A cnn mountainous city building-terrain-geology adaptation quantitative and landslide susceptibility prediction method fusing multi-source data

Abstract

This invention belongs to the field of risk prediction technology for mountainous cities, and discloses a CNN-based method for quantifying and predicting landslide susceptibility in mountainous cities by integrating multi-source data. The specific steps are as follows: Step 1: Multi-source data collection and processing, collecting remote sensing image data, geological data, topographic data, and building data. This invention achieves deep fusion of multi-source data by integrating remote sensing images, geological maps, topographic data, building distribution, and other multi-dimensional data; it automatically extracts deep features from the data through the convolutional and pooling layers of CNN, eliminating the need for manual feature design and improving feature extraction capabilities; through the construction of a CNN model that integrates heterogeneous data fusion and multi-source data fusion, it can directly capture spatial neighborhood relationships, improving spatial correlation; compared with traditional techniques, this method significantly improves the accuracy and robustness of landslide susceptibility prediction in complex mountainous cities through multi-source data complementarity and deep feature extraction.

Description

Technical Field

[0001] This invention belongs to the field of risk prediction technology for mountainous cities, specifically a CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by integrating multi-source data. Background Technology

[0002] CNN mountain city risk prediction refers to the use of convolutional neural network (CNN) technology, combined with the unique topography, geology and building data of mountain cities, to quantitatively assess the suitability of buildings in mountain cities with topography and geological conditions, and predict the susceptibility of geological disasters such as landslides, thereby providing a scientific basis for safety planning and disaster prevention in mountain cities.

[0003] Existing prediction methods, in practical applications, mainly rely on a single data source for landslide susceptibility assessment, which has the following drawbacks: Limited data dimensionality; traditional methods utilize only a single type of data, making it difficult to comprehensively reflect the complex interactions between buildings, topography, and geology, leading to biased assessment results; insufficient feature extraction capabilities; classical statistical models (such as frequency ratio method and logistic regression) or shallow machine learning models (such as SVM and random forest) cannot automatically learn deep nonlinear features in the data, especially when dealing with high-dimensional multi-source data; lack of spatial correlation; insufficient consideration of the spatial neighborhood relationship between landslide occurrence and the surrounding environment, resulting in inaccurate local risk assessments; and weak model generalization ability; in complex mountainous cities with significant differences in geological conditions, existing models struggle to adapt to different regional data distributions, leading to overfitting or underfitting. Summary of the Invention

[0004] The purpose of this invention is to provide a CNN-based method for adapting mountainous urban buildings, terrain, and geology to multiple data sources and predicting landslide susceptibility, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides the following technical solution: a CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data, the specific steps of which are as follows:

[0006] Step 1: Multi-source data collection and processing

[0007] Collect remote sensing image data, geological data, topographic data, and building data, and then preprocess the remote sensing image data, geological data, topographic data, and building data to unify the data format and spatial resolution:

[0008] Register remote sensing images and geological maps, keeping the error within one pixel; unify the spatial resolution of all data and use bilinear interpolation to maintain data continuity; for uninvestigated areas in the geological map, supplement lithological properties using kriging interpolation or neighborhood mean method.

[0009] Step 2: Heterogeneous Data Fusion

[0010] Different data sources are normalized, and the normalized multi-source data are spatially superimposed according to their spatial location to construct a unified multi-source data cube;

[0011] Step 3: Construct a sliding window sample set

[0012] A training sample set is constructed by extracting multi-source data features from local areas using a sliding window.

[0013] Step 4: Design the CNN model architecture

[0014] Construct a CNN model suitable for multi-source data fusion, including an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer;

[0015] Step 5: Model Training

[0016] The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Then, the CNN model is trained by the backpropagation algorithm, and the hyperparameters are optimized to improve performance.

[0017] Step Six: Extraction of Landslide Susceptibility Evaluation Indicators

[0018] Key factors influencing landslide susceptibility are extracted from the constructed CNN model, and a landslide susceptibility evaluation index system is built.

