Geological disaster monitoring system and method based on remote sensing data

Abstract

The invention relates to the technical field of geological disaster monitoring, and particularly discloses a geological disaster monitoring system and method based on remote sensing data. The method comprises the following steps: acquiring remote sensing, SAR, LiDAR, Internet of Things sensing and humanistic geography multi-source data, and obtaining a fusion data set through a pixel-feature-decision three-level fusion architecture after edge calculation and cloud classification preprocessing; comprehensive features are extracted, a disaster type classifier self-adaptively dispatches a special prediction model of a model pool, historical data prediction is combined, and the model is dynamically fine-tuned through actually measured data; a special algorithm is adopted to adapt to a complex scene, and point location level early warning and linkage with secondary disaster prediction are realized based on differential thresholds. The system comprises a multi-source data processing module, a three-level fusion module, a self-adaptive prediction module and the like. The technical problems that in existing monitoring, multi-source data fusion depth is insufficient, AI model generalization and updating capacity is poor, complex scene adaptation is poor, early warning is extensive, and secondary disaster linkage is lacked are solved.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, specifically to a geological disaster monitoring system and method based on remote sensing data. Background Technology

[0002] Geological disasters (such as landslides, ground subsidence, and debris flows) are characterized by their suddenness and destructiveness. They not only directly threaten the safety of life and property, but also damage key infrastructure such as transportation and water conservancy, causing chain losses. Therefore, accurate and efficient dynamic monitoring and early warning of these disasters are of paramount importance.

[0003] Currently, relevant monitoring relies heavily on single remote sensing data acquisition and analysis, or on ground equipment deployment combined with simple data stitching. The former often employs optical remote sensing, single synthetic aperture radar (SAR), or basic lidar (LiDAR) technologies, while the latter supplements data through GNSS base stations and manual surveys. Although some solutions attempt to integrate multi-source data, they only remain at the surface overlay level.

[0004] However, in practical applications, these monitoring methods have obvious shortcomings: multi-source data lacks deep integration, making it impossible to fully explore the value of data correlation; prediction models are mostly general-purpose, making it difficult to adapt to the characteristics of different disaster types, and their accuracy is easily reduced due to data drift after long-term use, lacking a dynamic update mechanism; in complex scenarios such as high vegetation cover and steep terrain, data distortion or low coherence problems are prominent, resulting in a high rate of missed detection for small-scale disasters; early warnings are mostly regional qualitative prompts, lacking precise threshold settings at the point level, and do not consider the linkage effects of secondary disasters, making it difficult to support scientific prevention and control decisions.

[0005] To address the aforementioned issues and meet the needs of large-scale, high-precision, and all-scenario geological disaster monitoring, it is urgent to construct a monitoring solution that balances data fusion depth, model adaptability, scenario adaptability, and refined early warning. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention provides a geological disaster monitoring system and method based on remote sensing data, which improves monitoring accuracy, scientific rigor of early warning, and response efficiency by addressing issues such as shallow data fusion, weak scene adaptation, and crude early warning.

[0007] To achieve the above objectives, this invention proposes a geological disaster monitoring method based on remote sensing data, comprising: S1. Acquire multi-source data of the monitoring area, including remote sensing image data, SAR image data, LiDAR data, IoT sensor data and human geography data, and perform hierarchical preprocessing on the multi-source data through edge computing nodes; S2. Based on a three-level fusion architecture of pixel-level, feature-level, and decision-level, the preprocessed multi-source data is fused to obtain a fused dataset. S3. Perform feature extraction on the fused dataset to obtain a comprehensive feature set containing spectral features, texture features, elevation features and deformation features; S4. Call the preset model pool, which includes a landslide-specific prediction model, a ground subsidence-specific prediction model, and a debris flow-specific prediction model. Identify the disaster type in the monitoring area through a disaster type classifier, and adaptively schedule the corresponding special prediction model. Combine historical disaster data and comprehensive feature set to predict the impact range, disaster severity, and development trend of geological disasters. S5. Based on the error comparison between measured data and prediction results, trigger the model dynamic fine-tuning mechanism to incrementally update the currently scheduled dedicated prediction model; S6. The prediction results are adapted to complex scenarios by adopting the LiDARDEM-assisted SAR geometric correction algorithm and the combined algorithm of SAR super-resolution reconstruction and small target detection. S7. Construct a differentiated threshold system, and perform point-level quantitative evaluation on the prediction results after scenario adaptation processing based on the differentiated threshold system. When the evaluation result exceeds the corresponding threshold, trigger a disaster warning and link the secondary disaster prediction sub-model to generate a disaster chain risk assessment result.

