A landslide susceptibility mapping method based on insar coherence rainfall component and deep learning

Abstract

The application discloses a landslide susceptibility mapping method based on InSAR coherence rainfall component and deep learning, belongs to the technical field of synthetic aperture radar interferometry (InSAR) data application and geological disaster risk assessment, and comprises the following steps: firstly, separating NDVI influence to extract a coherent rainfall signal; secondly, analyzing coherence-rainfall-landslide correlation to construct a feature set; and finally, taking LGM as a core, comparing MLP, SVM, RF, XGBoost and 1D-CNN models, verifying the feasibility of replacing rainfall data with the coherence of the rainfall component, and generating a high-precision susceptibility zoning map. The application not only provides a high-resolution hydrological proxy variable construction paradigm for areas without dense rainfall stations, but also provides a new path of remote sensing mechanism and model innovation collaborative driving for intelligent prediction of geological disasters. The application has important significance for landslide monitoring and early warning in high mountain and valley areas and other geological disaster-prone areas.

Description

Technical Field

[0001] This invention belongs to the field of synthetic aperture radar interferometry (InSAR) data application and geological disaster risk assessment technology, specifically involving a landslide susceptibility mapping method based on InSAR coherent rainfall components and deep learning. Background Technology

[0002] Landslides, as a typical geological disaster characterized by their sudden onset and high destructiveness, seriously threaten the safety of people's lives and property and the stability of infrastructure. Conducting landslide susceptibility assessments is a crucial foundation for disaster risk prevention and control and land spatial planning. In recent years, with the development of geographic information systems and machine learning technologies, landslide susceptibility mapping methods based on multi-factor statistical modeling have been widely applied. Among numerous disaster-causing factors, rainfall is one of the most critical triggering factors for landslides in high mountain and canyon areas. Due to the dramatic topographic relief and inherently weak slope stability in these areas, coupled with conditions such as fractured rock masses and uneven vegetation cover, heavy rainfall can directly induce landslides through soil infiltration, increased pore water pressure, and weakened shear strength. Simultaneously, the erosion and infiltration of surface runoff can damage the structural integrity of the slope, making regional landslide disasters highly likely during periods of concentrated heavy rainfall or monsoon seasons. Therefore, numerous studies have incorporated indicators such as cumulative rainfall, rainfall intensity, or effective rainfall duration into susceptibility models.

[0003] However, traditional methods relying on rainfall data from ground-based rain gauge stations have significant limitations: firstly, rain gauge stations are unevenly distributed, especially in sparsely populated mountainous areas with complex terrain, making it difficult to accurately characterize the spatial heterogeneity of local rainfall; secondly, while existing remote sensing precipitation products have some coverage capabilities, their spatiotemporal resolution is still insufficient to capture the instantaneous impact of short-duration heavy rainfall events on slope instability. Furthermore, ground-based monitoring facilities are easily damaged after extreme weather events or earthquakes, further exacerbating the problem of missing rainfall data. These limitations severely restrict the applicability and reliability of landslide susceptibility models in data-scarce areas.

[0004] Against this backdrop, exploring alternative indicators that do not rely on traditional rainfall characteristics has become a key direction for improving the universality of landslide risk assessment. Synthetic aperture radar interferometry, due to its advantages such as all-weather operation, high spatial resolution, and sensitivity to micro-deformations of the Earth's surface, is widely used in surface deformation monitoring and early identification of geological hazards. With the rapid development of computer technology and the increasing maturity of machine learning algorithms, the integration of InSAR technology with multidisciplinary methods has become a core innovative direction in landslide disaster research. Related research continues to advance around InSAR data acquisition, processing methods, and evaluation models: In 2022, some scholars first constructed an InSAR time-series deformation monitoring and wavelet time-frequency analysis framework, using long-term landslide displacement data to establish a high-precision deformation reference surface, making up for the problem of insufficient ground monitoring coverage; In 2024, in response to the problem of few effective measurement points and low data processing efficiency in high mountain and canyon areas, other scholars innovatively proposed a distributed scattering interferometry method based on principal component analysis time dimension compression. By developing a long-series SAR image compression module and integrating it into the DSI data processing workflow, the dependence on the original data is reduced while retaining the core deformation information. This achieves a dual breakthrough in acquiring more effective measurement points and efficient data processing in long-term surface deformation monitoring in high mountain and canyon areas. Its performance is significantly better than SBAS and traditional DSI methods, providing a better technical solution for long-term InSAR surface deformation detection. In 2025, some scholars combined the surface deformation rate retrieved from Sentinel-1 data with embedded machine learning and multi-temporal interferometric SAR coupling method to dynamically correct the landslide susceptibility map and improve the model's adaptability.