[0019] Step 7: Quantization of Building-Topography-Geology Adaptation

[0020] Based on landslide susceptibility evaluation indicators, a building-topography-geology adaptation evaluation model is constructed.

[0021] Step 8: Model Validation and Dynamic Update

[0022] The model is validated using historical landslide data that was not used in the training process. Combined with UAV aerial surveys and geological drilling, the high-risk areas predicted by the model are verified. The latest remote sensing images and geological survey data are incorporated annually to retrain the model. The model parameters are adjusted based on the field validation results to form a closed-loop optimization of data-model-validation.

[0023] As a preferred technical solution of the present invention, the normalization process in step two is as follows: for continuous data, Min-Max normalization or Z-score standardization is used; for discrete data, one-hot encoding is used to convert N types of lithology into N-dimensional binary vectors.

[0024] As a preferred embodiment of the present invention, the specific method for constructing the training sample set in step three is as follows:

[0025] Window design:

[0026] Set the window size according to the average size of the landslide to ensure that the window contains a complete interactive unit of topography, geology and architecture.

[0027] Sample labels:

[0028] Positive samples: Data within the window centered on the historical landslide boundary is extracted and marked as 1 (prone to occurrence).

[0029] Negative samples: Random sampling is conducted in non-landslide areas to ensure a 1:1 ratio of positive to negative samples and avoid class imbalance.

[0030] Data augmentation:

[0031] 90% of the samples were tested 0 Rotation and horizontal flipping operations expand the dataset to three times its original size, improving the model's generalization ability.

[0032] As a preferred technical solution of the present invention, the design method of the CNN model architecture in step four is as follows:

[0033] Input layer: Receives a 50×50×C multi-source data cube;

[0034] Convolution module:

[0035] Convolutional layer 1: 32 3×3 convolutional kernels, stride 1, ReLU activation, extracting low-level features (edges, textures).

[0036] Convolutional layer 2: 64 3×3 convolutional kernels, stride 1, capturing intermediate features (lithological boundaries, topographic relief patterns).

[0037] Pooling layer: 2×2 max pooling, reducing the feature dimension to 25×25×64;

[0038] Fully connected layer:

[0039] Flatten layer: Flattens the pooling output into a 1D vector (25×25×64=40,000 dimensions).

[0040] Dense layer 1: 256 neurons, Dropout=0.5 to prevent overfitting;

[0041] Dense layer 2 (output layer): 1 neuron, sigmoid activation, outputs the slippage probability (0-1).

[0042] As a preferred embodiment of the present invention, the calculation formula for the cross-entropy loss function in step five is as follows:

[0043]

[0044] In the formula, This represents the binary cross-entropy loss value; This indicates the number of samples, i.e., the total number of samples in the batch of data; Indicates the first The true label of each sample, with a value of 0 or 1; The model predicts the first The probability value of a sample being a positive class is usually mapped to the [0,1] interval by the Sigmoid activation function;

[0045] When training the CNN model using the backpropagation algorithm described in step five:

[0046] Training configuration:

[0047] Optimizer: Adam (learning rate = 0.001, , );

[0048] Batch size: 32, number of iterations = 100;

[0049] Hyperparameter tuning:

[0050] Using Keras Tuner for mesh search, the optimization parameters include:

[0051] Number of convolutional kernels (32 / 64 / 128); learning rate (0.0001 / 0.001 / 0.01); dropout rate (0.3 / 0.5 / 0.7).

[0052] Early shutdown mechanism:

[0053] Monitor the loss on the validation set; if it does not decrease within 10 iterations, terminate training to prevent overfitting.

[0054] As a preferred technical solution of the present invention, the specific method for extracting the landslide susceptibility evaluation index in step six is ​​as follows:

[0055] Feature visualization: Grad-CAM is used to generate heat maps to identify the areas that contribute the most to the prediction results (near faults, steep slope areas).

[0056] Factor quantification: Combining geomechanical principles, key characteristics are quantified into susceptibility indicators: Lithological factor: weak rock masses are assigned a value of 1, and hard rock masses are assigned a value of 0; Topographic factor: slope > 35 0 Assign a value of 1 if the value is 1, otherwise assign a value of 0.