[0008] Preferably, in S1, the IoT sensor data includes rainfall data, soil moisture data, and groundwater level data; the human geography data includes building height data, road grade data, and underground pipeline distribution data. The hierarchical preprocessing includes: vegetation removal from LiDAR data, denoising of SAR image data, and preliminary denoising of remote sensing image data through edge computing nodes; the remaining preprocessing operations are completed in the cloud. The cloud preprocessing includes radiometric calibration, atmospheric correction, and geometric correction of remote sensing image data; filtering, phase unwrapping, and orbital error removal of SAR image data; and ground point extraction and digital elevation model construction of LiDAR data.

[0009] Preferably, in S2, the fusion process of the three-level fusion architecture includes: Pixel-level fusion: Digital elevation models generated from LiDAR data are used to assist phase unwrapping of SAR image data and eliminate low-coherence interference in vegetated areas; Feature-level fusion: The attention mechanism is used to perform weighted fusion of the key features corresponding to each data. The key features include the spectral features of remote sensing image data, the deformation features of SAR image data, and the elevation features of LiDAR data. Decision-level fusion: The disaster type determination results output by the disaster type classifier are used to integrate decision information from various data dimensions and optimize the effectiveness of the fused dataset.

[0010] Preferably, in S4, the landslide-specific prediction model is an improved U-Net + attention mechanism model, the ground subsidence-specific prediction model is an LSTM + XGBoost fusion model, and the debris flow-specific prediction model is a CNN + hydrological model coupled model. The disaster type classifier extracts terrain slope features, rainfall correlation features, and land cover features from the comprehensive feature set to achieve automatic disaster type identification and model scheduling.

[0011] Preferably, in S5, the model dynamic fine-tuning mechanism includes: At each preset interval, GNSS measured displacement data and manual survey data of the monitoring area are collected, and the error is calculated with the prediction results of the corresponding period. When the calculated error value exceeds the preset error threshold, federated learning is adopted to fine-tune the current dedicated prediction model in small batches using incrementally collected measured data, without retraining the entire model.

[0012] Preferably, in S6, the LiDARDEM-assisted SAR geometric correction algorithm includes: generating a slope aspect map and distortion risk map of the monitoring area based on the LiDAR digital elevation model; using multi-track SAR data fusion to eliminate interference in overlapping areas; and filling data gaps in the shadow area by interpolating the deformation trends of adjacent non-shadow areas. The differentiated threshold system sets different warning thresholds according to the scene type of the monitoring area. The scene types include residential areas, high-speed rail lines, riverbanks, and mountainous areas.

[0013] A geological disaster monitoring system based on remote sensing data is also proposed, including: The multi-source data acquisition and preprocessing module is used to acquire remote sensing image data, SAR image data, LiDAR data, IoT sensor data and human geography data, and complete hierarchical preprocessing through edge computing nodes and cloud processing units; The three-level fusion module is used to fuse preprocessed multi-source data based on a three-level fusion architecture of pixel level, feature level, and decision level, and output a fused dataset. The feature extraction module is used to extract features from the fused dataset to generate a comprehensive feature set that includes spectral features, texture features, elevation features and deformation features; The adaptive prediction module includes a model pool, a disaster type classifier, and a prediction execution unit. The model pool stores prediction models for landslides, ground subsidence, and debris flows. The disaster type classifier is used to identify disaster types and schedule corresponding models. The prediction execution unit is used to combine historical data and a comprehensive feature set to output prediction results. The dynamic calibration module is used to collect measured data, calculate the error between the measured data and the prediction results, and trigger incremental fine-tuning of the model. The scene adaptation module includes a distortion correction subunit and a small target detection subunit. The distortion correction subunit adopts the LiDARDEM-assisted SAR geometric correction algorithm, and the small target detection subunit adopts the SAR super-resolution reconstruction and YOLOv8-tiny combined algorithm. The early warning linkage module is used to construct a differentiated threshold system, conduct point-level quantitative assessment, trigger disaster early warning, and link with the secondary disaster prediction sub-model to generate disaster chain risk assessment results.