[0005] In recent years, in addition to deformation rate, InSAR derived parameters have also begun to be introduced into landslide susceptibility modeling. Among them, InSAR coherence, as a core parameter of interferogram quality, reflects the stability of the scattering characteristics of the same ground feature between two imaging sessions. Studies have shown that increased surface soil moisture content, vegetation disturbance, or small displacements can all lead to a decrease in coherence. In particular, after rainfall infiltration, changes in soil dielectric constant and loosening of particle structure significantly reduce the phase stability of radar signals, thus exhibiting low coherence characteristics. Therefore, time-series InSAR coherence can indirectly reflect changes in regional surface moisture status and stability, possessing the potential to serve as precursor information for landslides. Although some studies have attempted to use InSAR deformation rate or displacement time series for landslide susceptibility modeling, few studies have systematically explored the feasibility of using coherence as a substitute for rainfall. To this end, this paper proposes to introduce the average temporal coherence extracted from Sentinel-1 time series images into the landslide susceptibility evaluation system. By separating the coherence component mainly affected by rainfall to replace the traditional rainfall input, a deep learning model integrating multiple environmental factors is constructed. While current mainstream models such as random forests and extreme gradient boosting can handle multiple static factors and have strong resistance to overfitting, they are difficult to reflect the cumulative effects caused by multiple rainfalls or continuous surface disturbances before a landslide occurs. On the other hand, typical deep learning models such as convolutional neural networks and Transformers are good at capturing spatial patterns, but they usually do not make full use of the temporal correlation between multiple periods of InSAR data, making it difficult to reflect the process of surface stability evolving over time.

[0006] To address the aforementioned limitations, a Memory Gene Model (LGM) is proposed. This model achieves semantic compression of multi-source static inputs through "geological factor gene encoding" and introduces a temporal feature regulation mechanism to capture the comprehensive regulatory effect of Normalized Difference Vegetation Index (NDVI) and coherence time series on current landslide risk. Finally, the probability of landslide occurrence is output through the product fusion of geological potential and environmental regulatory factors, activated by a sigmoid function. Different feature combinations are used to quantitatively assess the predictive ability of decoupled coherence factors and their correlation and complementarity with rainfall factors. This method uses the rainfall-induced InSAR coherence dynamic response as a rainfall proxy variable to construct a fully remote sensing-driven landslide susceptibility assessment framework that does not require ground rainfall. Furthermore, the LGM model enables the collaborative modeling of static instability potential and dynamic triggering disturbances, improving predictive performance in rainless scenarios. Summary of the Invention

[0007] This invention aims to address the challenges of traditional landslide susceptibility assessments, which rely on low-resolution rainfall data, making it difficult to accurately characterize local hydrological disturbances in complex mountainous areas, and the difficulty of existing models in effectively integrating static geological factors with dynamic remote sensing time-series features.

[0008] To achieve the above-mentioned technical objectives and effects, the present invention provides the following technical solution:

[0009] A landslide susceptibility mapping method based on InSAR coherent rainfall components and deep learning, the method comprising the following steps:

[0010] S1. Data Preparation: Collect and organize multi-source data related to landslide research, including historical landslide data, topographic data, basic geographical and geological data, remote sensing data, and meteorological data, in order to construct a basic dataset for landslide susceptibility mapping;

[0011] S2. Coherence Extraction: The SAR image data is registered, interferometric pairs are screened, flattened and filtered. Based on the processed interferometric images, the InSAR coherence coefficient of the study area is calculated, a coherence distribution map with a spatial resolution of 30 meters is generated, a coherence separation model is constructed, the influence of NDVI on coherence is separated, and the coherence signal after removing vegetation interference is obtained.

[0012] S3. Multi-model construction and multi-dataset training: The memory gene model is selected as the core model, and multiple machine learning and deep learning models such as MLP, SVM, RF, XGBoost, and 1D-CNN are introduced to construct a coherence dataset containing rainfall components, a dataset containing raw coherence, and a dataset containing rainfall data. The machine learning and deep learning models are trained on the three types of datasets respectively to complete the construction of multiple landslide susceptibility assessment models.

