A landslide disaster early warning method and system based on multi-source remote sensing data

Abstract

The application belongs to the field of geological disaster prevention, and discloses a landslide disaster early warning method and system based on multi-source remote sensing data, wherein the method comprises the following steps: obtaining original information about landslide disasters from multi-source data sources; performing resolution alignment on the original information to obtain a standardized data set; constructing a multi-source data fusion model based on the standardized data set; generating a dynamic distribution map of disaster risks by using the multi-source data fusion model; and completing early warning of landslide disasters by using the dynamic distribution map. The application realizes minute-level and meter-level precision landslide risk dynamic monitoring by coupling multi-source remote sensing data fusion and deep learning, significantly improves the warning timeliness and accuracy, reduces false positives and false negatives, and provides reliable decision support for disaster prevention and reduction.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prevention and control, specifically to an early warning method and system for landslide disasters based on multi-source remote sensing data. Background Technology

[0002] Landslides, as a serious geological hazard, pose a significant threat to human life, property, and the ecological environment, making their research crucial in the field of natural disaster prevention and control. Landslides often involve complex natural factors and dynamic changes, making timely and accurate early warnings key to minimizing losses. However, many current early warning methods have significant shortcomings in data integration and dynamic analysis, especially when facing complex terrain and variable environments, making it difficult to comprehensively capture early warning signals, thus limiting the accuracy and timeliness of early warnings.

[0003] Against this backdrop, the field of landslide disaster early warning faces significant technical challenges. The core issue lies in how to effectively integrate data from different sources, such as satellite imagery, drone monitoring, and information collected by ground equipment. These data vary greatly in format, spatiotemporal resolution, and coverage; without unified management and efficient integration, it is difficult to form a comprehensive assessment of disaster risk. At a deeper level, due to a lack of comprehensive analytical capabilities for multi-dimensional information such as geological features, rainfall changes, and slope conditions, early warning systems often fail to accurately determine the critical conditions for disaster occurrence, thus missing the optimal intervention window. These two problems are interconnected; insufficient data fusion directly limits the in-depth mining and application of multi-dimensional information.

[0004] Therefore, how to construct an early warning system that can integrate multi-source data and comprehensively analyze multi-dimensional information has become a key issue in early warning research on landslide disasters. Summary of the Invention

[0005] To address the above technical problems, this invention provides an early warning method for landslide disasters based on multi-source remote sensing data, the method comprising:

[0006] By obtaining raw information about landslide disasters from multiple data sources;

[0007] The original information is resolution aligned to obtain a standardized dataset;

[0008] Based on the standardized dataset, a multi-source data fusion model is constructed;

[0009] Using the multi-source data fusion model, a dynamic distribution map of disaster risk is generated;

[0010] The dynamic distribution map is used to complete the early warning of landslide disasters.

[0011] Preferably, the method for obtaining the raw information includes: acquiring raw information from satellite imagery, UAV monitoring, and ground equipment from multi-source data; performing data extraction operations to obtain a preliminary data set; classifying the data according to the characteristics of different formats based on the preliminary data set, dividing the data into image-related and equipment-related categories, and determining the classified data units; and using the classified data units, performing a format standardization conversion operation to map the image-related and equipment-related data to a pre-established structured framework to obtain standardized data units.

[0012] Preferably, the method for obtaining the standardized dataset includes:

[0013] The resolution of the original information is adjusted using spatial interpolation.

[0014] By setting threshold standards, the quality of the adjusted original information is verified to obtain comprehensive data.

[0015] Finally, the support vector machine algorithm is used to classify the comprehensive dataset, resulting in the standardized dataset.

[0016] Preferably, the method for performing the classification process includes: setting a set of multi-source remote sensing data samples, each sample x i Corresponding to a label y i For samples ∈{-1,+1}, find a hyperplane to separate samples with different labels:

[0017] w·x+b=0

[0018] Where w is the weight vector, which determines the direction of the hyperplane; x is the sample data point; and b is the bias term, which determines the position of the hyperplane.

[0019] To find the optimal hyperplane, we need to maximize the classification margin, that is:

[0020]

[0021] Meanwhile, to handle the case of nonlinear separability, a slack variable ζ is introduced. i And construct the following optimization problem:

[0022]

[0023] The constraints are satisfied:

[0024] y i (w·x i +b)≥1-ξ i

[0025] ξ i ≥0

[0026] Where C0 is the penalty parameter, used to balance the trade-off between classification margin and classification error; n is the number of samples.

