image data, geological exploration data, and meteorological and historical monitoring data; preprocessing includes data cleaning, format conversion, geometric correction, spatial registration, noise removal, image enhancement, data standardization, precise labeling of unstable slope areas, data enhancement, multi-source data fusion, and data set division; wherein the precise labeling of the unstable slope area is based on the slope, aspect, texture, and curvature characteristics of remote sensing images, lidar point clouds, and drone aerial image data, observation and analysis, and the use of GIS software to draw polygons for possible landslides or unstable slope areas to mark unstable slope areas.

[0014] A further technical solution of the present invention is as follows: the automated extraction process of unstable slope features in step S2 specifically includes:

[0015] S201. Multi-scale Convolutional Feature Extraction: Multiple convolutional layers are used to extract multi-scale local features of an image at different receptive fields.

[0016] S202. Multi-scale feature fusion: Downsample the local feature map extracted in step S201 using a pooling layer, using max pooling to reduce the image size and extract high-level abstract information from the image.

[0017] S203. Deep Residual Connections: Drawing on the idea of ​​ResNet, we introduce skip connections into the network; each residual block consists of two convolutional layers and one skip connection.

[0018] S204. Design an adaptive attention mechanism. Introduce a spatial attention module at different stages of the feature map. Dynamically adjust the importance weights of features at different spatial locations based on the saliency of the feature map, so that the model focuses more on regions and features related to unstable slope discrimination.

[0019] S205. Slope feature enhancement module: Integrates prior geological information into the deep learning feature learning process, and uses this module to enhance the extraction of key local details of unstable slopes.

[0020] A further technical solution of the present invention is as follows: in the process of automatically extracting unstable slope features in step S2, an adaptive loss function is used to focus on the unstable slope area, thereby improving the accuracy and robustness of feature extraction; the adaptive loss function includes three parts: cross entropy loss, regularization loss, and slope area loss. The adaptive loss function is as follows:

[0021] L total = αL ce +βL reg +γL instability ;

[0022] Where: L ce is the conventional cross entropy loss, used for classification;

[0023] L reg is the regularization loss, used to prevent the model from overfitting;

[0024] L instability It is a special loss function for unstable slope areas. This loss function depends on the complexity of the slope area and can be adaptively adjusted according to the significance of the slope area.

[0025] α, β, and γ in the loss function are hyperparameters used to adjust the weights of the three parts of the loss. Generally speaking, the value of γ should be large to increase attention to unstable slope areas.

[0026] A further technical solution of the present invention is as follows: in step S2, the SlopeNet model is trained end-to-end using the standard dataset preprocessed and divided in step S1, using a suitable Adam optimizer and a learning rate scheduling strategy, with the goal of minimizing the adaptive weighted loss function; during the training process, a validation set is used to monitor model performance and perform hyperparameter adjustments to prevent overfitting; and finally, the trained SlopeNet model weights are obtained.

[0027] A further technical solution of the present invention is as follows: the process of identifying the range and boundary of the unstable slope area based on the SlopeNet model in step S3 specifically includes:

[0028] S301. Data input: Input the slope area image data that is consistent with the training data format and has been pre-processed in step S1 but not labeled;

[0029] S302. Model inference: The input data from S301 is fed into the trained SlopeNet model for forward propagation. The model outputs a probability map of each pixel belonging to the unstable category.

[0030] S303. Post-processing: Post-process the output probability map by binarizing it using a threshold value to generate a binary mask, where a pixel value of 1 indicates an unstable slope area and a pixel value of 0 indicates a stable area. Set operations are then applied to obtain a smooth and reasonable segmentation result.

[0031] S304. Result output: Output a pixel-level binary segmentation map and vector polygon to represent the identification result of the range and boundary of the unstable slope area.

[0032] A further technical solution of the present invention is as follows: the process of establishing an intelligent analysis model of unstable slope evolution trend based on LSTM in step S4 specifically includes:

[0033] S401. Historical spatial feature extraction: For slope regions with time series data, repeat steps S1 and S3, using the trained SlopeNet model to identify and demarcate unstable slope regions at different historical time points. Quantitative spatial features are then extracted from these time-varying region segmentation results.

[0034] S402. Historical monitoring data synchronization: collecting historical monitoring data synchronized with the above spatial feature extraction time point;

[0035] S403. Time series construction: The spatial feature vector extracted from each historical time point is concatenated with the corresponding historical monitoring data vector to form a comprehensive feature vector. These comprehensive feature vectors are arranged in chronological order to construct a multivariate time series dataset.