[0013] S4. Model Performance Evaluation and Result Analysis: Output the evaluation results of each landslide susceptibility assessment model, using accuracy, recall, F1 score, and ROC curve as evaluation indicators to complete the quantitative evaluation of the performance of each model; generate a high-precision landslide susceptibility zoning map, compare the landslide susceptibility assessment effects of different datasets and different models, verify the feasibility and superiority of using rainfall component coherence to replace rainfall data for landslide susceptibility assessment, and finally form a complete high-precision landslide susceptibility assessment process.

[0014] Furthermore, the multi-source data in step S1 specifically includes:

[0015] Historical landslide data, and stratified sampling to obtain an equal number of non-landslide points;

[0016] Topographic data was obtained from DEM data on the geospatial data cloud platform, extracting slope, aspect, surface relief, profile curvature, and plane curvature.

[0017] Basic geographic and geological data, including lithological data from the National Geological Archives, road and water system data from OpenStreetMap, and fault zone data from the Seismic Fault Exploration Data Center;

[0018] Remote sensing data were used to extract NDVI from 30m spatial resolution Landsat-8 images from the U.S. Geological Survey. InSAR deformation rate and coherence data were calculated using Sentinel-1 satellite C-band images.

[0019] Meteorological data, using rainfall data from the National Tibetan Plateau Scientific Data Center, was used to construct a basic dataset for landslide susceptibility mapping.

[0020] Furthermore, step S2 specifically includes the following steps:

[0021] (1) Based on the assumption that “the spatiotemporal distribution of vegetation cover and rainfall are relatively independent, and their influence on coherence can be approximately linearly superimposed,” a component decomposition model of total coherence is established:

[0022]

[0023] in, It is coherence. It is a coherence component based on the impact of rainfall. It is a coherence component based on the influence of NDVI;

[0024] (2) Construct a model for the dominant coherence component of NDVI. Spatiotemporally match monthly Landsat-8 NDVI data with monthly average coherence data of InSAR during the same period to construct a synchronous dataset. Combine the land surface classification data to mask non-vegetation areas and perform linear fitting on annual NDVI and coherence scatter data. Extract the slope of the fitted line. (Characterizing the intensity of NDVI's influence on coherence) and intercept (Characterizing the effect of other environmental factors on coherence when NDVI is 0), the model is established by taking the mean of the slope and intercept respectively:

[0025]

[0026] in, NDVI value at the pixel scale;

[0027] (3) Extract the precipitation-dominant coherence component using the difference method. : The modeled NDVI influence components From the original observation coherence By subtracting from the middle, we can obtain the coherence variation component dominated by the rainfall process:

[0028]

[0029] Further, step S3 specifically involves:

[0030] (1) Construct a memory gene model, regard landslide risk as the product of "geological potential" and "environmental regulation", and construct a prediction framework that combines physical meaning and data-driven capability;

[0031] (2) Construct landslide potential with static characteristics and output the probability of instability within the geological environment;

[0032] (3) The coherence components of time-series NDVI and rainfall effects The external trigger strength is calculated using dynamic feature input.

[0033] (4) The probability of landslide occurrence for each pixel is obtained by compressing the product of the internal instability probability of the geological environment and the external trigger intensity regulation of dynamic characteristics using the Sigmoid function.

[0034] Furthermore, step S4 specifically includes the following steps:

[0035] (1) Output the evaluation results of each landslide susceptibility evaluation model, and use accuracy, recall, F1 value and ROC curve as core evaluation indicators to complete the quantitative analysis and ranking of the performance of each model.

[0036] (2) Based on the evaluation results of each model, a high-precision landslide susceptibility zoning map was drawn to clearly present the spatial distribution characteristics of different landslide susceptibility levels in the study area;

[0037] (3) Compare the landslide susceptibility assessment effects of datasets with different feature combinations and different models, and focus on verifying the feasibility and superiority of using rainfall component coherence to replace rainfall data for landslide susceptibility assessment.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This invention proposes to quantitatively separate high-resolution coherent components mainly affected by rainfall from InSAR coherent time-series signals, using them as rainfall proxy variables; and to construct a memory gene model (LGM), which, through the mechanisms of "geological gene encoding" and "temporal environmental regulation," collaboratively models the inherent instability potential of the region and the dynamic disturbances induced by rainfall, thereby achieving high-precision, high-resolution landslide susceptibility mapping without the need for ground rainfall data. Attached Figure Description