[0027] Preferably, the method for constructing the multi-source data fusion model includes: adopting a coupled architecture of a three-dimensional convolutional neural network and physical constraints: firstly, using a 3-DCNN as the main body, end-to-end learning is performed on the resolution-aligned multi-source data; the network front end uses three consecutive spatial-spectral-temporal convolutional blocks, each block extracting local spatiotemporal features by a convolutional kernel, with kernel sizes of 3×3×3, 5×5×3, and 7×7×5, and strides decreasing sequentially to ensure that fine-grained information is not lost while expanding the receptive field; the convolutional feature maps are batch normalized and activated by PReLU to alleviate gradient vanishing and accelerate convergence; then, a cross-source attention gating module is introduced to perform weighted fusion of satellite, UAV, and ground features using learnable query vectors.

[0028] Preferably, the method for generating the dynamic distribution map includes:

[0029] Stack all probability tensors from the past 24 hours on the time axis to form a four-dimensional spacetime volume;

[0030] Input a lightweight gated recurrent unit network and output a continuous curve of the risk index changing over time for the next hour;

[0031] It is coupled with an external driving field to achieve real-time coupling of "external triggering strength - internal instability probability";

[0032] An adaptive kernel density estimation method is used to spatially smooth the exponential field and perform color mapping.

[0033] The present invention also provides an early warning system for landslide disasters based on multi-source remote sensing data. The system is used to implement the above method and includes: a data acquisition module, an alignment module, a construction module, a generation module, and an early warning module.

[0034] The acquisition module is used to obtain raw information about landslide disasters from multiple data sources;

[0035] The alignment module is used to perform resolution alignment on the original information to obtain a standardized dataset;

[0036] The building module is used to construct a multi-source data fusion model based on the standardized dataset;

[0037] The generation module is used to generate a dynamic distribution map of disaster risk using the multi-source data fusion model;

[0038] The early warning module is used to complete the early warning of landslide disasters using the dynamic distribution map.

[0039] Preferably, the workflow of the acquisition module includes: acquiring raw information from satellite imagery, UAV monitoring, and ground equipment from multi-source data; performing data extraction operations to obtain a preliminary data set; classifying the data according to the characteristics of different formats based on the preliminary data set, dividing the data into image-related and equipment-related categories, and determining the classified data units; and using the classified data units, performing a format standardization conversion operation to map the image-related and equipment-related data to a pre-established structured framework to obtain standardized data units.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention achieves dynamic monitoring of landslide risks with minute-level and meter-level accuracy through multi-source remote sensing data fusion and deep learning coupling, significantly improving the timeliness and accuracy of early warning, reducing false alarms and missed alarms, and providing reliable decision support for disaster prevention and mitigation. Attached Figure Description

[0042] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0045] Explanation of reference numerals in the attached figures:

[0046] 1010, Processor; 1020, Memory; 1030, Input / Output Interface; 1040, Communication Interface; 1050, Bus. Detailed Implementation

[0047] 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.

[0048] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] Example 1

[0051] As the background technology indicates, the core problem with existing technologies lies in how to effectively integrate data from different sources, such as satellite imagery, drone monitoring, and information collected by ground equipment. These data differ significantly in format, spatiotemporal resolution, and coverage; without unified management and efficient integration, it is difficult to form a comprehensive assessment of disaster risks.

[0052] This invention provides an early warning method for landslide disasters based on multi-source remote sensing data.

[0053] S1. Obtain raw information about landslide hazards from multiple data sources.

[0054] Raw information from satellite imagery, UAV monitoring, and ground equipment is obtained from multi-source data. Data extraction is performed to obtain preliminary data sets. Based on the characteristics of different formats, the data is classified into image-based and equipment-based categories, determining the classified data units. Using these classified data units, a format standardization conversion operation is performed to map the image-based and equipment-based data to a pre-established structured framework, obtaining standardized data units. If any fields are missing in a standardized data unit, they are filled in by comparing with preset field templates to obtain complete data units. For complete data units, a support vector machine algorithm is applied to evaluate data quality. If the evaluation result is below a preset threshold, the data is cleaned to obtain qualified data units. Based on the qualified data units, a data integration operation is performed to fuse the multi-source data according to a unified structure, determining the final integrated data set. The integrated data set generates a structured data storage format, which is then archived according to preset storage rules, completing the data processing workflow.

[0055] Specifically, the classification and processing stage divides the data into image and device categories. The image category includes satellite and UAV data, while the device category consists of ground sensor data. Within the classified data units, image data is identified by spatial resolution and timestamp, while device data is identified by device ID and acquisition frequency. In the format standardization conversion, image data is mapped to a structured framework, standardized to GeoTIFF format, including latitude and longitude ranges such as 116.3°E to 116.5°E; device data is converted to JSON format, including fields such as device ID A001 and timestamp 2023-10-01 08:00:00.

[0056] S2. Align the original information with resolution to obtain a standardized dataset.

