Spatial-temporal multi-scale feature fusion deep learning landslide susceptibility evaluation method based on digital object dual drive

Abstract

The invention discloses a time-space multi-scale feature fusion deep learning landslide susceptibility evaluation method based on number object dual drive, and the method comprises the steps: constructing a time sequence branch + Swindow-Transformer (upper branch) + CNN (lower branch) time-space fusion DL-LSA model, combining an AFF multi-scale space fusion module and a TSF time-space fusion module, and carrying out the deep learning of landslide susceptibility. A complete technical scheme of high-quality sample screening, deep spatial-temporal feature extraction, multi-scale adaptive fusion and dynamic probability prediction is formed, and normal form upgrading of geological disaster risk management from single-drive vector object double-drive and from static evaluation to dynamic early warning is promoted. Comprising the following steps: 1, acquiring influence factors (X); 2, constructing an initial sample set (X-Y pairs); 3, preprocessing the influence factor (X); and 4, influence factor coding (generating a standardized feature X '). And 5, providing physical constraints (optimizing a sample set X '-Y) based on the P-LSA model. And 6, constructing a space-time fusion DL-LSA model. And 7, designing a feature fusion module. And 8, outputting a landslide susceptibility evaluation result.

Description

Technical Field

[0001] This invention belongs to the field of landslide geological hazard risk assessment. Background Technology

[0002] Landslides are among the most destructive disasters globally, posing a serious threat to human life and property. According to the Global Natural Disaster Assessment Report, landslides have become increasingly frequent in recent years, resulting in a significant increase in the number of deaths, affected populations, and economic losses. However, less than 20% of landslide sites have been identified, making the risk prevention and control situation extremely challenging.

[0003] Landslide susceptibility (LS) refers to the probability of a landslide occurring in a region under specific geological conditions. Landslide susceptibility assessment (LSA) provides crucial support for decision-making processes such as regional land spatial planning and engineering site selection optimization, playing an important role in disaster prevention and mitigation. Based on the assessment method, LSA can be divided into physics-based models (P-LSA) and data-driven models (D-LSA), with D-LSA models having long held a dominant position.

[0004] Deep learning (DL), constructed from multi-layered neural networks, can automatically extract feature information (surface deformation features) from raw data (such as SAR images) with little or no preprocessing, capturing nonlinear spatiotemporal correlations. It has become the core technology of state-of-the-art D-LSA models. Convolutional neural networks (CNNs) are suitable for image recognition, object recognition, and other fields. Transformer models, based on attention mechanisms, are adapted for multimodal information fusion and are mainly used in computer vision. Long Short-Term Memory (LSTM) networks excel in temporal data processing, controlling information flow through gating mechanisms to improve model training efficiency and prediction performance.

[0005] The southeastern coastal region of China is a high-incidence area for landslides, characterized by dense hills, intersecting coastal zones, and widespread distribution of soft soil and thick layers of granite residual soil. Earthquakes and extreme weather events (such as typhoons and torrential rains) are frequent, and engineering activities are active. Under the combined effects of these multiple factors, landslide geological hazards are characterized by wide distribution, strong concealment, and high time variability. Therefore, traditional LSA models have poor adaptability, posing challenges to their application in this region. 1. Landslides are widespread in the southeastern coastal areas and mostly occur in remote mountainous regions, resulting in scarce data. 2. Landslides are mainly shallow and small in volume, making them highly concealed. Combined with a high-humidity vegetation environment, the monitoring capabilities of optical / SAR images are limited, which can easily lead to "missed detection and misjudgment" and a large deviation from the actual situation. 3. The disaster-prone environment in this region is complex, and the D-LSA model suffers from the "black box effect," making it prone to overfitting and resulting in low accuracy in scenarios with scarce data; the P-LSA model relies on a large number of physical parameters, and the parameters have strong spatiotemporal heterogeneity, limiting its generalization ability. Neither of these models is suitable for this region. 4. The region has fragile geological conditions and is highly susceptible to environmental changes (mainly heavy rainfall). It is a time-varying system, while existing LSA models are mostly static evaluations, and the results are out of sync with the actual disaster-causing process. 5. The landslides in this region simultaneously exhibit macroscopic patterns of clustered landslides and microscopic mechanisms of deformation and fissures leading to disasters, with significant size differences. However, the existing LSA model uses a single grid size and cannot take into account the multi-scale characteristics. 6. High-dimensional data (such as InSAR and LiDAR data) cause the computational complexity of traditional Transformer models to increase quadratically, resulting in low operating efficiency and difficulty in engineering applications.

[0006] The core shortcomings of existing technologies are: scarce and biased samples, insufficient model interpretability, inconsistency between static evaluation and dynamic disaster-causing process, difficulty in multi-scale feature fusion, and high computational burden. There is an urgent need for a high-precision LSA method adapted to the southeastern coastal region. Summary of the Invention

[0007] In view of the technical problems in the background technology, this invention proposes a landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical data. This method aims to overcome the bottlenecks of existing technologies and solve a series of key problems faced by existing LSA technology in practical applications in this region, including data scarcity, large bias, inconsistency between static evaluation and dynamic disaster-causing processes, difficulty in integrating multi-scale and multi-modal data, and high computational burden. It significantly improves the accuracy and reliability of LSA for landslide risk assessment under complex geological conditions, provides more scientific decision support for landslide prevention and mitigation, and has important engineering application prospects and scientific value.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A landslide susceptibility assessment method based on spatiotemporal multi-scale feature fusion driven by both data and physical factors includes the following steps: S1. Obtaining the Impact Factor (X) Five major categories (16 items in total) of various influencing factors causing landslide disasters were identified, and the process was moved to S2.

[0009] S2. Construction of the initial sample set (XY pairs) The original positive sample (Y=1) data is obtained from landslide points surveyed in the study area / historical records, and the positive sample is expanded based on SMOTE. Negative samples (Y=0) are generated by randomly selecting points within the buffer. The positive and negative samples are merged to form the initial sample set (XY pairs), which is then processed in S3.

[0010] S3, Impact Factor (X) Preprocessing The correlation coefficient and VIF variance inflation factor were used to determine the correlation, and redundant influencing factors were removed until |rij| of all influencing factors in the initial sample set of S2 was less than 0.8 and VIF was less than 10.

[0011] S4. Impact Factor Coding (Generating Standardized Feature X') The discrete and continuous types of influencing factors are coded separately to generate standardized features X', and the rationality is verified by the frequency ratio method to form a sample set (X'-Y pairs).

[0012] S5. Provide physical constraints based on the P-LSA model (optimize the sample set X'-Y). Key mechanical parameters are correlated with features (X'), Fs is solved by a modified formula, prior constraints are provided by the P-LSA model, susceptibility levels are divided according to the magnitude of Fs values, and positive / negative samples in S4 that fall within the range of extremely high / low susceptibility are retained to form a high-quality X'-Y sample set.