[0040] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0041] Figure 1This is a flowchart of the method in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the LGM model in an embodiment of the present invention;

[0043] Figure 3 This is a graph showing the number of extremely low coherence pixels and rainfall in an embodiment of the present invention;

[0044] Figure 4 This is a graph showing the number of low-coherence pixels and rainfall in an embodiment of the present invention;

[0045] Figure 5 This is a graph showing the number of pixels with moderate coherence and rainfall in an embodiment of the present invention;

[0046] Figure 6 This is a graph showing the number of highly coherent pixels and rainfall in an embodiment of the present invention;

[0047] Figure 7 This is a graph showing the number of highly coherent pixels and rainfall in an embodiment of the present invention;

[0048] Figure 8 This is a graph showing the coherence relationship between NDVI and InSAR from 2021 to 2023 in an embodiment of the present invention.

[0049] Figure 9 This is a graph showing the correlation analysis results of different datasets in an embodiment of the present invention;

[0050] Figure 10 This is a graph showing the statistical results of the accuracy metrics of each model in the embodiments of the present invention under different datasets;

[0051] Figure 11 These are ROC curves of different models under various datasets in the embodiments of this invention;

[0052] Figure 12 This is a landslide susceptibility mapping under different combinations of models and input features in embodiments of the present invention. Detailed Implementation

[0053] 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.

[0054] like Figure 1As shown, this embodiment provides a landslide susceptibility mapping method based on InSAR coherent rainfall components and deep learning. By using the rainfall-induced InSAR coherent dynamic response as a high-resolution rainfall proxy variable, a fully remote sensing-driven landslide susceptibility assessment framework is constructed that does not rely on ground rainfall observations. The method includes the following steps:

[0055] Step S1: Data preparation: Collect and organize multi-source data related to landslide research, including (1) historical landslide data, and obtain an equal number of non-landslide points by stratified sampling; (2) topographic data, using DEM data from the geospatial data cloud platform to extract slope, aspect, surface relief, profile curvature and plane curvature; (3) basic geographic and geological data, including lithological data from the National Geological Archives, road and water system data from OpenStreetMap, and fault zone data from the Earthquake Active Fault Exploration Data Center; (4) remote sensing data, selecting NDVI from 30m spatial resolution Landsat-8 images from the U.S. Geological Survey, and using 89 scenes of descending orbit and 142 scenes of ascending orbit Sentinel-1 satellite C-band images from January 2021 to December 2023 to calculate InSAR deformation rate and coherence data; (5) meteorological data, using rainfall data from the National Qinghai-Tibet Plateau Scientific Data Center; construct a basic dataset for landslide susceptibility mapping.

[0056] Step S2: Coherence Extraction: The Sentinel-1 satellite C-band SAR image data is registered, interferometric pair screening, flattening, and filtered. Based on the processed interferometric image, the InSAR coherence coefficient of the study area is calculated, and a coherence distribution map with a spatial resolution of 30 meters is generated. A coherence separation model is constructed to separate the influence of NDVI on coherence, and the coherence signal after removing vegetation interference is obtained, laying the foundation for subsequent hydrological proxy.

[0057] Step S3: Multi-model construction and multi-dataset training: LGM is selected as the core model, and multiple machine learning and deep learning models such as MLP, SVM, RF, XGBoost, and 1D-CNN are introduced. Three types of datasets are constructed: a coherence dataset containing rainfall components, a dataset containing raw coherence, and a dataset containing rainfall data. Based on the above three types of datasets, the machine learning and deep learning models are trained respectively to complete the construction of multiple landslide susceptibility assessment models.

[0058] Step S4: Model Performance Evaluation and Result Analysis: Output the evaluation results of each landslide susceptibility assessment model, using accuracy, recall, F1 score, and ROC curve as evaluation indicators to complete the quantitative evaluation of the performance of each model; generate a high-precision landslide susceptibility zoning map, compare the landslide susceptibility assessment effects of different datasets and different models, verify the feasibility and superiority of using rainfall component coherence to replace rainfall data for landslide susceptibility assessment, and finally form a complete high-precision landslide susceptibility assessment process.