[0014] Preferably, the multi-source data acquisition and preprocessing module includes: The remote sensing data subunit is used to receive periodic scan data from medium-resolution satellites, key monitoring data from high-resolution satellites, synthetic aperture radar data, and airborne LiDAR data from UAVs. The Internet of Things (IoT) sub-unit is used to connect to rainfall sensors, soil moisture sensors, and groundwater level sensors to collect corresponding IoT sensor data. The humanities data interface is used to connect to the city's GIS system to obtain data on building height, road grade, and underground pipeline distribution. The edge processing unit is used to remove vegetation from LiDAR data, denoise SAR data, and perform preliminary noise reduction on remote sensing images. The cloud processing unit is used to perform radiometric calibration, atmospheric correction, and geometric correction of remote sensing images; SAR data filtering, phase unwrapping, orbital error removal; and LiDAR data ground point extraction and digital elevation model construction.

[0015] Preferably, in the model pool of the adaptive prediction module, the landslide-specific prediction model is an improved U-Net + attention mechanism model with an embedded edge enhancement module; the ground subsidence-specific prediction model is a fusion model of LSTM time-series prediction model and XGBoost factor analysis model; the debris flow-specific prediction model is a coupled model of CNN landslide source area identification model and hydrological runoff calculation model; and the disaster type classifier realizes automatic disaster type determination and model scheduling by extracting topographic, rainfall and surface cover correlation features from the comprehensive feature set.

[0016] Preferably, the distortion correction subunit of the scene adaptation module generates a slope aspect map and a distortion risk map through a LiDAR digital elevation model, eliminates overlapping interference by fusing ascending and descending SAR data, and fills in the shadow area data by interpolating the deformation trend of adjacent areas. The secondary disaster prediction sub-model of the early warning linkage module includes a landslide dam prediction sub-model, which is used to estimate the rate of rise of the landslide dam water level and the risk of dam failure based on debris flow accumulation, river width and flow data.

[0017] Therefore, this invention proposes a geological disaster monitoring system and method based on remote sensing data, the beneficial effects of which are as follows: (1) By integrating multi-source data into a three-level deep fusion and designing an adaptive dedicated model pool, the value of data correlation is fully explored, and the dynamic fine-tuning mechanism avoids accuracy decay, thereby improving model adaptability and monitoring stability.

[0018] (2) Solve the problems of data distortion in complex scenarios and missed detection of small-scale disasters. Point-level accurate early warning combined with secondary disaster linkage improves the comprehensiveness of monitoring and the scientific nature of early warning, and supports efficient prevention and control decision-making.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of a geological disaster monitoring system and method based on remote sensing data according to the present invention. Detailed Implementation

[0021] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figure 1 As shown, the present invention provides a geological disaster monitoring method based on remote sensing data, comprising: S1. Acquire multi-source data of the monitoring area, including remote sensing image data, SAR image data, LiDAR data, IoT sensor data and human geography data, and perform hierarchical preprocessing on the multi-source data through edge computing nodes; IoT sensor data includes rainfall data, soil moisture data, and groundwater level data; human geography data includes building height data, road classification data, and underground pipeline distribution data. The hierarchical preprocessing includes: vegetation removal of LiDAR data, denoising of SAR image data, and preliminary denoising of remote sensing image data through edge computing nodes; the remaining preprocessing operations are completed in the cloud; cloud preprocessing includes radiometric calibration, atmospheric correction and geometric correction of remote sensing image data, filtering, phase unwrapping and orbital error removal of SAR image data, and ground point extraction and digital elevation model construction of LiDAR data.

[0024] S2. Based on a three-level fusion architecture of pixel-level, feature-level, and decision-level, the preprocessed multi-source data is fused to obtain a fused dataset. The convergence process of the three-tier convergence architecture includes: Pixel-level fusion: Digital elevation models generated from LiDAR data are used to assist phase unwrapping of SAR image data and eliminate low-coherence interference in vegetated areas; Feature-level fusion: The attention mechanism is used to perform weighted fusion of key features corresponding to each data. Key features include spectral features of remote sensing image data, deformation features of SAR image data, and elevation features of LiDAR data. Decision-level fusion: The disaster type determination results output by the disaster type classifier are used to integrate decision information from various data dimensions and optimize the effectiveness of the fused dataset.

[0025] S3. Perform feature extraction on the fused dataset to obtain a comprehensive feature set containing spectral features, texture features, elevation features and deformation features; S4. Call the preset model pool, which includes landslide-specific prediction models, ground subsidence-specific prediction models, and debris flow-specific prediction models. Identify the disaster type in the monitoring area through a disaster type classifier, and adaptively schedule the corresponding special prediction models. Combine historical disaster data and comprehensive feature set to predict the impact range, disaster severity, and development trend of geological disasters. The landslide-specific prediction model is an improved U-Net + attention mechanism model, the land subsidence-specific prediction model is an LSTM + XGBoost fusion model, and the debris flow-specific prediction model is a CNN + hydrological model coupled model. The disaster type classifier extracts topographic slope features, rainfall correlation features, and land cover features from the comprehensive feature set to achieve automatic disaster type identification and model scheduling.