[0036] S404. Design of an intelligent analysis model for the evolution trend of unstable slopes: The intelligent analysis model for the evolution trend of unstable slopes adopts an LSTM model architecture, including an input layer, an LSTM layer, a Dropout layer, a fully connected layer, and an output layer. The input layer receives multivariate time series data expanded by time steps. The LSTM unit contains a gating mechanism that can effectively learn and memorize long-term dependent information. The Dropout layer is used to randomly inactivate some neurons between LSTM layers to prevent overfitting. The fully connected layer is responsible for mapping the output of the LSTM layer to the final prediction target. The output layer uses a sigmoid activation function to output probabilities to predict whether the slope will be unstable in the future.

[0037] A further technical solution of the present invention is as follows: the unstable slope evolution trend intelligent analysis model outputs the prediction result in step S5 and the identification result in step S3;

[0038] The prediction results are based on multiple time-step data from the past, predicting the slope state at a future time point, as well as the slope state at future sequence time points. The prediction result output is a time series, and the model outputs probability values ​​between 0 and 1 at multiple time points, representing the slope evolution trend and the probability of instability over a period of time in the future, where 0 indicates no instability and 1 indicates instability.

[0039] The recognition result in step S3 includes outputting a pixel-level binary mask and an unstable region vector boundary, which indicates the spatial extent of the unstable slope region and can be superimposed on the original image or DEM for display;

[0040] The prediction results output by the unstable slope evolution trend intelligent analysis model and the identification results in step S3 are both displayed visually. Specifically, the vector boundaries of the identified unstable areas are superimposed on the 3D terrain model for visualization. The pixel-level prediction probabilities from the SlopeNet identification results are superimposed on the slope map using heat maps or graded coloring. Time series graphs or bar charts are used to display historical changes in the area of ​​unstable areas, key monitoring data curves, and the time-varying curve of the future instability probability of the unstable slope evolution trend intelligent analysis model, providing dynamic evolution trend information. The predicted probabilities are then combined with preset warning thresholds, and different warning levels are indicated by different colors or markers.

[0041] A preferred technical solution of the present invention: the spatial features extracted in S401 include: the area of ​​each unstable region, the shape characteristics of the unstable region, i.e., the aspect ratio, the terrain feature statistics within the unstable region, and the change of the unstable region over time; the terrain feature statistics within the unstable region include the average slope, the maximum slope, the surface roughness, and the undulation; the change of the unstable region over time includes the area change and the boundary change; the historical monitoring data of S402 includes the accumulated precipitation, the groundwater level, and the displacement of the measuring point.

[0042] Compared with the prior art, the present invention has the following significant beneficial effects:

[0043] (1) The present invention has high precision and robustness: through the deep fusion of multi-source data and the specially designed SlopeNet model (including multi-scale feature extraction, attention mechanism, targeted feature enhancement module and adaptive weighted loss), it can more comprehensively and accurately capture the subtle features and complex patterns of unstable slopes, significantly improving the pixel-level recognition accuracy of unstable slope areas and the robustness in complex geological environments; the adaptive loss function particularly enhances the model's learning ability for high-risk and complex areas.

[0044] (2) The present invention is automated and efficient: the entire identification and detection process is highly automated, reducing dependence on manual experience and heavy labor, greatly improving the efficiency of slope exploration and monitoring, and is particularly suitable for rapid screening of large areas.

[0045] (3) The present invention has pixel-level refined recognition: SlopeNet realizes pixel-level classification based on the idea of ​​image segmentation, which can accurately depict the boundaries and scope of unstable slope areas, providing accurate spatial information for subsequent detailed surveys, monitoring and management.

[0046] (4) The present invention can realize intelligent evolution trend prediction: it creatively combines the quantitative features extracted from the historical spatial recognition results with traditional monitoring data as the input of the time series model LSTM, realizes intelligent analysis and probabilistic prediction of the future evolution trend of unstable slopes, provides the possibility of early warning, and moves from "identifying the current situation" to "predicting the future."

[0047] (5) The present invention has strong adaptability: the deep learning model has strong generalization ability. After training with diversified data, it has good adaptability to slopes of different types and environments. DETAILED DESCRIPTION

[0048] The present invention is further described below with reference to the embodiments. The technical solutions presented in the embodiments are specific solutions of the embodiments of the present invention and are not intended to limit the scope of the invention claimed for protection. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] The embodiment provides a method for identifying and detecting unstable slope features based on deep learning, which specifically includes the following steps:

[0050] S1. Collect image data of slope areas from multiple sources (drone image data, satellite remote sensing image data, geological exploration data) and perform comprehensive preprocessing on the collected data, including data cleaning, format conversion, geometric correction, spatial registration, noise removal, image enhancement, data standardization, precise manual annotation of unstable slope areas (as training labels for deep learning models), data enhancement, and multi-source data fusion. Finally, construct a standard dataset for deep learning model training, validation, and testing.