[0057] Inconsistent resolution is a common problem when processing satellite imagery and UAV monitoring data. To address this, spatial interpolation is used to adjust the data. The principle of spatial interpolation is to estimate the pixel values ​​of unknown areas based on known pixel data, thereby achieving resolution uniformity. Assuming the satellite imagery resolution is 10 meters per pixel and the UAV data resolution is 2 meters per pixel, interpolation algorithms can be used to improve the resolution of the satellite imagery to a level close to that of the UAV data, generating a unified preliminary image dataset that facilitates subsequent data overlay and comparative analysis.

[0058] When detecting local missing or anomalies in image data, this can be determined by comparing and analyzing the features of the resolution-aligned data. For example, if the brightness value of the image data suddenly deviates from the normal range in a certain area, it can be inferred that there may be cloud cover or data acquisition errors. In this case, neighborhood interpolation technology can fill in the missing area based on the average value of surrounding pixels, ensuring the integrity of the image dataset. This process helps improve data continuity and lays the foundation for subsequent processing.

[0059] To verify the quality of image data, a threshold standard for brightness or sharpness is set. If the data in a certain area falls below this threshold, for example, if the sharpness score is below 80, secondary interpolation is required to improve the data quality. This operation ensures the reliability of the standardized image dataset and reduces analytical bias caused by quality issues.

[0060] Finally, the Support Vector Machine (SVM) algorithm is used to classify the comprehensive dataset. Since satellite imagery data emphasizes large-area coverage, while UAV data focuses more on detail, the algorithm can classify the features of both to determine a standardized dataset. Specific steps include:

[0061] Set a set of multi-source remote sensing data samples, each sample x i Corresponding to a label y i ∈{-1,+1} indicates whether the data quality is acceptable. The goal of SVM is to find a hyperplane that separates samples with different labels. This hyperplane is represented by the following formula:

[0062] w·x+b=0

[0063] Where w is the weight vector, which determines the direction of the hyperplane; x is the sample data point; and b is the bias term, which determines the position of the hyperplane.

[0064] To find the optimal hyperplane, we need to maximize the classification margin, that is:

[0065]

[0066] This can be achieved by minimizing This is achieved by... Meanwhile, to handle the case of nonlinear separability, a slack variable ξ is introduced. i And construct the following optimization problem:

[0067]

[0068] The constraints are satisfied:

[0069] y i (w·x i +b)≥1-ξ i

[0070] ξ i ≥0

[0071] Where C0 is the penalty parameter, used to balance the trade-off between classification margin and classification error; n is the number of samples. By solving the above optimization problem, the optimal weight vector w and bias term b can be obtained, thereby achieving classification evaluation of data quality.

[0072] S3. Construct a multi-source data fusion model based on a standardized dataset.

[0073] After completing the construction of the standardized dataset, the core task of this step is to deeply couple multi-source remote sensing features within a unified spatiotemporal framework. This involves preserving both the large-scale low-frequency information of satellite imagery and the high-frequency details of UAV and ground sensors, thereby obtaining a fusion model capable of characterizing the entire process of slope instability. To this end, this embodiment employs a coupled architecture of a "space-spectrum-temporal" three-dimensional convolutional neural network (3-DCNN) with physical constraints: first, the 3-DCNN serves as the main body, performing end-to-end learning on the resolution-aligned multi-source data. The network input is a five-dimensional tensor. Where H and W are spatial dimensions, T is the time series length, C is the number of spectral or physical parameter channels, and S is the data source identifier; the network front-end uses three consecutive spatial-spectral-temporal convolutional blocks, each consisting of k convolutional kernels. τ，σ，ρ Local spatiotemporal feature extraction is performed using convolutional kernel sizes of 3×3×3, 5×5×3, and 7×7×5, with progressively decreasing strides to ensure that fine-grained information is not lost while expanding the receptive field. The convolutional feature maps are then subjected to batch normalization and PReLU activation to mitigate gradient vanishing and accelerate convergence. Subsequently, a cross-source attention gating module is introduced, utilizing learnable query vectors. Weighted fusion of satellite, drone, and ground features is performed, with the attention weights as follows:

[0074]

[0075] in, It is the local safety factor predicted by the network; c′ is the effective cohesion; φ′ is the effective internal friction angle; γ, γ w These are the unit weights of soil and water, respectively; h, h w These represent the slope height and saturation thickness, respectively; θ represents the slope.

[0076] This regularization term forces the network to learn a representation that conforms to the mechanical mechanism by minimizing the difference between the predicted value and the calculated value of the physical model; at the same time, it introduces a spatiotemporal consistency loss to encourage the prediction results of adjacent spatiotemporal locations to remain smooth.