[0057] Indicator screening: Redundant indicators were eliminated using the Pearson correlation coefficient.

[0058] As a preferred embodiment of the present invention, the construction of the building-topography-geology adaptation evaluation model in step seven includes:

[0059] Fit function: A weighted scoring model is designed by combining the probability of landslide susceptibility and the requirements for building safety.

[0060]

[0061] In the formula, , The weighting coefficients are determined using the analytic hierarchy process. Indicates the probability of landslide susceptibility; Indicates building safety requirements;

[0062] Level Classification:

[0063] Based on the compatibility score, it is divided into three levels: High compatibility (≥0.7): Suitable for construction; Medium compatibility (0.4-0.7): Requires engineering reinforcement; Low compatibility (≤0.4): Construction is prohibited.

[0064] Spatial mapping:

[0065] The adaptation results are overlaid onto the city's 3D model to generate a risk heat map, identifying high-risk building clusters.

[0066] As a preferred embodiment of the present invention, the process for deriving the landslide susceptibility probability is as follows:

[0067] Data fusion and feature extraction:

[0068] In step six, using the multi-source data cube constructed in steps one to four, Grad-CAM is used to extract the regional features that contribute the most to landslide prediction;

[0069] Based on the principles of geomechanics, key features are quantified into susceptibility indicators;

[0070] Model prediction:

[0071] Using the CNN model trained in step five, input the multi-source data cube and output the probability of landslide occurrence for each pixel;

[0072] The model ensures prediction accuracy through historical landslide data and cross-validation;

[0073] Probability mapping:

[0074] The probability values ​​output by the model are superimposed onto the 3D urban model to generate a landslide susceptibility heat map, identifying high-risk areas. >0.7 indicates a high-risk area.

[0075] As a preferred embodiment of the present invention, the method for obtaining the building safety requirements is as follows:

[0076] Building characteristics indicators The safety requirements are quantified by analyzing the building's geometric parameters, and then processed using linear normalization.

[0077]

[0078] In the formula, Indicates building height; Indicates building density; Indicates the base area; , , Represents the weight, take , , ;

[0079] Structural safety indicators The rating is directly assigned based on the standard rating, reflecting the structure's disaster resistance capability.

[0080]

[0081] Use functional indicators Security requirements are adjusted by weighting functional features, with higher requirements for high-risk applications.

[0082]

[0083] Adjustment logic: Demand for labor-intensive buildings increases by 25%; demand for lifeline engineering projects increases by 50%.

[0084] Comprehensive calculation of basic indicators: The three types of indicators are weighted and integrated to obtain the basic value of building safety requirements.

[0085]

[0086] In the formula, Indicates building safety requirements; , , This indicates that the weighting coefficients were determined using the analytic hierarchy process (AHP).

[0087] The beneficial effects of this invention are as follows:

[0088] This invention integrates multi-dimensional data such as remote sensing imagery, geological maps, topographic data, and building distribution to comprehensively characterize the compatibility of buildings, topography, and geology, thereby achieving deep fusion of multi-source data and providing more comprehensive evaluation results. It automatically extracts deep features from the data through the convolutional and pooling layers of a CNN, eliminating the need for manual feature design and improving feature extraction capabilities. The construction of a CNN model that integrates heterogeneous and multi-source data directly captures spatial neighborhood relationships, enhancing spatial correlation. Compared to traditional techniques, this method significantly improves the accuracy and robustness of landslide susceptibility prediction in complex mountainous cities through multi-source data complementarity and deep feature extraction. Attached Figure Description

[0089] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0090] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] like Figure 1 As shown in the figure, this invention provides a CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data. The specific steps are as follows:

[0092] Step 1: Multi-source data collection and processing

[0093] Collect remote sensing image data, geological data, topographic data, and building data, and then preprocess the remote sensing image data, geological data, topographic data, and building data to unify the data format and spatial resolution:

[0094] Register remote sensing images and geological maps, keeping the error within one pixel; unify the spatial resolution of all data and use bilinear interpolation to maintain data continuity; for uninvestigated areas in the geological map, supplement lithological properties using kriging interpolation or neighborhood mean method.