[0051] Satellite remote sensing imagery data acquisition: Use high-resolution satellite imagery (e.g., Google Earth Engine, Sentinel-2, etc.) to obtain two-dimensional ground images of the slope area. The acquired images must cover the entire extent of the slope, including possible landslide areas, fracture zones, and adjacent terrain features.

[0052] Drone image data collection: Use drones equipped with high-resolution cameras to capture aerial images of slopes. Drones can provide low-altitude, high-resolution images, making them particularly useful in hard-to-reach areas. Drones can also acquire multi-dimensional data such as 3D point cloud data and vegetation coverage from slopes by using lidar or multispectral cameras.

[0053] Geological exploration data collection: Combined with traditional geological exploration data, such as soil type, groundwater level, rock and soil samples, through ground drilling and geological profile analysis, to assist in the assessment of slope stability;

[0054] Meteorological and historical monitoring data: Collect meteorological data (such as precipitation, temperature, and wind speed) and historical landslide monitoring data. This data is valuable for understanding slope dynamics and the triggering conditions for landslides.

[0055] After data collection in step S1, the raw data needs to be preprocessed so that the deep learning model can efficiently extract useful features from it. The preprocessing step specifically includes the following processing steps:

[0056] Data cleaning: Remove irrelevant data and outliers to ensure the quality of input data; for example, remove clouds, haze, and other parts that have a greater impact on remote sensing images.

[0057] Format conversion: Converting data from different sources into a unified format for easy processing by deep learning models; for example, converting remote sensing images from different formats (such as .tif, .jpg) into a standard input format such as PNG or JPEG.

[0058] Image geometric correction: For remote sensing images, geometric correction is required to ensure that the slope areas in the image correspond to the ground truth coordinates; this step includes image registration, geographic coordinate transformation, and image distortion removal.

[0059] Data registration: Multi-source data (such as remote sensing images and drone images) need to be spatially registered, that is, different data sources are converted to the same coordinate system for subsequent fusion and analysis.

[0060] Noise removal: Remote sensing images are often affected by factors such as weather, sensor accuracy, and lighting changes, resulting in noise in the data. Filtering algorithms (such as median filtering and Gaussian filtering) can be used to denoise the data.

[0061] Image enhancement: To improve image quality, image enhancement methods such as histogram equalization and contrast stretching are used. By enhancing the contrast of slope features (such as cracks and signs of landslides), deep learning models can better identify these key features.

[0062] Data standardization: Data from different sources may have different dimensions and scales and must be standardized. For example, all input features can be converted to a uniform scale range (such as between 0 and 1) through the mean-variance normalization method to ensure that the training process of the deep learning model is not affected by data scale differences.

[0063] Unstable slope marking: Manual marking is used. Based on the slope, aspect, texture, curvature and other characteristics of remote sensing images, lidar point clouds and drone aerial images, polygons are drawn for possible landslide or unstable slope areas using GIS software such as ArcGIS, QGIS, and ENVI. Unstable slope areas are manually marked.

[0064] Data augmentation: To enhance the robustness of the model, data augmentation techniques are used to perform transformations such as rotation, translation, cropping, and scaling on the original images. These transformations can generate diverse training samples and help the model generalize better.

[0065] Multi-source data fusion: Data from different sources, such as remote sensing images and drone aerial photography, are integrated to form a comprehensive dataset containing spatial information (such as slope surface topography), temporal information (such as meteorological data and precipitation), and three-dimensional information (slope, aspect, etc.), while ensuring that the coordinates of all input data are aligned.

[0066] Dataset Partitioning: Divide the preprocessed dataset into a training set, a validation set, and a test set, typically with a 70%:15%:15% split ratio. Use the training set for model training, the validation set for hyperparameter adjustment, and the test set for model performance evaluation.