[0077]

[0078] in, Let β be the neighborhood set; β be the kernel width parameter; the overall optimization objective is:

[0079]

[0080] Mission loss Cross-entropy is used to measure the classification error between landslide and non-landslide pixels. The weight hyperparameters λ1 and λ2 are adaptively searched on the validation set through Bayesian optimization. The AdamW optimizer is used during the training phase with an initial learning rate of 10. -3 In conjunction with a cosine annealing strategy, the temperature is gradually reduced to 10 over 50 epochs. -5 The batch size was set to 16 to balance memory constraints and gradient stability. To prevent overfitting, a 0.3 Dropout layer was added at the end of the network, and online enhancements with random rotation, translation, and brightness perturbation were applied to the input tensor. After training, the high-dimensional feature map output by the network was reduced to a two-dimensional risk probability field through 1×1 convolution.

[0081]

[0082] Then, spatial post-processing is performed using a Conditional Random Field (CRF). The CRF energy function is:

[0083] E(z)=∑ i ψ u (z i )+∑ i＜j ψ p (z i ,z j )

[0084] Among them, the monistic force From the CNN output, the binary potential encourages consistent labeling in homogeneous regions through pixel location and spectral differences, and its expression is:

[0085]

[0086] Parameters μ, σ α σ β Determined through grid search.

[0087] The resulting multi-source data fusion model not only achieves optimal coupling of cross-source information in a high-dimensional feature space, but also outputs an interpretable landslide risk probability map under the constraints of physical mechanisms, providing high-precision input for the subsequent generation of dynamic distribution maps and early warning.

[0088] S4. Utilize a multi-source data fusion model to generate a dynamic distribution map of disaster risks.

[0089] Whenever a new satellite image is received by a ground station, a drone transmits high-resolution image clusters along a set route, or ground sensors push water content, pore water pressure, and tilt angle readings at a frequency of minutes, these data are first processed through preliminary steps, and then cut into five-dimensional tensors that match the input size of the multi-source data fusion model and directly sent to GPU memory.

[0090] At this point, the forward propagation of the 3D convolutional network is completely frozen, and backpropagation is no longer performed. It only outputs the posterior probability tensor of the landslide at the current moment with a millisecond delay. The value of this tensor is between 0 and 1, and its physical meaning is the instantaneous probability of slope instability corresponding to that pixel under the current combined meteorological, geological, and human disturbance conditions. However, a single inference can only provide a snapshot of "this moment". In order to generate a truly dynamic distribution map, the system stacks all probability tensors from the past 24 hours on the time axis to form a four-dimensional spatiotemporal volume, and immediately inputs it into a lightweight gated recurrent unit (GRU) network. The GRU passes the memory between adjacent time steps through the hidden state. The weights of its update gate and reset gate have been fine-tuned during the training phase specifically for the landslide evolution characteristics, so that the model can remember the cumulative seepage effect brought by continuous rainfall and also forget the instantaneous disturbances caused by brief vibrations. Thus, it outputs a continuous curve of the risk index changing with time for the next hour at each spatial grid point.

[0091] This curve does not exist in isolation, but is immediately coupled with the external driving field—the 0-3 hour rainfall forecast field extrapolated from radar echoes, the temperature gradient field interpolated from meteorological stations, and the real-time peak acceleration field from the seismic network are all uniformly downscaled to a spatial grid with the same spatial grid as the probability tensor. Through pixel-by-pixel Hadamard product and deep fusion with the risk evolution tensor, real-time coupling of "external triggering intensity - internal instability probability" is achieved. The coupled comprehensive risk index is then mapped to the 0-1 range by a single-layer perceptron; the closer the index is to 1, the higher the probability of a landslide. To make the risk index intuitively readable, the system uses adaptive kernel density estimation to spatially smooth the index field. Then, based on the index magnitude, the color is mapped to a blue-green-yellow-orange-red gradient band, with the band's transparency gradually decreasing over time, creating a visual effect similar to "ripples." Each heatmap frame is immediately overlaid on the 3D terrain tiles after generation and pushed to the browser via WebGL. Users can drag the time slider to rewind to any historical moment or observe the dynamic extrapolation of risk over the next hour.

[0092] To support high-concurrency access, the backend uses Redis to cache the most recent ten frames of heatmaps, and CDN nodes distribute them locally, ensuring that users across the country can pull the latest spatiotemporal distribution map within 1 second. Thus, step S4 completes the closed loop of the entire chain from the raw information flow to dynamic visualization. The output disaster risk dynamic distribution map not only has a minute-level refresh rate and meter-level spatial resolution, but also integrates physical and mechanical constraints and meteorological drivers, truly realizing the integration of "data-model-visualization". This provides a real-time digital foundation that can be queried, traced back and compared at any time for the early warning of landslide disasters in subsequent steps.

[0093] S5. Utilize dynamic distribution maps to complete early warning of landslide disasters.

[0094] Finally, through dynamic distribution maps, we can accurately determine the risk trends in various regions and complete early warnings of landslide disasters.