[0013] S6. Construction of Spatiotemporal Fusion DL-LSA Model Based on the S4 normalized feature X', an LSTM network is used to capture temporal features and generate a global temporal feature vector S. At the same time, the upper and lower branches of "Swin-Transformer+CNN" are used to extract global (Gl)-local spatial features (Ll) and it undergoes multiple rounds of learning iteration.

[0014] S7, Feature Fusion Module Design On the output side of the spatiotemporal fusion DL-LSA model, feature fusion is further completed by sequentially connecting the AFF module and the TSF module. The AFF module realizes the fusion of multi-scale spatial features (Gl, Ll) to form spatial fusion features (Hl), and the TSF module realizes the fusion of spatiotemporal features. S8. Output landslide susceptibility assessment results Output the model's predicted probability, plot the evaluation results, classify the monitored area into different levels, and verify the results based on the S5 sample set labels.

[0015] The beneficial effects of this invention are: To address the four core pain points in the assessment of landslide susceptibility in the southeast coastal region caused by rainfall—namely, scarce and biased samples, significant dynamic disaster-causing effects coupled with statically outdated assessment methods, difficulties in multi-scale spatial feature fusion and high computational cost, and insufficient fusion of heterogeneous spatiotemporal feature dimensions—this paper constructs a spatiotemporal fusion DL-LSA model based on "temporal branch + Swing-Transformer (upper branch) + CNN (lower branch)." Combined with AFF multi-scale spatial fusion and TSF spatiotemporal fusion modules, this forms a complete technical solution encompassing "high-quality sample screening, deep spatiotemporal feature extraction, multi-scale adaptive fusion, and dynamic probability prediction." This promotes a paradigm shift in geological disaster risk management from a single-driven approach to a dual-driven approach based on data and physical factors, and from static assessment to dynamic early warning. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Figure 2 This is a schematic diagram of the network structure of the present invention. Figure 3 for Figure 2 AFF structure diagram Figure 4 for Figure 2 Schematic diagram of TSF structure Figure 5 for Figure 4 Schematic diagram of MS-CAM Detailed Implementation The technical solution of the present invention will be further described in detail with reference to the accompanying drawings.

[0017] A landslide susceptibility assessment method based on spatiotemporal multi-scale feature fusion driven by both data and physical factors includes the following steps: S1. Obtaining the Impact Factor (X) Five major categories (16 items in total) of various influencing factors causing landslide disasters were identified, and the process was moved to S2.

[0018] S2. Construction of the initial sample set (XY pairs) The original positive sample (Y=1) data is obtained from landslide points surveyed in the study area / historical records, and the positive sample is expanded based on SMOTE. Negative samples (Y=0) are generated by randomly selecting points within the buffer. The positive and negative samples are merged to form the initial sample set (XY pairs), which is then processed in S3.

[0019] S3, Impact Factor (X) Preprocessing The correlation coefficient and VIF variance inflation factor were used to determine the correlation, and redundant influencing factors were removed until |rij| of all influencing factors in the initial sample set of S2 was less than 0.8 and VIF was less than 10.

[0020] S4. Impact Factor Coding (Generating Standardized Feature X') The discrete and continuous types of influencing factors are coded separately to generate standardized features X', and the rationality is verified by the frequency ratio method to form a sample set (X'-Y pairs).

[0021] S5. Provide physical constraints based on the P-LSA model (optimize the sample set X'-Y). Key mechanical parameters are correlated with features (X'), Fs is solved by a modified formula, prior constraints are provided by the P-LSA model, susceptibility levels are divided according to the magnitude of Fs values, and positive / negative samples in S4 that fall within the range of extremely high / low susceptibility are retained to form a high-quality X'-Y sample set.

[0022] S6. Construction of Spatiotemporal Fusion DL-LSA Model Based on the S4 normalized feature X', an LSTM network is used to capture temporal features and generate a global temporal feature vector S. At the same time, the upper and lower branches of "Swin-Transformer+CNN" are used to extract global (Gl)-local spatial features (Ll) and it undergoes multiple rounds of learning iteration.

[0023] S7, Feature Fusion Module Design On the output side of the spatiotemporal fusion DL-LSA model, feature fusion is further completed by sequentially connecting the AFF module and the TSF module. The AFF module realizes the fusion of multi-scale spatial features (Gl, Ll) to form spatial fusion features (Hl), and the TSF module realizes the fusion of spatiotemporal features. S8. Output landslide susceptibility assessment results Output the model's predicted probability, plot the evaluation results, classify the monitored area into different levels, and verify the results based on the S5 sample set labels.

[0024] Note: Sample label (Y): refers to the category identifier (dependent variable) of the sample. Landslide areas are labeled as 1 (positive sample) and non-landslide areas are labeled as 0 (negative sample). Sample set: A set of XY pairs consisting of "influence factor values ​​(X) + sample labels (Y)", which is the basic data for model training; Feature (X'): refers to the structured and effective information of the influencing factor (X) after preprocessing, redundancy removal, and hierarchical encoding, which can be directly input into the deep learning model and is transformed from X.

[0025] S1 Impact Factor (X) Acquisition Focusing on the disaster-causing mechanism of landslides in the southeast coastal area, data on 16 core LSA (Leading Sequence of Action) factors (X) in 5 categories were collected, as follows: Topographic factors (5 items): aspect, slope, elevation, profile curvature, and plan curvature; This data comes from digital elevation model (DEM) data acquired through aerial surveying or ground observation, such as satellite remote sensing, GNSS, Insar, and UAVs. Geological factors (1 item): Lithology; This data comes from regional geological survey results, borehole and exploration data. Hydrological factors (5 items): 7-day effective precipitation, rainfall intensity, distance from the river system, topographic humidity index (TWI), and flow intensity index (SPI). The data comes from meteorological station observations, official hydrological datasets, and DEM data. Land cover factors (3 items): Normalized Difference Vegetation Index (NDVI), distance from road, and land use type; The data comes from satellite remote sensing imagery, OpenStreetMap (OSM), and the Ministry of Natural Resources' "National Land Use Status Classification". Climate factors (2 items): annual average temperature, surface temperature.

[0026] The data comes from the climate dataset of the National Meteorological Information Center and the Landsat thermal infrared band of the GEE platform.

[0027] Note: The above-mentioned profile curvature, plane curvature, 7-day effective precipitation, rainfall intensity, topographic humidity index (TWI), flow intensity index (SPI), normalized difference vegetation index (NDVI), annual average temperature, and surface temperature are unique influencing factors designed for the landslide characteristics of the southeastern coastal area in this invention.

[0028] S2. Construct the initial sample set (XY pairs) Using "X value + Y label" as the core, a complete initial sample set is constructed to solve the problem of sample scarcity: S2.1 Collection and Augmentation of Positive Samples (Y=1) S2.1.1 Using the landslide points from the field survey / historical records in the study area as the original positive sample (number denoted as N), extract the values ​​of the corresponding 16 influencing factors (X) to form the original XY pairs (Y=1). S2.1.2. Use SMOTE oversampling to generate 29N positive samples. S2.1.2.1 For each positive sample (i=1, 2, ..., N), use Euclidean distance in the feature space to find its k nearest positive sample neighbors (k is usually 5). The Euclidean distance formula is: For feature dimension, , For the sample , exist The values ​​that can be taken on each feature.