[0067] S2. Construct and train a convolutional neural network (CNN)-based SlopeNet model. This model integrates multi-scale feature extraction, deep residual connections, an adaptive attention mechanism, and a feature enhancement module. By designing an adaptive loss function, it enhances the learning of unstable slope region features. The SlopeNet model is trained end-to-end using the standard dataset preprocessed in step S1, resulting in a model that can automatically extract unstable slope features, perform image segmentation, and then complete pixel-level classification. This paper proposes a "SlopeNet" deep learning algorithm for automated extraction of unstable slope features. This algorithm combines convolutional neural networks (CNNs) with multi-scale fusion techniques and uses an adaptive loss function to focus on unstable slope regions, thereby improving the accuracy and robustness of feature extraction. The automated extraction of unstable slope features specifically includes:

[0068] S201. Multi-scale Convolutional Feature Extraction Block: This block uses multiple convolutional layers to extract multi-scale local features of the image across different receptive fields. By performing convolution operations at different scales using convolution kernels of varying sizes (e.g., 3x3, 5x5, 7x7, etc.), it can capture subtle changes in slope topography (e.g., soil landslides, rock debris accumulation, etc.).

[0069] Input: Feature map of the previous layer (size is H′×W′×C′).

[0070] Output: Feature map output by the convolutional layer (size is H′×W′×C out , where C out is the number of output channels of the convolutional layer)

[0071] S202. Multi-scale feature fusion module: Combines the pooling layer to downsample the local feature map extracted in step S201, uses max pooling to reduce the image size, and extracts high-level abstract information of the image;

[0072] Input: Feature map output by the convolutional layer (size is H′×W′×C out );

[0073] Output: Pooled feature map (size is × ×C out ). And through the feature fusion layer, 1×1 convolution is used to fuse information from different scales to achieve multi-scale information fusion to retain a wider range of context information.

[0074] Input: Feature maps of multiple scales (size H′×W′×C1, H′×W′×C 2， ...); Output: fused feature map (size H′×W′×C fused ).

[0075] S203. Deep Residual Connections: Drawing on the ideas of ResNet, we introduce skip connections into the network. Residual connections allow information to be transferred directly across layers, facilitating the training of deeper networks, alleviating the vanishing gradient problem, accelerating convergence, and improving model performance. Each residual block consists of two convolutional layers and a skip connection.

[0076] Input: Input feature map of each residual block (size is H′×W′×Cresidual);

[0077] Output: Feature map after residual learning (size is H′×W′×Cresidual).

[0078] S204. Adaptive Attention Mechanism: A spatial attention module is introduced at different stages of the feature map. This attention mechanism dynamically adjusts the importance weights of features at different spatial locations based on the saliency of the feature map (regions associated with unstable slopes are typically more salient, such as obvious morphological changes and texture anomalies), allowing the model to focus more on regions and features relevant to unstable slope discrimination.

[0079] S205. Slope Feature Enhancement Modules integrate prior geological information into the deep learning feature learning process. This module is used to enhance the extraction of key local details of unstable slopes, such as potential cracks, step-downs, vegetation anomalies, and water pools.

[0080] In the process of automatically extracting unstable slope features in step S2, an adaptive loss function is used to focus on the unstable slope area, thereby improving the accuracy and robustness of feature extraction; the adaptive loss function includes three parts: cross entropy loss, regularization loss, and slope area loss:

[0081] Loss function: L total = αL ce +βLreg +γL instability

[0082] Where: L ce : Conventional cross entropy loss, used for classification.

[0083] L reg : Regularization loss, used to prevent the model from overfitting.

[0084] L instability : A special loss function for unstable slope regions. This loss function depends on the complexity of the slope region and can be adaptively adjusted based on the saliency of the slope region. For example, the saliency of the region can be calculated based on the slope change, texture complexity, etc.

[0085] The α, β, and γ in the loss function are hyperparameters used to adjust the weights of the three losses. Generally speaking, the value of γ should be large to increase the focus on unstable slope areas.

[0086] Input: The input data consists of three parts: (1) Prediction output: The output of the network is a predicted probability distribution map (size H′×W′×2), which indicates the probability of each pixel belonging to an unstable slope or a stable slope; (2) Ground Truth: The ground truth is a labeled map of size H′×W′, where each pixel value represents the true category of the pixel; (3) Slope region saliency map: The saliency map dynamically adjusts the loss function by calculating the complexity of the slope region in the image (such as slope, slope direction change, etc.); the size of the saliency map is the same as the input image (H′×W′).

[0087] Output: The final total loss value, which reflects the training error of the model and serves as the optimization target.

[0088] S3. Use the SlopeNet model trained in step S2 to perform inference (forward propagation) on new or unlabeled remote sensing image data to achieve pixel-level classification and image segmentation of slope areas, thereby accurately identifying the scope and boundaries of unstable slope areas.

[0089] Model training specifically includes:

[0090] (1) Data input:

[0091] Training data: a set of annotated images (size H×W×C) and their corresponding labels (size H×W);

[0092] Training batches: Divide the images into small batches, each batch contains multiple samples.