[0095] Step 2: Heterogeneous Data Fusion

[0096] Different data sources are normalized, and the normalized multi-source data are spatially superimposed according to their spatial location to construct a unified multi-source data cube;

[0097] Step 3: Construct a sliding window sample set

[0098] A training sample set is constructed by extracting multi-source data features from local areas using a sliding window.

[0099] Step 4: Design the CNN model architecture

[0100] Construct a CNN model suitable for multi-source data fusion, including an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer;

[0101] Step 5: Model Training

[0102] The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Then, the CNN model is trained by the backpropagation algorithm, and the hyperparameters are optimized to improve performance.

[0103] Step Six: Extraction of Landslide Susceptibility Evaluation Indicators

[0104] Key factors influencing landslide susceptibility are extracted from the constructed CNN model, and a landslide susceptibility evaluation index system is built.

[0105] Step 7: Quantization of Building-Topography-Geology Adaptation

[0106] Based on landslide susceptibility evaluation indicators, a building-topography-geology adaptation evaluation model is constructed.

[0107] Step 8: Model Validation and Dynamic Update

[0108] The model is validated using historical landslide data that was not used in the training process. Combined with UAV aerial surveys and geological drilling, the high-risk areas predicted by the model are verified. The latest remote sensing images and geological survey data are incorporated annually to retrain the model. The model parameters are adjusted based on the field validation results to form a closed-loop optimization of data-model-validation.

[0109] This method solves the problems of single data dimension, insufficient feature extraction capability and weak model generalization of traditional methods by deep fusion of multi-source data, CNN feature extraction and dynamic update mechanism. It achieves high-precision landslide prediction and building safety assessment in complex mountainous cities, and provides scientific decision support for urban planning.

[0110] In step two, the normalization process is as follows: for continuous data, Min-Max normalization or Z-score standardization is used; for discrete data, one-hot encoding is used to convert N types of lithology into N-dimensional binary vectors.

[0111] Min-Max normalization scales the data to the range [0,1], while Z-score standardization makes the mean of the data 0 and the variance 1. For continuous data (such as slope and elevation), Min-Max normalization or Z-score standardization can eliminate the influence of dimensions, accelerate model convergence, and avoid numerical instability. Discrete data (such as lithology categories) are inherently unordered. Directly encoding them with numbers (such as 1,2,3) will introduce incorrect order relationships (the model may mistakenly think that 2 is "greater" than 1). One-hot encoding explicitly represents the category through binary vectors, avoiding such misleading. It converts N types of lithology into N-dimensional binary vectors (such as sandstone = [1,0,0]), with each dimension representing a category, ensuring that the model can independently learn the features of each category.

[0112] The specific method for constructing the training sample set in step three is as follows:

[0113] Window design:

[0114] Set the window size according to the average size of the landslide to ensure that the window contains a complete interactive unit of topography, geology and architecture.

[0115] Sample labels:

[0116] Positive samples: Data within the window centered on the historical landslide boundary is extracted and marked as 1 (prone to occurrence).

[0117] Negative samples: Random sampling is conducted in non-landslide areas to ensure a 1:1 ratio of positive to negative samples and avoid class imbalance.

[0118] Data augmentation:

[0119] 90% of the samples were tested 0 Rotation and horizontal flipping operations expand the dataset to three times its original size, improving the model's generalization ability.

[0120] A sliding window is a rectangular (or higher-dimensional) region of fixed size that gradually covers the entire data range by sliding across the original data in steps. After each slide, the data within the window is extracted as a sample for subsequent feature calculations or directly as input to the model.

[0121] The design method for the CNN model architecture in step four is as follows:

[0122] Input layer: Receives a 50×50×C multi-source data cube;

[0123] Convolution module:

[0124] Convolutional layer 1: 32 3×3 convolutional kernels, stride 1, ReLU activation, extracting low-level features (edges, textures).