[0095] The technical solution of this invention achieves dynamic monitoring of landslide risk with minute-level and meter-level accuracy through multi-source remote sensing data fusion and deep learning coupling, which significantly improves the timeliness and accuracy of early warning, reduces false alarms and missed alarms, and provides reliable decision support for disaster prevention and mitigation.

[0096] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method.

[0097] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, it should be understood that the sequence number of each step in the above embodiments does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention. The actions or steps recorded in the claims can be performed in a different order than that in the above embodiments and can still achieve the desired result. In addition, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0098] Example 2

[0099] Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides an early warning system for landslide disasters based on multi-source remote sensing data, including: an acquisition module, an alignment module, a construction module, a generation module, and an early warning module; the acquisition module is used to obtain raw information about landslide disasters from multi-source data sources; the alignment module is used to perform resolution alignment on the raw information to obtain a standardized dataset; the construction module is used to construct a multi-source data fusion model based on the standardized dataset; the generation module is used to generate a dynamic distribution map of disaster risk using the multi-source data fusion model; and the early warning module is used to complete the early warning of landslide disasters using the dynamic distribution map.

[0100] The present invention will now be described in detail, with reference to this embodiment, how it solves the technical problems mentioned above.

[0101] First, the data acquisition module is used to obtain raw information about landslide disasters from multiple data sources.

[0102] Raw information from satellite imagery, UAV monitoring, and ground equipment is obtained from multi-source data. Data extraction is performed to obtain preliminary data sets. Based on the characteristics of different formats, the data is classified into image-based and equipment-based categories, determining the classified data units. Using these classified data units, a format standardization conversion operation is performed to map the image-based and equipment-based data to a pre-established structured framework, obtaining standardized data units. If any fields are missing in a standardized data unit, they are filled in by comparing with preset field templates to obtain complete data units. For complete data units, a support vector machine algorithm is applied to evaluate data quality. If the evaluation result is below a preset threshold, the data is cleaned to obtain qualified data units. Based on the qualified data units, a data integration operation is performed to fuse the multi-source data according to a unified structure, determining the final integrated data set. The integrated data set generates a structured data storage format, which is then archived according to preset storage rules, completing the data processing workflow.

[0103] Specifically, the classification and processing stage divides the data into image and device categories. The image category includes satellite and UAV data, while the device category consists of ground sensor data. Within the classified data units, image data is identified by spatial resolution and timestamp, while device data is identified by device ID and acquisition frequency. In the format standardization conversion, image data is mapped to a structured framework, standardized to GeoTIFF format, including latitude and longitude ranges such as 116.3°E to 116.5°E; device data is converted to JSON format, including fields such as device ID A001 and timestamp 2023-10-01 08:00:00.

[0104] The alignment module then performs resolution alignment on the original information to obtain a standardized dataset.

[0105] Inconsistent resolution is a common problem when processing satellite imagery and UAV monitoring data. To address this, spatial interpolation is used to adjust the data. The principle of spatial interpolation is to estimate the pixel values ​​of unknown areas based on known pixel data, thereby achieving resolution uniformity. Assuming the satellite imagery resolution is 10 meters per pixel and the UAV data resolution is 2 meters per pixel, interpolation algorithms can be used to improve the resolution of the satellite imagery to a level close to that of the UAV data, generating a unified preliminary image dataset that facilitates subsequent data overlay and comparative analysis.

[0106] When detecting local missing or anomalies in image data, this can be determined by comparing and analyzing the features of the resolution-aligned data. For example, if the brightness value of the image data suddenly deviates from the normal range in a certain area, it can be inferred that there may be cloud cover or data acquisition errors. In this case, neighborhood interpolation technology can fill in the missing area based on the average value of surrounding pixels, ensuring the integrity of the image dataset. This process helps improve data continuity and lays the foundation for subsequent processing.

[0107] To verify the quality of image data, a threshold standard for brightness or sharpness is set. If the data in a certain area falls below this threshold, for example, if the sharpness score is below 80, secondary interpolation is required to improve the data quality. This operation ensures the reliability of the standardized image dataset and reduces analytical bias caused by quality issues.

[0108] Finally, the Support Vector Machine (SVM) algorithm is used to classify the comprehensive dataset. Since satellite imagery data emphasizes large-area coverage, while UAV data focuses more on detail, the algorithm can classify the features of both to determine a standardized dataset. Specific steps include:

[0109] Set a set of multi-source remote sensing data samples, each sample x i Corresponding to a label y i ∈{-1, +1} indicates whether the data quality is acceptable. The goal of SVM is to find a hyperplane that separates samples with different labels. This hyperplane is represented by the following formula:

[0110] w·x+b=0

[0111] Where w is the weight vector, which determines the direction of the hyperplane; x is the sample data point; and b is the bias term, which determines the position of the hyperplane.