[0093] (2) Training process: Forward Propagation:

[0094] Input: image data for each batch (size B×H×W×C, where B is the batch size);

[0095] Output: prediction result (size is B×H′×W′×N).

[0096] (3) Loss calculation:

[0097] Input: predicted output and true label;

[0098] Output: calculated loss value (size B);

[0099] (4) Backward Propagation: By calculating the gradient of the loss function, backpropagating the error and updating the network parameters.

[0100] (5) Optimization algorithm: Use the Adam optimizer for training, calculate the gradient based on the loss value and update the network weights; input: calculated gradient and current network parameters; output: optimized network parameters.

[0101] (6) Model evaluation: Use evaluation indicators such as accuracy, precision, recall, and F1 score to evaluate the performance of the model on the test set.

[0102] Optimized network model: After multiple rounds of training, the final optimized model is obtained, which can accurately extract slope features from remote sensing images and identify unstable slope areas, including:

[0103] Input is unlabeled, pre-processed slope image data in the same format as the training data, which has been processed in step S1. This input data is fed into the trained SlopeNet model for forward propagation. The model outputs a probability map indicating that each pixel belongs to the unstable category. This output probability map is post-processed by binarizing it using a threshold to generate a binary mask, where a pixel value of 1 indicates an unstable slope region and a pixel value of 0 indicates a stable region. Set operations are then applied to obtain a smooth, reasonable segmentation result. Finally, a pixel-by-pixel binary segmentation map and vector polygons are output, representing the range and boundary of the unstable slope region.

[0104] S4. Based on the scope and boundaries of the unstable slope area obtained in step S3, combined with historical meteorological, hydrological, and geological monitoring data, a time series dataset is constructed. A long short-term memory (LSTM) time series deep learning model is used to train this time series dataset and establish an intelligent analysis model for the evolution trend of unstable slopes. This process is divided into data preparation, model design, and training.

[0105] (1) Data Preparation: In previous work, we used the Slopenet network to segment slope images and obtained labels for stable and unstable slope regions. We need to extract the spatial features of unstable slope regions from these labels. The segmentation result obtained by Slopenet is usually a binary image, where each pixel represents whether a certain area of ​​the slope is stable, with a label of 1 for stable areas and 0 for unstable areas. The spatial features of these unstable slope regions can be further analyzed by performing image analysis on the unstable areas, such as extracting morphological features, slope changes, crack distribution, soil composition changes, etc., to construct a spatial feature vector, which serves as the input for the subsequent LSTM model.

[0106] Collect historical monitoring data related to unstable slopes. This data will be used to train the LSTM network, mainly including:

[0107] Slope displacement data: displacement data that can be measured by GPS, inclinometers and other equipment.

[0108] Precipitation: historical precipitation data, reflecting the impact of precipitation on slope stability.

[0109] Seismicity data: earthquake events and their intensity, reflecting the impact of external geological factors on slope stability.

[0110] Temperature data: If applicable, changes in air temperature may also affect slope stability.

[0111] Other geological data: such as groundwater level, ground vibration, etc.

[0112] These monitoring data are arranged in chronological order into time series as the input of the LSTM network; the spatial characteristics of the unstable slope area and the historical monitoring data are combined to ensure that the timestamps of the two types of data are aligned. In particular, in the input of time series data, the monitoring data at each time point is ensured to be consistent with the spatial characteristics of the corresponding time period.

[0113] (2) Design of an intelligent analysis model for the evolution trend of unstable slopes: The prediction model adopts the LSTM model architecture, and the input data should include time series data and spatial features.

[0114] For each time step t, the input data can be a vector containing the following: a. Historical monitoring data: such as displacement, precipitation, and seismic activity. b. Spatial features: The spatial feature vector of the unstable region segmented by Slopenet. After concatenating the time series data and spatial features, a single input vector containing multiple features is formed.

[0115] The LSTM model architecture includes input layer, LSTM layer, Dropout layer, fully connected layer, and output layer;

[0116] The size of the input layer depends on the dimensions of the historical data. For example, if you are using 10 time steps of data for forecasting, and each time step contains several features (such as displacement, precipitation, etc.), then the dimension of the input layer is (10, n), where n is the number of features (for example, 4 features).

[0117] LSTM layer 1: Contains 64 units and uses the tanh activation function. The goal of the LSTM layer is to capture long-term dependencies in time series.

[0118] LSTM layer 2: A further LSTM layer is stacked to enhance the model's representation capabilities. The number of units can be set to 32.