[0029] S2.1.2.2, Randomly select a sample from the k nearest neighbors. New samples are synthesized through linear interpolation: in, Original positive samples; : Random nearest neighbor sample; δ is a random number between 0 and 1 (e.g., δ=0.3).

[0030] S2.1.2.3 Repeat the above process until 29N synthetic samples are generated.

[0031] S2.2 Selection of negative samples (Y=0) S2.2.1. Using the geographical location of the expanded positive sample as the center, divide an 800-meter buffer zone (based on the geological survey data of the area, the buffer zone is a potential landslide area and is not included in the selection of negative samples); S2.2.2 In the area outside the buffer zone, randomly select points at a ratio of 5 times the number of positive samples (maintaining a positive sample: negative sample ratio of 1:5 to achieve class balance between positive and negative samples), and mark them as negative samples (label Y=0). S2.3 Initial Sample Set Generation: Merge positive and negative samples to form an initial XY pair sample set of "influence factor value (X) + sample label (Y)", providing basic data for subsequent S3 preprocessing.

[0032] S3. Preprocess the impact factor X in S1 to select the impact factors that are actually suitable for entering S4. Note: The preprocessing procedure is specifically designed for subsequent spatiotemporal fusion deep learning models. Its core objectives are: 1. To eliminate redundant factors and reduce the computational complexity of high-dimensional data; 2. To unify the data format and adapt to the tensor input requirements of DL models; 3. To eliminate dimensional differences, avoid oscillations during gradient descent optimization, and improve model convergence efficiency. Pearson correlation coefficient and VIF variance inflation factor are used in a complementary screening process. The specific steps are as follows: S3.1, Pearson correlation coefficient screening S3.1.1 Calculation of Pearson Correlation Coefficient Matrix Calculate any number of impact factors in the impact factor dataset X. and Pearson correlation coefficient among (i, j∈[1,16], i≠j) The formula is: in: : No. Influence factors in each sample Values; Impact Factor The mean; Total number of samples (the total number of landslide points + the total number of non-landslide points); S3.1.2 Correlation Judgment If | | ≥0.8 indicates a strong linear correlation between the two influencing factors, and factors with ambiguous physical meanings should be eliminated; S3.2, VIF Collinearity Screening S3.2.1 Constructing a linear regression model Using impact factor data X each factor With as the dependent variable and all other factors as independent variables, construct a linear regression model: = + + +...+ + +...+ + in: Dependent variable; (j≠i): Independent variable (except for) (All other influencing factors besides) Total number of impact factors; : Intercept term; : Regression coefficient; Error term.

[0033] S3.2.2 Calculate the coefficient of determination R² in: Factors in the k-th sample The original observations; Calculated using a linear regression model The estimated value; S3.2.3 Calculate the VIF value for each factor The formula for calculating VIF is: in: The coefficient of determination for the linear regression model in S3.2.2 S3.2.4 Collinearity Judgment If VIF ≥ 10 (5 in strict scenarios), it indicates that the factor has severe collinearity with other factors, and the influencing factor should be removed.

[0034] S3.3, Iterative Check Repeat steps 3.1-3.2 until all impact factors are covered. | < 0.8 and VIF < 10.

[0035] S4. The preprocessed set of influence factors X from S3 is divided into two categories based on spatial distribution characteristics: discrete influence factors and continuous influence factors. These are then coded to generate standardized features X'. The frequency ratio method is used to verify the rationality of the coding. S4.1 Discrete Influencing Factors (specifically, the lithology and land use items in S1): Based on the inherent classification system coding, the larger the coding value, the stronger the landslide promoting effect. The lithology is classified according to the rock type identified in the geological survey (such as "granite", "shale", "sandstone", etc.), which is an inherent classification system. The land use types are classified according to the classification standards of remote sensing image interpretation (such as "forest land", "cultivated land", "construction land", etc.), which is the inherent classification system.

[0036] S4.2, Continuous Influencing Factors (specifically referring to the 14 items in S1.1: slope, aspect, elevation, profile curvature, plane curvature, 7-day effective precipitation, rainfall intensity, distance from water system, TW, SPI, NDVI, distance from road, annual average temperature, and surface temperature): Classified using the natural discontinuity method, with larger coded values ​​indicating stronger promoting effects. S4.2.1 Data Sorting: Sort all sample values ​​of continuous impact factors in ascending order. S4.2.2 Breakpoint Calculation: The intervals are automatically divided (usually into 3 levels) using the statistical method of "minimizing within-class variance and maximizing between-class variance". S4.2.3 Standardized Coding: Assigning values ​​to the divided intervals; the larger the value, the stronger the promotion effect. Among them, rainfall intensity (I) reflects the typical characteristics of short-term torrential rain during the typhoon season in the southeast coastal areas. The calculation method is as follows: in Rainfall (mm) over a specific period. The duration of rainfall (h).

[0037] Effective cumulative rainfall over 7 days This reflects the "cumulative effect" (excluding ineffective losses such as surface runoff and evaporation). The calculation method is as follows: in, (Daily moisture content decay coefficient). Let be the rainfall (mm) on day i.

[0038] TWI calculation method: TWI = ln(As / tanβ) Where As: catchment area per unit contour line length (m² / m), reflecting the upstream water collection capacity; tanβ: tangent of the slope (dimensionless), reflecting the potential for water flow velocity; ln: natural logarithm function.

[0039] SPI calculation method: SPI = As×tanβ Where As: catchment area per unit contour line length (m² / m), i.e., cumulative flow × grid resolution; tanβ: tangent of slope (dimensionless). NDVI formula: Wherein, NIR: near-infrared reflectance; R: red reflectance. S4.2.3. To enhance the rationality, the frequency ratio (FR) method should be used for verification: Among them, The number of landslide grid cells in the j-th class of the i-th influence factor. This represents the total number of landslide grids in the study area. Let S be the total number of raster cells in the j-th category of the i-th impact factor, and S be the total number of raster cells in the study area. The frequency ratio of the i-th influence factor in the j-th category; like >1: This interval is positively correlated with landslides (the larger the value, the stronger the correlation). If 0.5 ≤ ≤ 1: The correlation is weak, but there is no need to merge them; like <0.5: This interval has almost no correlation with landslides and needs to be merged with adjacent intervals (e.g., merge "0-5°" and "5-15°" into "0-15°").