[0112] To find the optimal hyperplane, we need to maximize the classification margin, that is:

[0113]

[0114] This can be achieved by minimizing This is achieved by... Meanwhile, to handle the case of nonlinear separability, a slack variable ξ is introduced. i And construct the following optimization problem:

[0115]

[0116] The constraints are satisfied:

[0117] y i (w·x i +b)≥1-ξ i

[0118] ξ i ≥0

[0119] Where C0 is the penalty parameter, used to balance the trade-off between classification margin and classification error; n is the number of samples. By solving the above optimization problem, the optimal weight vector w and bias term b can be obtained, thereby achieving classification evaluation of data quality.

[0120] A multi-source data fusion model is built using building blocks based on a standardized dataset.

[0121] After completing the construction of the standardized dataset, the core task of this step is to deeply couple multi-source remote sensing features within a unified spatiotemporal framework. This involves preserving both the large-scale low-frequency information of satellite imagery and the high-frequency details of UAV and ground sensors, thereby obtaining a fusion model capable of characterizing the entire process of slope instability. To this end, this embodiment employs a coupled architecture of a "space-spectrum-temporal" three-dimensional convolutional neural network (3-DCNN) with physical constraints: first, the 3-DCNN serves as the main body, performing end-to-end learning on the resolution-aligned multi-source data. The network input is a five-dimensional tensor. Where H and W are spatial dimensions, T is the time series length, C is the number of spectral or physical parameter channels, and S is the data source identifier; the network front-end uses three consecutive spatial-spectral-temporal convolutional blocks, each consisting of k convolutional kernels. τ，σ，ρ Local spatiotemporal feature extraction is performed using convolutional kernel sizes of 3×3×3, 5×5×3, and 7×7×5, with progressively decreasing strides to ensure that fine-grained information is not lost while expanding the receptive field. The convolutional feature maps are then subjected to batch normalization and PReLU activation to mitigate gradient vanishing and accelerate convergence. Subsequently, a cross-source attention gating module is introduced, utilizing a learnable query vector. Weighted fusion of satellite, drone, and ground features is performed, with the attention weights as follows:

[0122]

[0123] in, It is the local safety factor predicted by the network; c′ is the effective cohesion; φ′ is the effective internal friction angle; γ, γ w These are the unit weights of soil and water, respectively; h, h w These represent the slope height and saturation thickness, respectively; θ represents the slope.

[0124] This regularization term forces the network to learn a representation that conforms to the mechanical mechanism by minimizing the difference between the predicted value and the calculated value of the physical model; at the same time, it introduces a spatiotemporal consistency loss to encourage the prediction results of adjacent spatiotemporal locations to remain smooth.

[0125]

[0126] in, Let β be the neighborhood set; β be the kernel width parameter; the overall optimization objective is:

[0127]

[0128] Mission loss Cross-entropy is used to measure the classification error between landslide and non-landslide pixels. The weight hyperparameters λ1 and λ2 are adaptively searched on the validation set through Bayesian optimization. The AdamW optimizer is used during the training phase with an initial learning rate of 10. -3 In conjunction with a cosine annealing strategy, the temperature is gradually reduced to 10 over 50 epochs. -5 The batch size was set to 16 to balance memory constraints and gradient stability. To prevent overfitting, a 0.3 Dropout layer was added at the end of the network, and online enhancements with random rotation, translation, and brightness perturbation were applied to the input tensor. After training, the high-dimensional feature map output by the network was reduced to a two-dimensional risk probability field through 1×1 convolution.

[0129]

[0130] Then, spatial post-processing is performed using a Conditional Random Field (CRF). The CRF energy function is:

[0131] E(z)=∑ i ψ u (z i )+∑ i＜j ψ p (z i , z j )

[0132] Among them, the monistic force From the CNN output, the binary potential encourages consistent labeling in homogeneous regions through pixel location and spectral differences, and its expression is:

[0133]

[0134] Parameters μ, σ ασ β Determined through grid search.

[0135] The resulting multi-source data fusion model not only achieves optimal coupling of cross-source information in a high-dimensional feature space, but also outputs an interpretable landslide risk probability map under the constraints of physical mechanisms, providing high-precision input for the subsequent generation of dynamic distribution maps and early warning.

[0136] The generation module uses a multi-source data fusion model to generate a dynamic distribution map of disaster risks.

[0137] Whenever a new satellite image is received by a ground station, a drone transmits high-resolution image clusters along a set route, or ground sensors push water content, pore water pressure, and tilt angle readings at a frequency of minutes, these data are first processed through preliminary steps, and then cut into five-dimensional tensors that match the input size of the multi-source data fusion model and directly sent to GPU memory.