[0119] Dropout layer: A dropout layer (e.g., 0.2 or 0.3) is added after each LSTM layer to prevent overfitting.

[0120] Dense Layer: The output of the LSTM layer is passed to a fully connected layer to further extract high-level features. The size of the output layer matches the prediction target. For regression problems, the output layer consists of one node; for classification problems, the output layer consists of multiple nodes (e.g., unstable / stable).

[0121] The output layer uses a sigmoid or softmax activation function (depending on whether the problem is regression or classification); if it is a regression problem, the output is the predicted value of the slope displacement; if it is a classification problem, the output is the probability of whether an instability event will occur in the future.

[0122] Model hyperparameters

[0123] Number of LSTM layers: You can choose a 1-2 layer LSTM structure, depending on the complexity of the data.

[0124] Number of LSTM units: 64, 128, etc.; can be adjusted according to the task.

[0125] Learning rate: The learning rate can be set to 0.001 and the Adam optimizer is used.

[0126] Batch Size: Generally 32 or 64, the specific value is adjusted according to the amount of training data.

[0127] Number of training rounds (Epochs): For example, train for 50-100 epochs, depending on the performance of the model on the validation set.

[0128] (3) Model training process: The data is divided into a training set and a validation set, with a general ratio of 80% for training and 20% for validation. The validation set is used to monitor whether the model is overfitting and adjust hyperparameters. Historical monitoring data and spatial features of unstable slopes are used as input for model training. Model parameters are optimized by minimizing the prediction error (such as mean squared error (MSE) or cross entropy loss). Grid search or random search is used to adjust hyperparameters such as learning rate, number of LSTM units, and dropout ratio to find the optimal model.

[0129] S5. Using the intelligent analysis model of unstable slope evolution trends trained in step S4, input the current and recent spatial characteristics of unstable slopes and historical monitoring data series to predict the evolution trend of unstable slopes in the future, that is, the probability of instability.

[0130] Model prediction and application: Using the trained unstable slope evolution trend intelligent analysis model, inputting new historical monitoring data and spatial features, the model will output the slope stability evolution trend for a period of time in the future. For example:

[0131] Predict future slope displacement values;

[0132] Predict whether a slope instability event is likely to occur in the future.

[0133] Based on the prediction results output by the model, a threshold can be set to determine whether the slope is about to become unstable. If the prediction value exceeds the threshold (such as excessive slope displacement changes, excessive precipitation, etc.), the corresponding early warning system can be activated.

[0134] Use mean squared error (MSE) or accuracy (if it is a classification problem) to evaluate the performance of the prediction. The model can be further optimized through cross-validation.

[0135] The final output of the model includes the recognition results in step S3 and the prediction results in step S5, both of which are output in a visual way to provide intuitive evaluation and early warning information.

[0136] The prediction results are based on multiple time-step data from the past, predicting the slope state at a future time point, as well as the slope state at future sequence time points. The prediction result output is a time series, and the model outputs probability values ​​between 0 and 1 at multiple time points, representing the slope evolution trend and the probability of instability over a period of time in the future, where 0 indicates no instability and 1 indicates instability.

[0137] The recognition result in step S3 includes outputting a pixel-level binary mask and an unstable region vector boundary, which indicates the spatial extent of the unstable slope region and can be superimposed on the original image or DEM for display;

[0138] The prediction results output by the unstable slope evolution trend intelligent analysis model and the identification results in step S3 are both displayed visually. Specifically, the vector boundaries of the identified unstable areas are superimposed on the 3D terrain model for visualization. The pixel-level prediction probabilities from the SlopeNet identification results are superimposed on the slope map using heat maps or graded coloring. Time series graphs or bar charts are used to display historical changes in the area of ​​unstable areas, key monitoring data curves, and the time-varying curve of the future instability probability of the unstable slope evolution trend intelligent analysis model, providing dynamic evolution trend information. The predicted probabilities are then combined with preset warning thresholds, and different warning levels are indicated by different colors or markers.

[0139] For example, a bar chart is used to display the classification results for each time step. The following table shows the prediction results for each week in the future:

[0140]

[0141] The predicted probability values ​​can also be displayed as a bar chart, where each bar at a time point represents the probability of slope instability at that time point. The predicted probability value can be directly compared to a threshold value (e.g., 0.7). If the probability is greater than the threshold, it can be displayed as "instability"; if it is less than the threshold, it can be displayed as "stable."

[0142] This paper proposes a deep learning-based method for identifying and detecting unstable slope features. By integrating multi-source data with convolutional neural networks and long-short-term memory networks, it achieves automated identification and detection of unstable slopes. This method boasts high accuracy and efficiency, providing effective support for geotechnical engineering investigations, landslide disaster prediction, and environmental protection, and possesses broad application prospects and practical value.