[0125] Convolutional layer 2: 64 3×3 convolutional kernels, stride 1, capturing intermediate features (lithological boundaries, topographic relief patterns).

[0126] Pooling layer: 2×2 max pooling, reducing the feature dimension to 25×25×64;

[0127] Fully connected layer:

[0128] Flatten layer: Flattens the pooling output into a 1D vector (25×25×64=40,000 dimensions).

[0129] Dense layer 1: 256 neurons, Dropout=0.5 to prevent overfitting;

[0130] Dense layer 2 (output layer): 1 neuron, sigmoid activation, outputs the slippage probability (0-1).

[0131] The 40,000-dimensional vector flattened from pooling output may cause an explosion of parameters in the fully connected layer (256×40,000≈10 million weights), increasing the risk of overfitting. Adding global average pooling before Flattening compresses 25×25×64 to 1×1×64, which can significantly reduce the number of parameters. If local information needs to be preserved, 1×1 convolution can be used to reduce the dimensionality first (e.g., from 64 channels to 32 channels) before Flattening.

[0132] The formula for calculating the cross-entropy loss function in step five is as follows:

[0133]

[0134] In the formula, This represents the binary cross-entropy loss value; This indicates the number of samples, i.e., the total number of samples in the batch of data; Indicates the first The true label of each sample, with a value of 0 or 1; The model predicts the first The probability value of a sample being a positive class is usually mapped to the [0,1] interval by the Sigmoid activation function;

[0135] When training the CNN model using the backpropagation algorithm in step five:

[0136] Training configuration:

[0137] Optimizer: Adam (learning rate = 0.001, , );

[0138] Batch size: 32, number of iterations = 100;

[0139] Hyperparameter tuning:

[0140] Using Keras Tuner for mesh search, the optimization parameters include:

[0141] Number of convolutional kernels (32 / 64 / 128); learning rate (0.0001 / 0.001 / 0.01); dropout rate (0.3 / 0.5 / 0.7).

[0142] Early shutdown mechanism:

[0143] Monitor the loss on the validation set; if it does not decrease within 10 iterations, terminate training to prevent overfitting.

[0144] Adam is a widely used adaptive optimization algorithm for training neural networks (such as CNNs and RNNs). It combines the ideas of momentum and adaptive learning rate, and can efficiently handle non-stationary objective functions (such as loss surfaces in deep learning), performing well in most tasks. Keras Tuner is an open-source Python library for automating hyperparameter optimization. It aims to help developers efficiently find the best combination of hyperparameters for deep learning models. Users define the hyperparameter search space (such as learning rate, number of layers, number of neurons, etc.), and Keras Tuner automatically tries different combinations to find the optimal solution by training and evaluating model performance.

[0145] The specific method for extracting landslide susceptibility evaluation indicators in step six is ​​as follows:

[0146] Feature visualization: Grad-CAM is used to generate heat maps to identify the areas that contribute the most to the prediction results (near faults, steep slope areas).

[0147] Factor quantification: Combining geomechanical principles, key characteristics are quantified into susceptibility indicators: Lithological factor: weak rock masses are assigned a value of 1, and hard rock masses are assigned a value of 0; Topographic factor: slope > 35 0 Assign a value of 1 if the value is 1, otherwise assign a value of 0.

[0148] Indicator screening: Redundant indicators were eliminated using the Pearson correlation coefficient.

[0149] The core principle of Grad-CAM is as follows: Gradient calculation: The gradient of the target class score with respect to the feature map of the last convolutional layer is calculated through backpropagation. These gradients reflect the importance of the feature map to the classification result. Global average pooling: Global average pooling is performed on the gradients to obtain the weights of each feature map. These weights represent the degree of contribution of the feature map to the classification result. Weighted summation: The weights are multiplied by the feature map of the last convolutional layer to obtain weighted feature maps. These weighted feature maps highlight the regions that have the greatest impact on the classification result. Visualization: The weighted feature maps are upsampled to the size of the input image to generate a heatmap. The red areas in the heatmap represent the parts that have the greatest impact on the classification result, and the blue areas represent the parts that have a smaller impact.