[0040] S5: Construct a P-LSA model (infinite slope stability analysis) and perform LSA. Select reasonable positive / negative samples from areas with extremely high / low susceptibility to form a high-quality sample set (X'-Y sample pairs). S5.1, P-LSA Parameters and Characteristics (X'): 6 Key Mechanical Parameters (c', , h , All of these are derived from the quantization assignment of feature (X'). Note: Effectively avoid the "black box problem" of machine learning and ensure complete coupling between physical computation and the impact factor system: S5.2 Correction formula for P-LSA model (based on weighted influence factor fusion comprehensive judgment value) in: Land use correction coefficient (-0.2 for construction land, +0.15 for forest land); Vegetation cover correction factor (NDVI > 0.5, take +0.1; < 0.3, take -0.1). Rainfall intensity correction factor (+0.2 for >30mm / h, 0 for ≤30mm / h).

[0041] S5.3, using Classification of susceptibility levels S5.4 Sample Screening Retain areas with extremely high susceptibility ( Positive samples within the range (Y=1), and extremely low regions ( The negative samples (Y=0) within the range are used to form a high-quality X'-Y sample set.

[0042] S6. Spatiotemporal Fusion DL-LSA Model Construction A "temporal branch + spatial dual-path branch" architecture is constructed. Based on the standardized features (X') generated by S4, a feature map is built, and deep spatiotemporal features are extracted autonomously to achieve dynamic dependency capture and full coverage of multi-scale spatial features. S6.1, Timing Branching (LSTM Captures Dynamic Dependencies); S6.1.1, Time Series Data Structuring The dynamic influencing factor in S4 is: 7-day effective precipitation (P) e The standardized features (X') corresponding to rainfall intensity (I) are converted into a structured feature vector at time t: in: Standardized characteristics of the 7-day effective precipitation at time t; Standardized characteristics of rainfall intensity at time t; For feature dimension ( ).

[0043] S6.1.2 Data Normalization For the original impact factors (characteristics) Min-max normalization is performed to eliminate dimensional differences and adapt to LSTM input requirements: in: The standardized features corresponding to the 7-day effective precipitation and rainfall intensity in S4 are the global minimum / maximum values.

[0044] S6.1.3, Constructing sequences using the sliding window method: S6.1.3.1. Set the sliding window size W = 7 days (to adapt to the cumulative disaster effect of rainfall), the sliding step is 1 day, and slice the continuous time series features into input sequences of fixed length (7). Each sequence corresponds to one evaluation day, and the window contains feature data for 7 consecutive days: in The input sequence matrix of the LSTM is... (k=0,1,...,6) represents the time series feature vector after normalization for the previous k days, used to evaluate the current day; To evaluate the normalized time series feature vector for the day, m=2 is the feature dimension.

[0045] S6.1.3.2, Map the daily feature vector within the window to the target dimension using a fully connected layer: in: This is the mapping weight matrix; It is the bias vector; It is the ReLU activation function; d represents the mapped daily feature vector; d represents the mapped feature dimension.

[0046] S6.1.3.3 Constructing the input matrix of LSTM: in: The feature vectors mapped over 7 days within the sliding window are arranged in chronological order and directly input into the LSTM for time series modeling. This is the single-day feature vector after mapping for that day.

[0047] S6.1.4, LSTM Temporal Feature Modeling LSTM dynamically updates cell states through a gating mechanism, capturing the temporal dependencies of daily features within a window, as shown in the following formula: Forget Gate (Proportion of cell states retained in history): Input gate (the proportion of new information updated to the cell state): Candidate cell states (candidate values ​​for new information): Cell state update (fusion of historical state and new information): Output gate (output ratio in hidden state): Hidden state (output feature at the current moment): in: : Daily LSTM hidden state (dimension d); Features entered on the same day; : LSTM cell state (dimension d); : Gating weight matrix (dimension d); : Gating bias vector (dimension d); Hadamard product (element-wise multiplication); Sigmoid activation function (outputs 0~1, controls the proportion of information flow). Hyperbolic tangent function (output -1 to 1, adjust information intensity).

[0048] Output: After modeling with 7 days of data within a sliding window, the LSTM layer outputs the complete set of hidden states for the entire sequence: in: A W×d hidden state matrix representing the entire sequence, containing the hidden states for 7 days within a window; (all dimensions are d).

[0049] S6.1.5 Attention Mechanism: Global Temporal Feature Aggregation By using attention-weighted aggregation, key date features within the window are highlighted to generate global time-series features: S6.1.5.1 Calculate daily hidden status Attention score: in, , , Here, k represents the attention parameter, and k represents the hidden layer dimension.

[0050] S6.1.5.2, Convert the scores into normalized weights using the Softmax function: S6.1.5.3, Weighted summation to generate global temporal feature vectors: S6.2, Spatial Branching Using the multi-channel normalized feature map (X') generated by S4 as input, a dual-path structure of "upper branch global spatial feature extraction + lower branch local spatial feature extraction" is constructed. Figure 2 These correspond to the clustering patterns and fissures (regional-level correlation features) and deformation mechanisms (pixel-level detail features) in the southeastern coastal region, respectively.

[0051] S 6.2.1, Upper Branch: Global Spatial Feature Extraction (Hierarchical Swing-Transformer Block Encoding) S 6.2.1.1, Patch segmentation and embedding (adapting input to Transformer) Multi-channel standardized feature maps (K is the total number of encoded features) is the input, and the data is divided into blocks without repetition. The patch (e.g., 2×2) is flattened into a vector sequence by sliding window (unfold) and arranged in spatial order. =Unfold(X', kernel_size=P, stride=P) The Unfold output shape is as follows: kernel_size is the kernel size, stride is the stride. Then through a linear mapping layer ( Embedded in high-dimensional space: The Linear layer adjusts the Patch vector dimension to the Transformer hidden dimension C (e.g., 96), and outputs... .

[0052] S 6.2.1.2, Hierarchical Swing Transformer Block Global Feature Extraction Through multi-level (four-level example setup) Swin Transformer Blocks (alternating window multi-head self-attention W-MSA and sliding window multi-head self-attention SW-MSA), combined with Patch Merging downsampling (halving the resolution and doubling the number of channels at each level), global features G0, G1, G2, and G3 are gradually generated. in: For the first Level input features; : No. Level global features (resolution of Corresponding to the number of channels, such as for , for ).

[0053] S 6.2.2, Lower Branch: Local Spatial Feature Extraction (Hierarchical CNN Encoding) S 6.2.2.1 Hierarchical CNN Local Feature Extraction The same standardized feature map After P×P patching, local details are extracted layer by layer through convolutional blocks consisting of "convolutional layer + batch normalization layer (BN) + ReLU activation function", as shown in the following structure: Where W is the convolution kernel weight and b is the bias term. This indicates a convolution operation. Output shape: Finally, multi-scale local features L0, L1, L2, and L3 are generated, aligned with the upper branch level. Their resolution is consistent with the corresponding global features G0-G3, ensuring that the dimensions of subsequent spatial feature fusion are fully adapted (e.g., G1 corresponds to L1, and CCNN is the number of convolution channels). S7 Feature Fusion Module Design (AFF Multi-Scale Spatial Feature Fusion + TSF Spatiotemporal Feature Fusion) The AFF module is connected in series to achieve spatial dual-branch feature fusion, and the TSF module is used to achieve spatiotemporal feature fusion. S7.1 Spatial Upper and Lower Branch Feature Fusion (AFF Module) (e.g.) Figure 2 (As shown) S7.1.1, Core Logic like Figure 2As shown, the global spatial features G0, G1, G2, G3 output by the upper branch and the local spatial features L0, L1, L2, L3 output by the lower branch correspond one-to-one. AFF modules are connected in series to achieve spatial feature fusion between the upper and lower branches: each AFF module is connected sequentially. The fused feature output by the previous AFF module (e.g., H0) is input together with the next-level global spatial feature (e.g., G1) and local spatial feature (e.g., L1) to obtain the current-level fused feature (e.g., H1). Fusion features H0, H1, H2, and H3 are generated one by one according to the above method.