[0059] Specifically, step S2 includes the following steps:

[0060] (1) InSAR coherence comprehensively reflects the stability of surface scattering characteristics, and its temporal series changes are regulated by multiple environmental factors such as NDVI and rainfall. Multiple scattering and volume scattering by vegetation canopy introduce additional phase perturbations, reducing coherence; rainfall, by changing soil moisture content, causes fluctuations in dielectric constant, perturbing the radar scattering mechanism and leading to a decrease in phase consistency. Since the original coherence is the result of the coupling of multiple factors, it is difficult to directly characterize the influence of a single hydrological process. Therefore, it is necessary to construct a quantifiable and separable coherence component model. Based on the assumption that "the spatiotemporal distribution of vegetation cover and rainfall in the study area are relatively independent, and their influence on coherence can be approximately linearly superimposed," a component decomposition model of total coherence is established:

[0061]

[0062] in, It is coherence. It is a coherence component based on the impact of rainfall. It is a coherence component based on the influence of NDVI.

[0063] (2) Construct an NDVI-dominated coherence component model. Using monthly Landsat-8 NDVI data from 2021 to 2023, and spatiotemporally matching the monthly average coherence data of InSAR during the same period, 36 sets of synchronized datasets were constructed. Water bodies, snow cover, and artificial building areas were masked using surface classification data, retaining only effective pixels in vegetation-covered areas to eliminate interference from non-vegetated regions. Linear fitting was performed on the annual NDVI and coherence scatter plot data, extracting the slope (characterizing the intensity of NDVI's influence on coherence) and intercept (characterizing the effect of other environmental factors on coherence when NDVI is 0). The average of the slope and intercept over the three years was then used to establish an NDVI-dominated coherence component model.

[0064]

[0065] in, This represents the coherence component driven solely by NDVI. The strength of the effect of NDVI on coherence. The NDVI value is at the pixel scale. The intercept represents the impact of other geographic features on coherence when NDVI is 0.

[0066] (3) Extraction of precipitation-dominant coherence components, in constructing NDVI-dominant coherence components Building upon this foundation, this study further extracts the rainfall-dominated coherence component using the difference method. Specifically, the modeled NDVI influence components will be... From the original observation coherence By subtracting from the middle, we can obtain the coherence variation component dominated by the rainfall process:

[0067]

[0068] Specifically, step S3 includes the following steps:

[0069] (1) A memory gene model (LGM) was constructed, which regards landslide risk as the product of "geological potential" and "environmental regulation", thus constructing a prediction framework that combines physical meaning and data-driven capabilities.

[0070] (2) Construct landslide potential with static characteristics and output the probability of instability within the geological environment.

[0071] (3) The coherence components of time-series NDVI and rainfall effects The external trigger strength is calculated using dynamic feature input.

[0072] (4) Risk fusion and probability output: Finally, the landslide probability of each pixel is obtained by compressing the product of the internal instability probability of the geological environment and the external trigger intensity regulation of dynamic characteristics through the Sigmoid function.

[0073] Specifically, step S4 includes the following steps:

[0074] (1) Output the evaluation results of each landslide susceptibility evaluation model, and use accuracy, recall, F1 value and ROC curve as core evaluation indicators to complete the quantitative analysis and ranking of the performance of each model.

[0075] (2) Based on the evaluation results of each model, a high-precision landslide susceptibility zoning map was drawn to clearly present the spatial distribution characteristics of different landslide susceptibility levels in the study area.

[0076] (3) Compare the landslide susceptibility assessment effects of datasets with different feature combinations and different models, and focus on verifying the feasibility and superiority of using rainfall component coherence to replace rainfall data for landslide susceptibility assessment.

[0077] The invention will now be further described with reference to the accompanying drawings.

[0078] Figure 2 The diagram illustrates the principle of the LGM model. As can be seen from the diagram, one branch encodes static geological factors as "instability genes," while the other branch integrates rainfall coherence components and NDVI temporal characteristics to form "environmental memory." The two are multiplied and then output as landslide probability via Sigmoid, reflecting the core idea of ​​collaborative modeling of innate potential and acquired triggers.

[0079] Figure 3 The map shows the number of extremely low coherence pixels and rainfall. It can be observed that the number of extremely low coherence pixels is significantly positively correlated with rainfall over time: the number of extremely low coherence pixels increases significantly during peak rainfall periods (e.g., 2021Q3, 2022Q3, 2023Q3), indicating that heavy rainfall leads to increased surface water content and changes in scattering characteristics, thus causing a decrease in coherence. This verifies that rainfall is one of the main factors affecting InSAR coherence.