[0138] At this point, the forward propagation of the 3D convolutional network is completely frozen, and backpropagation is no longer performed. It only outputs the posterior probability tensor of the landslide at the current moment with a millisecond delay. The value of this tensor is between 0 and 1, and its physical meaning is the instantaneous probability of slope instability corresponding to that pixel under the current combined meteorological, geological, and human disturbance conditions. However, a single inference can only provide a snapshot of "this moment". In order to generate a truly dynamic distribution map, the system stacks all probability tensors from the past 24 hours on the time axis to form a four-dimensional spatiotemporal volume, and immediately inputs it into a lightweight gated recurrent unit (GRU) network. The GRU passes the memory between adjacent time steps through the hidden state. The weights of its update gate and reset gate have been fine-tuned during the training phase specifically for the landslide evolution characteristics, so that the model can remember the cumulative seepage effect brought by continuous rainfall and also forget the instantaneous disturbances caused by brief vibrations. Thus, it outputs a continuous curve of the risk index changing with time for the next hour at each spatial grid point.

[0139] This curve does not exist in isolation, but is immediately coupled with the external driving field—the 0-3 hour rainfall forecast field extrapolated from radar echoes, the temperature gradient field interpolated from meteorological stations, and the real-time peak acceleration field from the seismic network are all uniformly downscaled to a spatial grid with the same spatial grid as the probability tensor. Through pixel-by-pixel Hadamard product and deep fusion with the risk evolution tensor, real-time coupling of "external triggering intensity - internal instability probability" is achieved. The coupled comprehensive risk index is then mapped to the 0-1 range by a single-layer perceptron; the closer the index is to 1, the higher the probability of a landslide. To make the risk index intuitively readable, the system uses adaptive kernel density estimation to spatially smooth the index field. Then, based on the index magnitude, the color is mapped to a blue-green-yellow-orange-red gradient band, with the band's transparency gradually decreasing over time, creating a visual effect similar to "ripples." Each heatmap frame is immediately overlaid on the 3D terrain tiles after generation and pushed to the browser via WebGL. Users can drag the time slider to rewind to any historical moment or observe the dynamic extrapolation of risk over the next hour.

[0140] To support high-concurrency access, the backend uses Redis to cache the most recent ten frames of heatmaps, and CDN nodes distribute them locally, ensuring that users across the country can pull the latest spatiotemporal distribution map within 1 second. Thus, step S4 completes the closed loop of the entire chain from the raw information flow to dynamic visualization. The output disaster risk dynamic distribution map not only has a minute-level refresh rate and meter-level spatial resolution, but also integrates physical and mechanical constraints and meteorological drivers, truly realizing the integration of "data-model-visualization". This provides a real-time digital foundation that can be queried, traced back and compared at any time for the early warning of landslide disasters in subsequent steps.

[0141] Finally, the early warning module utilizes a dynamic distribution map to complete the early warning of landslide disasters.

[0142] Finally, through dynamic distribution maps, we can accurately determine the risk trends in various regions and complete early warnings of landslide disasters.

[0143] The system described in the above embodiments is used to implement the corresponding early warning method for landslide disasters based on multi-source remote sensing data in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0144] It should be noted that the aforementioned landslide disaster early warning system based on multi-source remote sensing data is embodied in the form of functional units. The term "module" here can be implemented in software and / or hardware, without specific limitations.

[0145] For example, a "module" can be a software program, hardware circuit, or a combination of both that implements the above functions. Hardware circuits may include application-specific integrated circuits (ASICs), electronic circuits, processors (e.g., shared processors, proprietary processors, or group processors) and memory for executing one or more software or firmware programs, combined logic circuits, and / or other suitable components that support the described functions.

[0146] Example 3

[0147] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the landslide disaster early warning method based on multi-source remote sensing data as described in any of the above embodiments.

[0148] Figure 2 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0149] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0150] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0151] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0152] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB (Universal Serial Bus), network cable, etc.) or wireless means (such as mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0153] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0154] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0155] The system described in the above embodiments is used to implement the corresponding early warning method for landslide disasters based on multi-source remote sensing data in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0156] Example 4

[0157] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the landslide disaster early warning method based on multi-source remote sensing data as described in any of the above embodiments.