[0054] S7.1.2 Initial Feature Fusion (Static Addition) First, consider the global features at the same level. and local features Element-by-element addition forms the initial fusion feature of the hierarchy. The formula is as follows: S7.1.3 Multi-Scale Channel Attention Extraction (MS-CAM) (Dynamic Feature Extraction) For initial fusion features The design employs a dual-path structure: a global path (compressing spatial information and capturing global dependencies) and a local path (preserving spatial information and capturing local details). Channel features are extracted separately for each path, providing a feature foundation for subsequent AFF weight generation. Figure 3 ) S7.1.3.1, Global Path The lightweight design of "dimensionality reduction-activation-dimensionality enhancement" reduces computational complexity and redundant information interference, highlights key channel features, and enhances the effectiveness of channel attention.

[0055] S7.1.3.1.1, Firstly, regarding Global average pooling (GAP) is performed to compress the input features H×W×C into 1×1×C, where each channel feature is represented as the average value across the entire spatial region. i is the feature map row index, j is the feature map column index, and c is the feature channel index; for The space is height and width, and C is the number of channels.

[0056] S7.1.3.1.2 Channel-by-channel 1×1 convolution dimensionality reduction reduces the number of channels from C to C / r (where r is the dimensionality reduction scaling factor (e.g., r=2), as shown in the following formula: Where W1 is a convolution weight matrix of shape (C / r) × C.

[0057] S7.1.3.1.3. A ReLU activation function is placed in the middle of each channel to achieve non-linear mapping and enhance the model's ability to express features. The formula is as follows: BN represents batch normalization operation.

[0058] S7.1.3.1.4, Channel 1×1 convolution dimensionality upscaling: The number of channels is increased from C / r back to the original number of channels C, restoring the feature dimension and obtaining the global path output features. .

[0059] W2 is a convolution weight matrix of shape C × (C / r). S7.1.3.2, Local Path Without compressing spatial information, directly... Perform the operation of "1×1 convolution dimensionality reduction - BN normalization - ReLU activation - 1×1 convolution dimensionality increase" to obtain local path output features. .

[0060] For 1×1 dimensionless convolution weights (C→C / r); The weights are 1×1 up-dimensional convolution weights (C / r→C).

[0061] S7.1.4, Weighted fusion of dual-path output features (e.g.) Figure 5 (As shown) Based on the Multi-Scale Channel Attention (MS-CAM) mechanism, the channel weights A are generated by fusing dual-path output features, thereby dynamically weighting the initial fused features, strengthening key features, suppressing redundant information, and realizing the dynamic fusion of global and local features.

[0062] S7.1.4.1, Generating Channel Attention Weights Output the global path Local path output After normalization, the values ​​are summed and activated by the Sigmoid function to generate channel attention weights (A) ranging from 0 to 1 (a larger value indicates that the channel feature is more important), as shown in the following formula: Where σ(·) is the Sigmoid activation function; BN represents the normalization of the data.

[0063] S7.1.4.2, Channel Weighting S7.1.4.3, Residual Connection By using residual connections, the original global features are... and local features Compared with the channel weighting feature of S7.1.4.2 The features are added together to ensure the integrity of the original feature information, thereby enhancing the training stability of the network. The final output is the fused feature of the current layer's AFF module. The formula is as follows: in, The fused features (such as H0, H1, H2, H3) output by the current-level AFF module, and the global features of the next layer. and local features Together they serve as input for the next AFF module.

[0064] S7.2 Spatiotemporal Feature Fusion (TSF Module) The final spatial fusion feature H3 output by S7.1 is deeply coupled with the global temporal feature S output by S6.1.5 to solve the problem of dimensional heterogeneity between spatial features (high-dimensional tensors) and temporal features (one-dimensional vectors), and outputs a unified representation Z that fuses spatiotemporal disaster information.

[0065] S7.2.1 Channel alignment and semantic dimension matching By using a fully connected layer, the global temporal feature S output by S6.1.5 is projected to the same channel dimension as the final spatial fusion feature H3 output by S7.1, thus eliminating semantic heterogeneity. in, The projection weight matrix is... For projection bias term; d represents the time series features aligned to the channel dimensions, and d represents the dimension of the global time series feature S.

[0066] S7.2.2, Temporal Feature Broadcasting, Spatial Dimension Expansion Time-series features aligned with channel dimensions Broadcasting and copying in the spatial dimension covers all spatial locations of H3, achieving pixel-level spatiotemporal feature matching, as shown in the following formula: Among them, Broadcast ( This refers to a space broadcast operation; the timing characteristics after broadcasting. Example: If the spatial resolution of H3 is 10×10×C, then S' will be copied 100 times (10×10).

[0067] S7.2.3 Initial Fusion of Spatiotemporal Features The timing characteristics after broadcasting The composite feature is obtained by adding it to the spatial fusion feature H3 to achieve preliminary fusion. : S7.2.4, Gated Weighted Fusion, Dynamic Feature Filtering S7.2.4.1, Gating Weight Calculation in, Time-series feature-specific learnable channel weight matrix (applied to) (to achieve deep semantic mapping of temporal features) Spatial feature-specific learnable channel weight matrix (applied to) (to achieve deep semantic mapping of spatial features) : Bias term; The Sigmoid activation function outputs the same signal as... The gating weight G (value 0~1, pixel-level dynamic selection of key disaster-causing features) is perfectly matched in dimension.

[0068] S7.2.4.2, Weighted Fusion Gating weights With initial fusion features Element-wise dot product enhances key spatiotemporal features and suppresses redundant information: Element-wise dot product operation S7.2.5, Residual Connection Gated weighted features With original spatial features and time series characteristics By adding them together, the original information is preserved and the fusion effect is enhanced, resulting in the final spatiotemporal integrated features. .

[0069] S8. Landslide Susceptibility Assessment: Probability Calculation, Result Mapping, Grading and Verification Analysis S8.1. The spatiotemporal fusion DL-LSA model trained in S6 and S7 is used to evaluate each grid cell in the study area and update the probability value of landslide occurrence for each cell. S8.2 Based on the probabilities obtained in S8.1, perform LSA result mapping, grading, and verification analysis. S8.2.1. Correspond the landslide occurrence probability values ​​to the geographical location of the study area to generate a landslide susceptibility probability distribution map.