[0080] Figure 4 This is a map showing the number of low-coherence pixels and rainfall. The map shows that the number of low-coherence pixels fluctuates with seasonal rainfall; the more abundant the rainfall, the greater the number of low-coherence pixels.

[0081] Figure 5 The study presents a map showing the number of moderately coherent pixels and rainfall. The results show a significant positive correlation between the number of moderately coherent pixels and rainfall in each quarter from 2021 to 2023. The number of pixels peaked in the seasons with high rainfall (e.g., Q3 2021, Q2 2022, Q3 2023) and troughed in the seasons with lower rainfall (e.g., Q3 2021, Q4 2022), generally fluctuating synchronously with rainfall.

[0082] Figure 6 The map shows the number of highly coherent pixels and rainfall in the study area. From 2021 to 2023, the number of highly coherent pixels and rainfall showed a negative correlation trend in each quarter. The number of pixels decreased significantly in quarters with peak rainfall (e.g., Q3 2021, Q3 2022, Q3 2023), while the number of pixels was relatively higher in quarters with less rainfall (e.g., Q1 2021, Q4 2022, Q1 2023), and overall decreased with increasing rainfall.

[0083] Figure 7 The map shows the number of highly coherent pixels and rainfall. From 2021 to 2023, the number of highly coherent pixels was negatively correlated with rainfall. The number of pixels decreased significantly during peak rainfall periods (such as Q3 of 2021, Q3 of 2022, and Q3 of 2023), while it was relatively higher during the driest seasons, and decreased overall as rainfall increased.

[0084] Figure 8The study presents the relationship between NDVI and InSAR coherence from 2021 to 2023. The results show a significant negative correlation between coherence and NDVI from 2021 to 2023, with fitting slopes of -0.5542, -0.4648, and -0.4442, respectively, and the intercept decreasing slightly year by year. Higher NDVI correlates with lower coherence, and the negative correlation strength weakens slightly with each year, exhibiting an overall stable negative linear trend.

[0085] Figure 9 This is a correlation analysis result chart of different datasets. The correlation coefficient heatmap of the three datasets shows that there are stable statistical associations between most environmental and terrain features, and the high consistency across different datasets indicates that the coherence after decoupling has less redundancy in the dataset combined with other environmental factors, providing better input features for subsequent modeling.

[0086] Figure 10 The graph shows the statistical results of the accuracy indicators of each model under different datasets. The rainfall impact coherence component dataset constructed in this study showed the best prediction performance among all the comparative models, fully verifying its innovation and effectiveness as a high-resolution hydrological proxy variable.

[0087] Figure 11 These are ROC curves for different models on various datasets. Under the optimal feature combination, the LGM model achieves an AUC of 0.97, and its Accuracy, Recall, and F1 scores are all superior to other models. Its robustness stems from the accurate modeling of nonlinear mechanisms and the ability to fuse multi-source features.

[0088] Figure 12 This dataset provides landslide susceptibility mapping under different model and input feature combinations. The experimental group using decoupled rainfall influence coherence components exhibits a more refined and continuous high-risk patch structure in the Lancang River main stream and its tributary valleys. This advantage stems from two core characteristics of dataset a:

[0089] (1) Inheriting the high spatial resolution of approximately 30m from InSAR data, it is able to capture local hydrological anomalies within river erosion zones;

[0090] (2) By subtracting the vegetation influence dominated by NDVI, non-rainfall signal interference was effectively suppressed, making the hydrological disturbance response purer.

[0091] In summary, the method of this invention not only provides a high-resolution hydrological proxy variable construction paradigm for areas without dense rain gauge stations, but also offers a new path for intelligent prediction of geological disasters driven by the synergistic approach of "remote sensing mechanism and model innovation".