[0026] S5. Based on the error comparison between measured data and prediction results, trigger the model dynamic fine-tuning mechanism to incrementally update the currently scheduled dedicated prediction model; The model's dynamic fine-tuning mechanism includes: At each preset interval, GNSS measured displacement data and manual survey data of the monitoring area are collected, and the error is calculated with the prediction results of the corresponding period. When the calculated error value exceeds the preset error threshold, federated learning is adopted to fine-tune the current dedicated prediction model in small batches using incrementally collected measured data, without retraining the entire model.

[0027] S6. The prediction results are adapted to complex scenarios by adopting the LiDARDEM-assisted SAR geometric correction algorithm and the combined algorithm of SAR super-resolution reconstruction and small target detection. The LiDARDEM-assisted SAR geometric correction algorithm includes: generating slope and aspect maps and distortion risk maps of the monitoring area based on the LiDAR digital elevation model; using multi-track SAR data fusion to eliminate interference in overlapping areas; and filling data gaps in shadowed areas by interpolating the deformation trends of adjacent non-shadowed areas. The differentiated threshold system sets different warning thresholds based on the scene type of the monitoring area, including residential areas, high-speed rail lines, riverbanks, and mountainous areas.

[0028] S7. Construct a differentiated threshold system, and perform point-level quantitative evaluation on the prediction results after scenario adaptation processing based on the differentiated threshold system. When the evaluation result exceeds the corresponding threshold, trigger a disaster warning and link the secondary disaster prediction sub-model to generate a disaster chain risk assessment result.

[0029] This invention also proposes a geological disaster monitoring system based on remote sensing data, comprising: The multi-source data acquisition and preprocessing module is used to acquire remote sensing image data, SAR image data, LiDAR data, IoT sensor data and human geography data, and complete hierarchical preprocessing through edge computing nodes and cloud processing units; The three-level fusion module is used to fuse preprocessed multi-source data based on a three-level fusion architecture of pixel level, feature level, and decision level, and output a fused dataset. The feature extraction module is used to extract features from the fused dataset to generate a comprehensive feature set that includes spectral features, texture features, elevation features and deformation features; The adaptive prediction module includes a model pool, a disaster type classifier, and a prediction execution unit. The model pool stores prediction models for landslides, ground subsidence, and debris flows. The disaster type classifier is used to identify disaster types and schedule corresponding models. The prediction execution unit is used to combine historical data and a comprehensive feature set to output prediction results. The dynamic calibration module is used to collect measured data, calculate the error between the measured data and the prediction results, and trigger incremental fine-tuning of the model. The scene adaptation module includes a distortion correction subunit and a small target detection subunit. The distortion correction subunit adopts the LiDARDEM-assisted SAR geometric correction algorithm, and the small target detection subunit adopts the SAR super-resolution reconstruction and YOLOv8-tiny combined algorithm. The early warning linkage module is used to construct a differentiated threshold system, conduct point-level quantitative assessment, trigger disaster early warning, and link with the secondary disaster prediction sub-model to generate disaster chain risk assessment results.

[0030] The multi-source data acquisition and preprocessing module includes: The remote sensing data subunit is used to receive periodic scan data from medium-resolution satellites, key monitoring data from high-resolution satellites, synthetic aperture radar data, and airborne LiDAR data from UAVs. The Internet of Things (IoT) sub-unit is used to connect to rainfall sensors, soil moisture sensors, and groundwater level sensors to collect corresponding IoT sensor data. The humanities data interface is used to connect to the city's GIS system to obtain data on building height, road grade, and underground pipeline distribution. The edge processing unit is used to remove vegetation from LiDAR data, denoise SAR data, and perform preliminary noise reduction on remote sensing images. The cloud processing unit is used to perform radiometric calibration, atmospheric correction, and geometric correction of remote sensing images; SAR data filtering, phase unwrapping, orbital error removal; and LiDAR data ground point extraction and digital elevation model construction.