[0070] S8.2.2, using the natural breakpoint method, the resulting map is divided into five susceptibility levels: extremely low, low, medium, high, and extremely high, to obtain the final zoning map; S8.2.3. Organize the final zoning map data into a table, using the high-quality sample labels output in S5. The validity of the LSA is verified based on the area ratio, landslide point ratio, and landslide ratio of the final zoning map. Once the verification is successful, it becomes the final LSA result.

Claims

1. A landslide susceptibility assessment method based on deep learning-driven spatiotemporal multi-scale feature fusion using both data and physical factors, characterized in that... Includes the following steps: Step S1: Obtaining the impact factor X Five major categories of factors contributing to landslide disasters were identified, and the process proceeded to step S2. Step S2: Construction of the initial sample set XY pairs The original positive sample Y=1 data is taken from the landslide points in the field survey / historical records of the study area, and the positive sample is expanded based on SMOTE; negative samples Y=0 are generated by randomly selecting points in the buffer, and the positive and negative samples are merged to form the initial sample set XY pair, and then proceed to step S3; Step S3: Preprocessing of Impact Factor X The correlation was determined by the Pearson correlation coefficient and the VIF variance inflation factor. Redundant influencing factors were eliminated until |rij| of all influencing factors in the initial sample set of step S2 was less than 0.8 and VIF was less than 10. Step S4: Influence factor coding generates standardized feature X' The two types of influencing factors, discrete and continuous, are coded respectively to generate standardized features X', and the rationality is verified by the frequency ratio method to form a sample set X'-Y pair; Step S5: Provide a physical constraint optimization sample set X'-Y based on the P-LSA model. Key mechanical parameters are associated with feature X'. Fs is solved by a modified formula. Prior constraints are provided by the P-LSA model. Susceptibility levels are divided according to the magnitude of Fs value. Positive / negative samples that fall within the extremely high / low susceptibility range in step S4 are retained to form a high-quality X'-Y sample set. Step S6: Construction of the Spatiotemporal Fusion DL-LSA Model Based on the standardized feature X' in step S4, an LSTM network is used to capture temporal features and generate a global temporal feature vector S. At the same time, the upper and lower branches of "Swin-Transformer+CNN" are used to extract global Gl-local spatial features Ll, ​​and multiple rounds of learning iterations are performed. Step S7: Feature Fusion Module Design On the output side of the spatiotemporal fusion DL-LSA model, feature fusion is further completed by sequentially connecting the AFF module and the TSF module. The AFF module realizes the fusion of multi-scale spatial features Gl and Ll to form spatial fusion feature Hl. The TSF module realizes spatiotemporal feature fusion. Step S8: Output the landslide susceptibility assessment results Output the model's predicted probability, plot the evaluation results, classify the monitored area into different levels, and verify the results based on the sample set labels from step S5.

2. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S1, the impact factor X acquisition focuses on the disaster-causing mechanism of landslides in the southeast coastal area, and collects 16 LSA core impact factor (X) data in 5 categories.

3. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S2, an initial sample set XY is constructed with "X value + Y label" as the core to build a complete initial sample set and solve the problem of sample scarcity: S2.1, positive sample Y=1 collection and expansion. S2.1.1 Using the landslide points from the field survey / historical records in the study area as the original positive sample, extract the values ​​of the corresponding 16 influencing factors X to form the original XY pairs. S2.1.

2. Use SMOTE oversampling to generate 29N positive samples. S2.1.2.1 For each positive sample (i=1,2,…,N), use Euclidean distance in the feature space to find its k nearest positive sample neighbors. The formula for Euclidean distance is: For feature dimension, , For the sample , exist The values ​​that can be taken on each feature; S2.1.2.2, Randomly select a sample from the k nearest neighbors. New samples are synthesized through linear interpolation: in, Original positive samples; : Random nearest neighbor sample; δ is a random number between 0 and 1; S2.1.2.3 Repeat the above process until 29N synthetic samples are generated; S2.2 Selection of negative sample Y=0 S2.2.

1. Divide the area into an 800-meter buffer zone centered on the geographical location of the expanded positive sample; S2.2.2 In the area outside the buffer zone, randomly select points at a rate of 5 times the number of positive samples and mark them as negative samples with the label Y=0; S2.3 Initial Sample Set Generation: Merge positive and negative samples to form an initial XY pair sample set of "influence factor value X + sample label Y", providing basic data for subsequent S3 preprocessing.

4. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S3, the impact factor X from step S1 is preprocessed to screen out the impact factors that are actually suitable for entering step S4; the specific steps are as follows: S3.1, Pearson correlation coefficient screening S3.1.1 Calculation of Pearson Correlation Coefficient Matrix Calculate any number of impact factors in the impact factor dataset X. and Pearson correlation coefficient between The formula is: in: : No. Influence factors in each sample Values; Impact Factor The mean; Total number of samples; S3.1.2 Correlation Judgment If | | ≥0.8 indicates a strong linear correlation between the two influencing factors, and factors with ambiguous physical meanings should be eliminated; S3.2, VIF Collinearity Screening S3.2.1 Constructing a linear regression model Using impact factor data X each factor With as the dependent variable and all other factors as independent variables, construct a linear regression model: = + + +...+ + +...+ + in: Dependent variable; : Independent variable; Total number of impact factors; : Intercept term; : Regression coefficient; Error term; S3.2.2 Calculate the coefficient of determination R² in: Factors in the k-th sample The original observations; Calculated using a linear regression model The estimated value; S3.2.3 Calculate the VIF value for each factor The formula for calculating VIF is: in: The coefficient of determination for the linear regression model in S3.2.2 S3.2.4 Collinearity Judgment If VIF ≥ 10, it indicates that the factor has severe collinearity with other factors, and the influencing factor should be removed. S3.3, Iterative Check Repeat steps 3.1-3.2 until all impact factors are covered. | < 0.8 and VIF < 10.

5. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S4, the preprocessed set of influence factors X from step S3 is divided into two categories—discrete influence factors and continuous influence factors—based on their spatial distribution characteristics for encoding, generating standardized features X'. The frequency ratio method is then used to verify the rationality of the encoding. S4.1 Discrete Influence Factor: Based on the inherent classification system coding, the larger the coding value, the stronger the landslide promoting effect. S4.2 Continuous impact factors are classified using the natural breakpoint method; the larger the coding value, the stronger the promoting effect. S4.2.1 Data Sorting: Sort all sample values ​​of continuous impact factors in ascending order. S4.2.2, Breakpoint Calculation: The interval is automatically divided using the statistical method of "minimizing within-class variance and maximizing between-class variance"; S4.2.3 Standardized Coding: Assigning values ​​to the divided intervals; the larger the value, the stronger the promotion effect. Among them, rainfall intensity I reflects the typical characteristics of short-term torrential rain during the typhoon season in the southeast coastal areas. Calculation method: in The rainfall amount in mm during a certain period is [missing information]. The duration of rainfall is h; Effective cumulative rainfall over 7 days Reflecting the "cumulative effect" (excluding ineffective losses such as surface runoff and evaporation), the calculation method is as follows: in, (Daily moisture content decay coefficient). Let the rainfall on day i be mm; TWI calculation method: TWI = ln(As / tanβ) Where As: catchment area per unit contour line length (m² / m), reflecting upstream water collection capacity; tanβ: tangent of slope, reflecting potential water flow velocity; ln: natural logarithm function; SPI calculation method: SPI = As×tanβ Where As: catchment area per unit contour line length (m² / m), i.e., cumulative flow value × grid resolution; tanβ: tangent of slope; NDVI formula: Wherein, NIR: near-infrared reflectance; R: red reflectance. To enhance its validity, the frequency ratio (FR) method can be used for verification. Among them, The number of landslide grid cells in the j-th class of the i-th influence factor. This represents the total number of landslide grids in the study area. Let S be the total number of raster cells in the j-th category of the i-th impact factor, and S be the total number of raster cells in the study area. The frequency ratio of the i-th influence factor in the j-th category; like >1: This interval is positively correlated with landslides; If 0.5 ≤ ≤ 1: The correlation is weak, but there is no need to merge them; like <0.5: This interval has almost no correlation with landslides and needs to be merged with adjacent intervals.

6. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S5, a P-LSA model is constructed to perform LSA, and reasonable positive / negative samples are selected from regions with extremely high / low susceptibility to form a high-quality sample set X'-Y sample pairs. S5.1 Correlation between P-LSA parameters and characteristic X': 6 key mechanical parameters c' , h , All of these are derived from the quantization assignment of feature X'. S5.2 Correction formula for P-LSA model (based on weighted influence factor fusion comprehensive judgment value) in: Land use correction coefficient (-0.2 for construction land, +0.15 for forest land); Vegetation cover correction factor; Rainfall intensity correction factor; S5.4 Sample Screening Preserve areas with extremely high susceptibility Positive samples Y=1 within the range, and extremely low regions Negative samples Y=0 within the range form a high-quality X'-Y sample set.

7. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S6, the spatiotemporal fusion DL-LSA model constructs a "temporal branch + spatial dual-path branch" architecture, builds a feature map based on the standardized feature X' generated in S4, and autonomously extracts deep spatiotemporal features to achieve dynamic dependency capture and full coverage of multi-scale spatial features. S6.1, Timing Branch; S6.1.1, Time Series Data Structuring The dynamic influencing factor in S4 is: 7-day effective precipitation P. e The standardized feature X' corresponding to rainfall intensity I is converted into a structured feature vector at time t: in: Standardized characteristics of the 7-day effective precipitation at time t; Standardized characteristics of rainfall intensity at time t; For feature dimension ( ); S6.1.2 Data Normalization For the original impact factor Min-max normalization is performed to eliminate dimensional differences and adapt to LSTM input requirements: in: : Standardized feature global minimum / maximum values ​​corresponding to 7-day effective precipitation and rainfall intensity in S4; S6.1.3, Constructing sequences using the sliding window method: S6.1.3.

1. Set the sliding window size W = 7 days (to adapt to the cumulative disaster effect of rainfall), the sliding step is 1 day, and slice the continuous time series features into input sequences of fixed length (7). Each sequence corresponds to one evaluation day, and the window contains feature data for 7 consecutive days: in The input sequence matrix of the LSTM is... (k=0,1,...,6) represents the time series feature vector after normalization for the previous k days, used to evaluate the current day; To evaluate the normalized time-series feature vector for the day, m=2 is the feature dimension; S6.1.3.2, Map the daily feature vector within the window to the target dimension using a fully connected layer: in: This is the mapping weight matrix; It is the bias vector; It is the ReLU activation function; d represents the mapped daily feature vector; d is the mapped feature dimension. S6.1.3.3 Constructing the input matrix of LSTM: in: The feature vectors mapped over 7 days within the sliding window are arranged in chronological order and directly input into the LSTM for time series modeling. This is the single-day feature vector after mapping for that day; S6.1.4, LSTM Temporal Feature Modeling LSTM dynamically updates cell states through a gating mechanism, capturing the temporal dependencies of daily features within a window, as shown in the following formula: Forgotten Gate: Input Gate: Candidate cell status: Cell status update: Output gate: Hidden state: in: : Daily LSTM hidden state; Features entered on the same day; : LSTM cell status; Gating weight matrix; : Gating bias vector; : Hadama accumulation; Sigmoid activation function Hyperbolic tangent function; Output: After modeling with 7 days of data within a sliding window, the LSTM layer outputs the complete set of hidden states for the entire sequence: in: : A W×d hidden state matrix for the entire sequence, containing the hidden states for 7 days within the window; All dimensions are d; S6.1.5 Attention Mechanism: Global Temporal Feature Aggregation By using attention-weighted aggregation, key date features within the window are highlighted to generate global time-series features: S6.1.5.1 Calculate daily hidden status Attention score: in, , , Here, k represents the attention parameter, and k represents the hidden layer dimension. S6.1.5.2, Convert the scores into normalized weights using the Softmax function: S6.1.5.3, Weighted summation to generate global temporal feature vectors: S6.2, Spatial Branching Using the multi-channel standardized feature map X' generated by S4 as input, a dual-path structure of "upper branch global spatial feature extraction + lower branch local spatial feature extraction" is constructed, which respectively correspond to the clustering pattern and crack and deformation mechanism in the southeast coastal area; S 6.2.1, Upper Branch: Global Spatial Feature Extraction S 6.2.1.1, Patch Segmentation and Embedding Multi-channel standardized feature maps For input, divide into blocks without repetition. The patch is flattened into a vector sequence by sliding window blocks and arranged in spatial order. =Unfold(X’, kernel_size=P, stride=P) The Unfold output shape is as follows: kernel_size is the kernel size, stride is the stride. Then, it is embedded into a high-dimensional space through a linear mapping layer: The Linear layer adjusts the Patch vector dimension to the Transformer hidden dimension C, and outputs... ; S 6.2.1.2, Hierarchical Swing Transformer Block Global Feature Extraction By combining multiple levels of Patch Merging downsampling (halving the resolution and doubling the number of channels at each level), global features G0, G1, G2, and G3 are gradually generated: in: For the first Level input features; : No. Global features at the level; S 6.2.2, Lower Branch: Local Spatial Feature Extraction S 6.2.2.1 Hierarchical CNN Local Feature Extraction The same standardized feature map After P×P patching, local details are extracted layer by layer through convolutional blocks consisting of "convolutional layers + batch normalization (BN) layers + ReLU activation function", as shown in the following structure: Where W is the convolution kernel weight and b is the bias term. Indicates the convolution operation; Output shape: Finally, multi-scale local features L0, L1, L2, and L3, aligned with the upper branch level, are generated, with resolutions consistent with the corresponding global features G0-G3, ensuring that the dimensions of subsequent spatial feature fusion are fully adapted.

8. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S7, the feature fusion module is designed. The AFF module is connected in series to achieve spatial dual-branch feature fusion, and the TSF module is used to achieve spatiotemporal feature fusion. S7.1 Spatial Upper and Lower Branch Feature Fusion AFF Module S7.1.1, Core Logic The global spatial features G0, G1, G2, G3 output by the upper branch and the local spatial features L0, L1, L2, L3 output by the lower branch correspond one-to-one. The AFF modules are connected in series to achieve the fusion of spatial features between the upper and lower branches: the AFF modules are connected sequentially, and the fused features output by the previous AFF module are input together with the global and local spatial features of the next level to obtain the fused features of this level; the fused features H0, H1, H2, H3 are generated one by one according to the above method. S7.1.2 Initial Feature Fusion First, consider the global features at the same level. and local features Element-by-element addition forms the initial fusion feature of the hierarchy. The formula is as follows: S7.1.3 Multi-scale Channel Attention Extraction Channel Feature MS-CAM For initial fusion features The design incorporates a dual-path structure with global and local paths to extract channel features, providing a feature foundation for subsequent AFF weight generation. S7.1.3.1, Global Path The lightweight design of "dimensionality reduction-activation-dimensionality enhancement" reduces computational complexity and redundant information interference, highlights key channel features, and enhances the effectiveness of channel attention. S7.1.3.1.1, Firstly, regarding Global average pooling (GAP) is performed to compress the input features H×W×C into 1×1×C, with each channel feature represented as the average value across the entire spatial region. i is the feature map row index, j is the feature map column index, and c is the feature channel index; for The space is height and width, and C is the number of channels; S7.1.3.1.2 Channel-by-channel 1×1 convolution dimensionality reduction reduces the number of channels from C to C / r, as shown in the following formula: in, W1 is a convolution weight matrix of shape (C / r) × C; S7.1.3.1.

3. A ReLU activation function is placed in the middle of each channel to achieve non-linear mapping and enhance the model's ability to express features. The formula is as follows: BN represents batch normalization operation; S7.1.3.1.4, Channel 1×1 convolution dimensionality upscaling: The number of channels is increased from C / r back to the original number of channels C, restoring the feature dimension and obtaining the global path output features. ; W2 is a convolution weight matrix of shape C × (C / r). S7.1.3.2, Local Path Without compressing spatial information, directly... The operation of "1×1 convolution dimensionality reduction - BN normalization - ReLU activation - 1×1 convolution dimensionality increase" is performed to obtain local path output features. ; The weights for the 1×1 dimensionality-reduced convolution are C→C / r; The weights for the 1×1 up-dimensional convolution are C / r→C; S7.1.4, Dual-path output feature weighted fusion Based on the multi-scale channel attention MS-CAM mechanism, the dual-path output features are fused to generate channel weights A, thereby achieving dynamic weighting of the initial fused features, strengthening key features, suppressing redundant information, and realizing dynamic fusion of global and local features. S7.1.4.1, Generating Channel Attention Weights Output the global path Local path output After normalization, the weights are summed and activated by the Sigmoid function to generate channel attention weights A ranging from 0 to 1, as shown in the following formula: Where σ(·) is the Sigmoid activation function; BN represents the normalization of the data; S7.1.4.2, Channel Weighting S7.1.4.3, Residual Connection By using residual connections, the original global features are... and local features Compared with the channel weighting feature of S7.1.4.2 The features are added together to ensure the integrity of the original feature information, thereby enhancing the training stability of the network; finally, the fused features of the current layer's AFF module are output. The formula is as follows: in, The fused features output by the current-level AFF module, combined with the global features of the next layer. and local features Together they serve as input to the next AFF module; S7.2, Spatiotemporal Feature Fusion TSF Module The final spatial fusion feature H3 output by S7.1 is deeply coupled with the global temporal feature S output by S6.1.5 to solve the problem of dimensional heterogeneity of "spatial features are high-dimensional tensors and temporal features are one-dimensional vectors", and outputs a unified representation Z that fuses spatiotemporal disaster information. S7.2.1 Channel alignment and semantic dimension matching By using a fully connected layer, the global temporal feature S output by S6.1.5 is projected to the same channel dimension as the final spatial fusion feature H3 output by S7.1, thus eliminating semantic heterogeneity. in, The projection weight matrix is... For projection bias term; The time series features are aligned to the channel dimensions, and d is the dimension of the global time series feature S. S7.2.2, Temporal Feature Broadcasting, Spatial Dimension Expansion Time-series features aligned with channel dimensions Broadcasting and copying in the spatial dimension covers all spatial locations of H3, achieving pixel-level spatiotemporal feature matching, as shown in the following formula: Among them, Broadcast ( This refers to a space broadcast operation; the timing characteristics after broadcasting. Example: If the spatial resolution of H3 is 10×10×C, then S' will be copied 100 times; S7.2.3 Initial Fusion of Spatiotemporal Features The timing characteristics after broadcasting The composite feature is obtained by adding it to the spatial fusion feature H3 to achieve preliminary fusion. : S7.2.4, Gated Weighted Fusion, Dynamic Feature Filtering S7.2.4.1, Gating Weight Calculation in, A time-series feature-specific learnable channel weight matrix; Spatial feature-specific learnable channel weight matrix; : Bias term; The Sigmoid activation function outputs the same signal as... The gating weight G is dimensionally perfectly matched; S7.2.4.2, Weighted Fusion Gating weights With initial fusion features Element-wise dot product enhances key spatiotemporal features and suppresses redundant information: Element-wise dot product operation S7.2.5, Residual Connection Gated weighted features With original spatial features and time series characteristics By adding them together, the original information is preserved and the fusion effect is enhanced, resulting in the final spatiotemporal integrated features. ; 。 9. The landslide susceptibility assessment method based on deep learning for spatiotemporal multi-scale feature fusion driven by both data and physical factors, as described in claim 1, is characterized in that... In step S8, landslide susceptibility assessment includes probability calculation, result mapping, grade zoning, and verification analysis. S8.

1. The spatiotemporal fusion DL-LSA model trained in S6 and S7 is used to evaluate each grid cell in the study area and update the probability value of landslide occurrence for each cell. S8.2 Based on the probabilities obtained in S8.1, perform LSA result mapping, grading, and verification analysis. S8.2.

1. Correspond the landslide occurrence probability values ​​to the geographical location of the study area to generate a landslide susceptibility probability distribution map; S8.2.2, using the natural breakpoint method, the resulting map is divided into five susceptibility levels: extremely low, low, medium, high, and extremely high, to obtain the final zoning map; S8.2.

3. Organize the final zoning map data into a table, using the high-quality sample labels output in S5. The validity of the LSA is verified based on the area ratio, landslide point ratio, and landslide ratio of the final zoning map. Once the verification is successful, it becomes the final LSA result.

Images