[0150] The construction of the building-terrain-geological compatibility evaluation model in step seven includes:

[0151] Fit function: A weighted scoring model is designed by combining the probability of landslide susceptibility and the requirements for building safety.

[0152]

[0153] In the formula, , The weighting coefficients are determined using the analytic hierarchy process. Indicates the probability of landslide susceptibility; Indicates building safety requirements;

[0154] Level Classification:

[0155] Based on the compatibility score, it is divided into three levels: High compatibility (≥0.7): Suitable for construction; Medium compatibility (0.4-0.7): Requires engineering reinforcement; Low compatibility (≤0.4): Construction is prohibited.

[0156] Spatial mapping:

[0157] The adaptation results are overlaid onto the city's 3D model to generate a risk heat map, identifying high-risk building clusters.

[0158] Analytic Hierarchy Process (AHP) for Determining Weight Coefficients , The core steps are as follows:

[0159] Construct a hierarchical model: Goal layer: Define the ultimate goal of the decision; Criteria layer: List the key factors that affect the goal; Alternatives layer: List all options to be evaluated;

[0160] Constructing the judgment matrix: Pairwise comparison: For each element in the same layer, compare the relative importance of the elements in the same layer to a certain element in the upper layer through expert scoring or data quantification; Scaling system: Use the 1-9 scaling method to quantify the comparison results;

[0161] Calculate the weight vector: Calculate the largest eigenvalue of the judgment matrix and its corresponding eigenvector, normalize the eigenvector, and obtain the weight coefficient of each element;

[0162] Consistency check: Calculate the consistency index, find random consistency indicators, and calculate the consistency ratio;

[0163] Comprehensive weight calculation: Multiply the weight of the criterion layer with the weight of each criterion in the scheme layer to obtain the comprehensive weight of the scheme layer in relation to the target layer.

[0164] The process for deriving the probability of landslide susceptibility is as follows:

[0165] Data fusion and feature extraction:

[0166] In step six, using the multi-source data cube constructed in steps one to four, Grad-CAM is used to extract the regional features that contribute the most to landslide prediction;

[0167] Based on the principles of geomechanics, key features are quantified into susceptibility indicators;

[0168] Model prediction:

[0169] Using the CNN model trained in step five, input the multi-source data cube and output the probability of landslide occurrence for each pixel;

[0170] The model ensures prediction accuracy through historical landslide data and cross-validation;

[0171] Probability mapping:

[0172] The probability values ​​output by the model are superimposed onto the 3D urban model to generate a landslide susceptibility heat map, identifying high-risk areas. >0.7 indicates a high-risk area.

[0173] When a model ensures prediction accuracy through historical landslide data and cross-validation, the AUC value can be used to determine this. The AUC value represents the probability that the model's prediction probability for the positive sample is greater than its prediction probability for the negative sample when a positive sample and a negative sample are randomly selected. An AUC of 0.8 means that the model can correctly distinguish between positive and negative samples in 80% of cases.

[0174] The method for obtaining building safety requirements is as follows:

[0175] Building characteristics indicators The safety requirements are quantified by analyzing the building's geometric parameters, and then processed using linear normalization.

[0176]

[0177] In the formula, Indicates building height; Indicates building density; Indicates the base area; , , Represents the weight, take , , ;

[0178] Structural safety indicators The rating is directly assigned based on the standard rating, reflecting the structure's disaster resistance capability.

[0179]

[0180] Use functional indicators Security requirements are adjusted by weighting functional features, with higher requirements for high-risk applications.

[0181]

[0182] Adjustment logic: Demand for labor-intensive buildings increases by 25%; demand for lifeline engineering projects increases by 50%.

[0183] Comprehensive calculation of basic indicators: The three types of indicators are weighted and integrated to obtain the basic value of building safety requirements.

[0184]

[0185] In the formula, Indicates building safety requirements; , , This indicates that the weighting coefficients were determined using the analytic hierarchy process (AHP).