Claims

1. A method for identifying and detecting unstable slope features based on deep learning, characterized in that: The specific steps include: S1. Collect image data of slope areas from multiple sources and preprocess the collected image data to construct a standard dataset for deep learning model training, validation, and testing; S2. Build and train a convolutional neural network-based SlopeNet model. The SlopeNet model integrates multi-scale feature extraction, deep residual connections, an adaptive attention mechanism, and a slope feature enhancement module. It also enhances the learning of unstable slope region features by designing an adaptive loss function. The SlopeNet model is trained end-to-end using the standard dataset preprocessed in step S1 to automatically extract unstable slope features and perform image segmentation, thereby completing the pixel-level classification model. S3. Use the SlopeNet model trained in step S2 to perform inference on new or unlabeled remote sensing image data to achieve pixel-level classification and image segmentation of slope areas, thereby accurately identifying the scope and boundaries of unstable slope areas. S4. Based on the scope and boundaries of the unstable slope area obtained in step S3, combined with historical meteorological, hydrological, and geological monitoring data, a time series dataset is constructed. A long short-term memory (LSTM) time series deep learning model is used to train this time series dataset and establish an intelligent analysis model for the evolution trend of unstable slopes. S5. Using the intelligent analysis model of unstable slope evolution trends trained in step S4, input the current and recent spatial characteristics of unstable slopes and historical monitoring data series to predict the evolution trend of unstable slopes in the future, that is, the probability of instability.

2. The method for identifying and detecting unstable slope features based on deep learning according to claim 1, characterized in that: The method further includes the following steps: outputting the recognition result in step S3 and the prediction result in step S5 in a visual manner to provide intuitive evaluation and warning information.

3. The method for identifying and detecting unstable slope features based on deep learning according to claim 1 or 2, characterized in that: The multi-source data in step S1 include UAV image data, satellite remote sensing image data, geological exploration data and meteorological and historical monitoring data; preprocessing includes data cleaning, format conversion, geometric correction, spatial registration, noise removal, image enhancement, data standardization, precise labeling of unstable slope areas, data enhancement, multi-source data fusion and data set division; wherein the precise labeling of unstable slope areas is based on the slope, aspect, texture and curvature characteristics of remote sensing images, lidar point clouds and UAV aerial image data, observation and analysis, and the use of GIS software to draw polygons for possible landslides or unstable slope areas to mark unstable slope areas.

4. The method for identifying and detecting unstable slope features based on deep learning according to claim 1 or 2, characterized in that: The automated extraction process of unstable slope features in step S2 specifically includes: S201. Multi-scale Convolutional Feature Extraction: Multiple convolutional layers are used to extract multi-scale local features of an image at different receptive fields. S202. Multi-scale feature fusion: Downsample the local feature map extracted in step S201 using a pooling layer, using max pooling to reduce the image size and extract high-level abstract information from the image. S203. Deep Residual Connections: Drawing on the idea of ​​ResNet, we introduce skip connections into the network; each residual block consists of two convolutional layers and one skip connection. S204. Design an adaptive attention mechanism. Introduce a spatial attention module at different stages of the feature map. Dynamically adjust the importance weights of features at different spatial locations based on the saliency of the feature map, so that the model focuses more on regions and features related to unstable slope discrimination. S205. Slope feature enhancement module: Integrates prior geological information into the deep learning feature learning process, and uses this module to enhance the extraction of key local details of unstable slopes.

5. The method for identifying and detecting unstable slope features based on deep learning according to claim 1 or 2, characterized in that: In the step S2, during the automated extraction of unstable slope features, an adaptive loss function is used to focus on the unstable slope region, thereby improving the accuracy and robustness of feature extraction. The adaptive loss function includes three parts: cross entropy loss, regularization loss, and slope region loss. The adaptive loss function is as follows: L total = αL ce +βL reg +γL instability ; Where: L ce is the conventional cross entropy loss, used for classification; L reg is the regularization loss, used to prevent the model from overfitting; L instability It is a special loss function for unstable slope areas. This loss function depends on the complexity of the slope area and can be adaptively adjusted according to the significance of the slope area. α, β, and γ in the loss function are hyperparameters used to adjust the weights of the three parts of the loss. Generally speaking, the value of γ should be large to increase attention to unstable slope areas.