[0158] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0159] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the landslide disaster early warning method based on multi-source remote sensing data as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0160] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0161] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuitry) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0162] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0163] Therefore, the units of the various examples described in the embodiments of this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0164] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for early warning of landslide disasters based on multi-source remote sensing data, characterized in that, The method includes: By obtaining raw information about landslide disasters from multiple data sources; The original information is resolution aligned to obtain a standardized dataset; Based on the standardized dataset, a multi-source data fusion model is constructed. The steps include: adopting a coupled architecture of a 3D convolutional neural network and physical constraints: first, using a 3DCNN as the main body, end-to-end learning is performed on the resolution-aligned multi-source data; the network front end uses three consecutive spatiotemporal convolutional blocks, each block extracting local spatiotemporal features through convolutional kernels, with kernel sizes of 3×3×3, 5×5×3, and 7×7×5, and strides decreasing sequentially to ensure that fine-grained information is not lost while expanding the receptive field; the convolutional feature maps are batch normalized and activated by PReLU to alleviate gradient vanishing and accelerate convergence; then, a cross-source attention gating module is introduced to perform weighted fusion of satellite, UAV, and ground features using learnable query vectors; Using the multi-source data fusion model, a dynamic distribution map of disaster risk is generated. The steps include: stacking all probability tensors from the past 24 hours on the time axis to form a four-dimensional spatiotemporal volume; inputting a lightweight gated recurrent unit network to output a continuous curve of the risk index changing over time in the next hour; coupling with an external driving field to achieve real-time coupling of external triggering intensity and internal instability probability; and using adaptive kernel density estimation to spatially smooth the index field and perform color mapping. The dynamic distribution map is used to complete the early warning of landslide disasters.

2. The landslide disaster early warning method based on multi-source remote sensing data according to claim 1, characterized in that, The method for obtaining the raw information includes: acquiring raw information from satellite imagery, UAV monitoring, and ground equipment from multi-source data, performing data extraction operations to obtain a preliminary data set; classifying the data according to the characteristics of different formats based on the preliminary data set, dividing the data into image-type and equipment-type data, and determining the classified data units; using the classified data units, performing a format standardization conversion operation to map the image-type and equipment-type data to a pre-established structured framework to obtain standardized data units.

3. The landslide disaster early warning method based on multi-source remote sensing data according to claim 1, characterized in that, Methods for obtaining the standardized dataset include: The resolution of the original information is adjusted using spatial interpolation. By setting threshold standards, the quality of the adjusted original information is verified to obtain comprehensive data. Finally, the support vector machine algorithm is used to classify the comprehensive dataset, resulting in the standardized dataset.

4. The landslide disaster early warning method based on multi-source remote sensing data according to claim 3, characterized in that, The method for performing the classification process includes: setting a set of multi-source remote sensing data samples, each sample Corresponding to a tag By finding a hyperplane, samples with different labels can be separated: in, The weight vector determines the direction of the hyperplane; ... To find the optimal hyperplane, we need to maximize the classification margin, that is: Meanwhile, to handle the case of nonlinear separability, slack variables are introduced. And construct the following optimization problem: The constraints are satisfied: in, C 0 is the penalty parameter, used to balance the trade-off between classification margin and classification error; n It refers to the number of samples.

5. A landslide disaster early warning system based on multi-source remote sensing data, the system being used to implement the method described in any one of claims 1-4, characterized in that, include: The module includes a data acquisition module, an alignment module, a construction module, a generation module, and an early warning module. The acquisition module is used to obtain raw information about landslide disasters from multiple data sources; The alignment module is used to perform resolution alignment on the original information to obtain a standardized dataset; The construction module is used to build a multi-source data fusion model based on the standardized dataset. The process includes: adopting a coupled architecture of a three-dimensional convolutional neural network and physical constraints: first, using a 3-DCNN as the main body, end-to-end learning is performed on the resolution-aligned multi-source data; the network front end uses three consecutive spatiotemporal convolutional blocks, each block extracting local spatiotemporal features by a convolutional kernel, with kernel sizes of 3×3×3, 5×5×3, and 7×7×5, and strides decreasing sequentially to ensure that fine-grained information is not lost while expanding the receptive field; the convolutional feature maps are batch normalized and activated by PReLU to alleviate gradient vanishing and accelerate convergence; then, a cross-source attention gating module is introduced to perform weighted fusion of satellite, UAV, and ground features using learnable query vectors; The generation module is used to generate a dynamic distribution map of disaster risk using the multi-source data fusion model. The process includes: stacking all probability tensors from the past 24 hours on the time axis to form a four-dimensional spatiotemporal volume; inputting a lightweight gated recurrent unit network and outputting a continuous curve of the risk index changing with time in the next hour; coupling with an external driving field to achieve real-time coupling of external triggering intensity and internal instability probability; using adaptive kernel density estimation to spatially smooth the index field and perform color mapping. The early warning module is used to complete the early warning of landslide disasters using the dynamic distribution map.

6. The landslide disaster early warning system based on multi-source remote sensing data according to claim 5, characterized in that, The workflow of the acquisition module includes: acquiring raw information from satellite imagery, UAV monitoring, and ground equipment from multi-source data, performing data extraction operations to obtain a preliminary data set; classifying the data according to the characteristics of different formats based on the preliminary data set, dividing the data into image-based and equipment-based categories, and determining the classified data units; using the classified data units, performing a format standardization conversion operation to map the image-based and equipment-based data to a pre-established structured framework to obtain standardized data units.

Images