[0186] The structural safety indicators are based on the standards "Standard for Reliability Appraisal of Civil Buildings" (GB 50292-2015) and "Standard for Seismic Appraisal of Buildings" (GB 50023-2009). When S is 0-0.4, the safety requirement level is low, requiring maintenance of the status quo and regular inspections. When S is 0.4-0.7, the safety requirement level is medium, requiring local reinforcement and restriction of usage. When S is 0.7-1.0, the safety requirement level is high, requiring overall renovation or relocation. When S is 1.0-1.2, the safety requirement level is extremely high, requiring emergency evacuation and priority resource allocation.

[0187] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0188] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A CNN-based method for quantifying building-topography-geology adaptation and predicting landslide susceptibility in mountainous cities by integrating multi-source data, characterized in that... The specific steps are as follows: Step 1: Multi-source data collection and processing Collect remote sensing image data, geological data, topographic data, and building data, and then preprocess the remote sensing image data, geological data, topographic data, and building data to unify the data format and spatial resolution: Register remote sensing images and geological maps, keeping the error within one pixel; unify the spatial resolution of all data and use bilinear interpolation to maintain data continuity; for uninvestigated areas in the geological map, supplement lithological properties using kriging interpolation or neighborhood mean method. Step 2: Heterogeneous Data Fusion Different data sources are normalized, and the normalized multi-source data are spatially superimposed according to their spatial location to construct a unified multi-source data cube; Step 3: Construct a sliding window sample set A training sample set is constructed by extracting multi-source data features from local areas using a sliding window. Step 4: Design the CNN model architecture Construct a CNN model suitable for multi-source data fusion, including an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer; Step 5: Model Training The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Then, the CNN model is trained by the backpropagation algorithm, and the hyperparameters are optimized to improve performance. Step Six: Extraction of Landslide Susceptibility Evaluation Indicators Key factors influencing landslide susceptibility are extracted from the constructed CNN model, and a landslide susceptibility evaluation index system is built. Step 7: Quantization of Building-Topography-Geology Adaptation Based on landslide susceptibility evaluation indicators, a building-topography-geology adaptation evaluation model is constructed. Step 8: Model Validation and Dynamic Update The model is validated using historical landslide data that was not used in the training process. Combined with UAV aerial surveys and geological drilling, the high-risk areas predicted by the model are verified. The latest remote sensing images and geological survey data are incorporated annually to retrain the model. The model parameters are adjusted based on the field validation results to form a closed-loop optimization of data-model-validation.

2. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 1, characterized in that: The normalization process described in step two is as follows: for continuous data, Min-Max normalization or Z-score standardization is used; for discrete data, one-hot encoding is used to convert N types of lithology into N-dimensional binary vectors.

3. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 1, characterized in that: The specific method for constructing the training sample set mentioned in step three is as follows: Window design: Set the window size according to the average size of the landslide to ensure that the window contains a complete interactive unit of topography, geology and architecture. Sample labels: Positive samples: Data within the window centered on the historical landslide boundary is extracted and marked as 1 (prone to occurrence). Negative samples: Random sampling is conducted in non-landslide areas to ensure a 1:1 ratio of positive to negative samples and avoid class imbalance. Data augmentation: 90% of the samples were tested 0 Rotation and horizontal flipping operations expand the dataset to three times its original size, improving the model's generalization ability.

4. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 1, characterized in that: The design method of the CNN model architecture described in step four is as follows: Input layer: Receives a 50×50×C multi-source data cube; Convolution module: Convolutional layer 1: 32 3×3 convolutional kernels, stride 1, ReLU activation, extracting low-level features (edges, textures). Convolutional layer 2: 64 3×3 convolutional kernels, stride 1, capturing intermediate features (lithological boundaries, topographic relief patterns). Pooling layer: 2×2 max pooling, reducing the feature dimension to 25×25×64; Fully connected layer: Flatten layer: Flattens the pooling output into a 1D vector (25×25×64=40,000 dimensions). Dense layer 1: 256 neurons, Dropout=0.5 to prevent overfitting; Dense layer 2 (output layer): 1 neuron, sigmoid activation, outputs the slippage probability (0-1).

5. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 1, characterized in that: The formula for calculating the cross-entropy loss function mentioned in step five is as follows: In the formula, This represents the binary cross-entropy loss value; This indicates the number of samples, i.e., the total number of samples in the batch of data; Indicates the first The true label of each sample, with a value of 0 or 1; The model predicts the first The probability value of a sample being a positive class is usually mapped to the [0,1] interval by the Sigmoid activation function; When training the CNN model using the backpropagation algorithm described in step five: Training configuration: Optimizer: Adam (learning rate = 0.001, , ); Batch size: 32, number of iterations = 100; Hyperparameter tuning: Using Keras Tuner for mesh search, the optimization parameters include: Number of convolutional kernels (32 / 64 / 128); learning rate (0.0001 / 0.001 / 0.01); dropout rate (0.3 / 0.5 / 0.7). Early shutdown mechanism: Monitor the loss on the validation set; if it does not decrease within 10 iterations, terminate training to prevent overfitting.

6. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 1, characterized in that: The specific method for extracting landslide susceptibility evaluation indicators as described in step six is ​​as follows: Feature visualization: Grad-CAM is used to generate heat maps to identify the areas that contribute the most to the prediction results (near faults, steep slope areas). Factor quantification: Combining geomechanical principles, key characteristics are quantified into susceptibility indicators: Lithological factor: weak rock masses are assigned a value of 1, and hard rock masses are assigned a value of 0; Topographic factor: slope > 35 0 Assign a value of 1 if the value is 1, otherwise assign a value of 0. Indicator screening: Redundant indicators were eliminated using the Pearson correlation coefficient.

7. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 1, characterized in that: The construction of the building-topography-geology adaptation evaluation model described in step seven includes: Fit function: A weighted scoring model is designed by combining the probability of landslide susceptibility and the requirements for building safety. In the formula, , The weighting coefficients are determined using the analytic hierarchy process. Indicates the probability of landslide susceptibility; Indicates building safety requirements; Level Classification: Based on the compatibility score, it is divided into three levels: High compatibility (≥0.7): Suitable for construction; Medium compatibility (0.4-0.7): Requires engineering reinforcement; Low compatibility (≤0.4): Construction is prohibited. Spatial mapping: The adaptation results are overlaid onto the city's 3D model to generate a risk heat map, identifying high-risk building clusters.

8. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 7, characterized in that: The process for deriving the landslide susceptibility probability is as follows: Data fusion and feature extraction: In step six, using the multi-source data cube constructed in steps one to four, Grad-CAM is used to extract the regional features that contribute the most to landslide prediction; Based on the principles of geomechanics, key features are quantified into susceptibility indicators; Model prediction: Using the CNN model trained in step five, input the multi-source data cube and output the probability of landslide occurrence for each pixel; The model ensures prediction accuracy through historical landslide data and cross-validation; Probability mapping: The probability values ​​output by the model are superimposed onto the 3D urban model to generate a landslide susceptibility heat map, identifying high-risk areas. >0.7 indicates a high-risk area.

9. The CNN-based method for quantifying the building-topography-geology fit of mountainous cities and predicting landslide susceptibility by fusing multi-source data as described in claim 7, characterized in that: The method for obtaining the building safety requirements is as follows: Building characteristics indicators The safety requirements are quantified by analyzing the building's geometric parameters, and then processed using linear normalization. In the formula, Indicates building height; Indicates building density; Indicates the base area; , , Represents the weight, take , , ; Structural safety indicators The rating is directly assigned based on the standard rating, reflecting the structure's disaster resistance capability. Use functional indicators Security requirements are adjusted by weighting functional features, with higher requirements for high-risk applications. Adjustment logic: Demand for labor-intensive buildings increases by 25%; demand for lifeline engineering projects increases by 50%. Comprehensive calculation of basic indicators: The three types of indicators are weighted and integrated to obtain the basic value of building safety requirements. In the formula, Indicates building safety requirements; , , This indicates that the weighting coefficients were determined using the analytic hierarchy process (AHP).

Images