[0092] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0093] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A landslide susceptibility mapping method based on InSAR coherent rainfall components and deep learning, characterized in that, The method includes the following steps: S1. Data Preparation: Collect and organize multi-source data related to landslide research, including historical landslide data, topographic data, basic geographical and geological data, remote sensing data, and meteorological data, in order to construct a basic dataset for landslide susceptibility mapping; S2. Coherence Extraction: The SAR image data is registered, interferometric pairs are screened, flattened and filtered. Based on the processed interferometric images, the InSAR coherence coefficient of the study area is calculated, a coherence distribution map with a spatial resolution of 30 meters is generated, a coherence separation model is constructed, the influence of NDVI on coherence is separated, and the coherence signal after removing vegetation interference is obtained. S3. Multi-model construction and multi-dataset training: The memory gene model is selected as the core model, and multiple machine learning and deep learning models such as MLP, SVM, RF, XGBoost, and 1D-CNN are introduced to construct a coherence dataset containing rainfall components, a dataset containing raw coherence, and a dataset containing rainfall data. The machine learning and deep learning models are trained on the three types of datasets respectively to complete the construction of multiple landslide susceptibility assessment models. S4. Model Performance Evaluation and Result Analysis: Output the evaluation results of each landslide susceptibility assessment model, using accuracy, recall, F1 score, and ROC curve as evaluation indicators to complete the quantitative evaluation of the performance of each model; generate a high-precision landslide susceptibility zoning map, compare the landslide susceptibility assessment effects of different datasets and different models, verify the feasibility and superiority of using rainfall component coherence to replace rainfall data for landslide susceptibility assessment, and finally form a complete high-precision landslide susceptibility assessment process.

2. The method according to claim 1, characterized in that, The multi-source data in step S1 specifically includes: Historical landslide data, and stratified sampling to obtain an equal number of non-landslide points; Topographic data was obtained from DEM data on the geospatial data cloud platform, extracting slope, aspect, surface relief, profile curvature, and plane curvature. Basic geographic and geological data, including lithological data from the National Geological Archives, road and water system data from OpenStreetMap, and fault zone data from the Seismic Fault Exploration Data Center; Remote sensing data were used to extract NDVI from 30m spatial resolution Landsat-8 images from the U.S. Geological Survey. InSAR deformation rate and coherence data were calculated using Sentinel-1 satellite C-band images. Meteorological data, using rainfall data from the National Tibetan Plateau Scientific Data Center, was used to construct a basic dataset for landslide susceptibility mapping.

3. The method according to claim 1, characterized in that, Step S2 specifically includes the following steps: (1) Based on the assumption that "the spatiotemporal distribution of vegetation cover and rainfall are relatively independent, and their influence on coherence can be approximately linearly superimposed", a component decomposition model of total coherence is established: in, It is coherence. It is a coherence component based on the impact of rainfall. It is a coherence component based on the influence of NDVI; (2) Construct a model for the dominant coherence component of NDVI. Spatiotemporally match monthly Landsat-8 NDVI data with monthly average coherence data of InSAR during the same period to construct a synchronous dataset. Combine the land surface classification data to mask non-vegetation areas and perform linear fitting on annual NDVI and coherence scatter data. Extract the slope of the fitted line. With intercept Regarding the slope and intercept Take the mean values ​​separately and build a model: in, NDVI value at the pixel scale; (3) Extract the precipitation-dominant coherence component using the difference method. : The modeled NDVI influence components From the original observation coherence By subtracting from the middle, we can obtain the coherence variation component dominated by the rainfall process:

4. The method according to claim 1, characterized in that, Step S3 specifically involves: (1) Construct a memory gene model, regard landslide risk as the product of "geological potential" and "environmental regulation", and construct a prediction framework that combines physical meaning and data-driven capability; (2) Construct landslide potential with static characteristics and output the probability of instability within the geological environment; (3) The coherence components of time-series NDVI and rainfall effects The external trigger strength is calculated using dynamic feature input. (4) The probability of landslide occurrence for each pixel is obtained by compressing the product of the internal instability probability of the geological environment and the external trigger intensity regulation of dynamic characteristics using the Sigmoid function.

5. The method according to claim 1, characterized in that, Step S4 specifically includes the following steps: (1) Output the evaluation results of each landslide susceptibility evaluation model, and use accuracy, recall, F1 value and ROC curve as core evaluation indicators to complete the quantitative analysis and ranking of the performance of each model. (2) Based on the evaluation results of each model, a high-precision landslide susceptibility zoning map was drawn to clearly present the spatial distribution characteristics of different landslide susceptibility levels in the study area; (3) Compare the landslide susceptibility assessment effects of datasets with different feature combinations and different models, and focus on verifying the feasibility and superiority of using rainfall component coherence to replace rainfall data for landslide susceptibility assessment.

Images