6. The method for identifying and detecting unstable slope features based on deep learning according to claim 1 or 2, characterized in that: In step S2, the SlopeNet model is trained end-to-end using the standard dataset preprocessed and divided in step S1. The appropriate Adam optimizer and learning rate scheduling strategy are used to minimize the adaptive weighted loss function. During the training process, the validation set is used to monitor the model performance and perform hyperparameter adjustments to prevent overfitting. Finally, the trained SlopeNet model weights are obtained.

7. The method for identifying and detecting unstable slope features based on deep learning according to claim 1 or 2, characterized in that: The process of identifying the range and boundary of the unstable slope area based on the SlopeNet model in step S3 specifically includes: S301. Data input: Input the slope area image data that is consistent with the training data format and has been pre-processed in step S1 but not labeled; S302. Model inference: The input data from S301 is fed into the trained SlopeNet model for forward propagation. The model outputs a probability map of each pixel belonging to the unstable category. S303. Post-processing: Post-process the output probability map by binarizing it using a threshold value to generate a binary mask, where a pixel value of 1 indicates an unstable slope area and a pixel value of 0 indicates a stable area. Set operations are then applied to obtain a smooth and reasonable segmentation result. S304. Result output: Output a pixel-level binary segmentation map and vector polygon to represent the identification result of the range and boundary of the unstable slope area.

8. The method for identifying and detecting unstable slope features based on deep learning according to claim 1 or 2, characterized in that: The process of establishing the unstable slope evolution trend intelligent analysis model based on LSTM in step S4 specifically includes: S401. Historical spatial feature extraction: For slope regions with time series data, repeat steps S1 and S3, using the trained SlopeNet model to identify and demarcate unstable slope regions at different historical time points. Quantitative spatial features are then extracted from these time-varying region segmentation results. S402. Historical monitoring data synchronization: collecting historical monitoring data synchronized with the above spatial feature extraction time point; S403. Time series construction: The spatial feature vector extracted from each historical time point is concatenated with the corresponding historical monitoring data vector to form a comprehensive feature vector. These comprehensive feature vectors are arranged in chronological order to construct a multivariate time series dataset. S404. Design of an intelligent analysis model for the evolution trend of unstable slopes: The intelligent analysis model for the evolution trend of unstable slopes adopts an LSTM model architecture, including an input layer, an LSTM layer, a Dropout layer, a fully connected layer, and an output layer. The input layer receives multivariate time series data expanded by time steps. The LSTM unit contains a gating mechanism that can effectively learn and memorize long-term dependent information. The Dropout layer is used to randomly inactivate some neurons between LSTM layers to prevent overfitting. The fully connected layer is responsible for mapping the output of the LSTM layer to the final prediction target. The output layer uses a sigmoid activation function to output probabilities to predict whether the slope will be unstable in the future.

9. The method for identifying and detecting unstable slope features based on deep learning according to claim 2, wherein: The unstable slope evolution trend intelligent analysis model outputs the prediction result in step S5 and the identification result in step S3; The prediction result is based on multiple time step data in the past, and predicts the slope state at a future time point, as well as the slope state at future sequence time points; The prediction result output is a time series. The model outputs probability values ​​between 0 and 1 at multiple time points, indicating the slope evolution trend and the probability of instability in the future period of time, where 0 means no instability and 1 means instability. The recognition result in step S3 includes outputting a pixel-level binary mask and an unstable region vector boundary, which indicates the spatial extent of the unstable slope region and can be superimposed on the original image or DEM for display; The prediction results output by the unstable slope evolution trend intelligent analysis model and the identification results in step S3 are both displayed visually. Specifically, the vector boundaries of the identified unstable areas are superimposed on the 3D terrain model for visualization. The pixel-level prediction probabilities from the SlopeNet identification results are superimposed on the slope map using heat maps or graded coloring. Time series graphs or bar charts are used to display historical changes in the area of ​​unstable areas, key monitoring data curves, and the time-varying curve of the future instability probability of the unstable slope evolution trend intelligent analysis model, providing dynamic evolution trend information. The predicted probabilities are then combined with preset warning thresholds, and different warning levels are indicated by different colors or markers.

10. The method for identifying and detecting unstable slope features based on deep learning according to claim 7, characterized in that: The spatial features extracted in S401 include: the area of ​​each unstable region, the shape characteristics of the unstable region, i.e., the aspect ratio, the terrain feature statistics within the unstable region, and the change of the unstable region over time; the terrain feature statistics within the unstable region include the average slope, maximum slope, surface roughness, and undulation; the change of the unstable region over time includes area change and boundary change; the historical monitoring data in S402 includes precipitation accumulation, groundwater level, and measuring point displacement.

Images