[0031] In the model pool of the adaptive prediction module, the landslide-specific prediction model is an improved U-Net + attention mechanism model with an embedded edge enhancement module; the land subsidence-specific prediction model is a fusion model of LSTM time series prediction model and XGBoost factor analysis model; the debris flow-specific prediction model is a coupled model of CNN landslide source area identification model and hydrological runoff calculation model; the disaster type classifier realizes automatic disaster type determination and model scheduling by extracting the topographic, rainfall and surface cover correlation features from the comprehensive feature set.

[0032] The distortion correction subunit of the scene adaptation module generates slope aspect map and distortion risk map through LiDAR digital elevation model, eliminates overlapping interference by fusing ascending and descending SAR data, and fills the shadow area data by interpolating the deformation trend of adjacent areas. The secondary disaster prediction sub-model of the early warning linkage module includes a landslide dam prediction sub-model, which is used to estimate the rate of rise of the landslide dam water level and the risk of dam failure based on debris flow accumulation, river width and flow data.

[0033] This invention focuses on a geologically hazardous area in the southwestern mountainous region (with a total area of ​​approximately 500 km²). 2Taking the area as an example (covering low and medium mountains, river terraces, rural settlements, and provincial highways, with high vegetation coverage accounting for 42% and steep slopes with a gradient greater than 30° accounting for 28%), a special geological disaster monitoring project was carried out for a period of 6 months.

[0034] During the data acquisition phase, the Sentinel-1 SAR satellite (with both ascending and descending orbits and a revisit cycle of 6 days) was used to acquire deformation data, and the Gaofen-6 optical satellite (16m multispectral resolution) was used for periodic scanning. For three suspected anomaly areas on the river valley slopes, the Gaofen-2 satellite (0.8m panchromatic resolution) was used to conduct key monitoring once a week. A hexacopter UAV equipped with a RIEGLVUX-1UAV airborne LiDAR was used to conduct a full-coverage scan of the monitoring area at a flight altitude of 50m and a 100% point cloud overlap rate. Simultaneously, artificial corner reflectors were deployed at 12 key locations, including landslide hazard points and along provincial highways, to assist in data calibration. The system was also integrated with 15 automatic weather stations and 30 soil moisture sensors in the area (with a deployment density of approximately 17km). 2 ( / each), real-time collection of rainfall and soil volumetric moisture content data, and retrieval of regional 1:10,000 topographic data, building survey data and historical landslide records of the past 10 years through the urban GIS platform interface.

[0035] In the data preprocessing stage, edge computing nodes are deployed at regional monitoring stations. The progressive densification triangulation method is used to extract ground points and remove vegetation from LiDAR data. Gaussian filtering is used to perform preliminary denoising on SAR images, and radiometric calibration preprocessing is performed on optical remote sensing images. The cloud processing unit further performs atmospheric correction (6S model), geometric correction (based on ground control points), and noise reduction (median filtering) on ​​optical images. Goldstein filtering, phase unwrapping (branch cutting method), and orbital error removal (based on precise orbital data) are applied to SAR images. A 1m resolution digital elevation model (DEM) is constructed based on LiDAR ground points.

[0036] In the multi-source data fusion stage, pixel-level fusion uses the slope and aspect map generated by LiDARDEM to locate low-coherence areas in SAR images and assists in the phase unwrapping process to eliminate vegetation interference; feature-level fusion adopts a co-attention mechanism to weight and fuse the spectral features of optical images, the deformation features of SAR images, the elevation features of LiDAR, and the environmental features of the Internet of Things, dynamically allocating the weights of each feature; decision-level fusion uses a disaster type classifier (trained based on historical disaster data and the correlation features of terrain and rainfall) to automatically identify the two main disaster risks of landslides and ground subsidence in the region.

[0037] During the model application, for the 8 identified landslide hazard points, an improved U-Net + attention mechanism model with embedded edge enhancement module was scheduled to accurately extract the landslide body boundary and deformation trend. For areas along provincial highways and rural settlements, an LSTM (time window set to 30 days) + XGBoost fusion model was used to predict ground subsidence development by combining groundwater level and building load data. Measured displacement data (accuracy ±2mm) from 10 GNSS monitoring points were collected every 7 days, and the mean square error between the prediction results and the measured data was calculated. When the error exceeded 8%, federated learning was used to fine-tune the model with incremental data in small batches.

[0038] In terms of adapting to complex scenes, for steep slope areas, the data from ascending Sentinel-1A and descending Sentinel-1B orbits are fused, and overlapping phenomena are eliminated through angle complementarity. For shadowed areas, deformation trend interpolation of adjacent non-shadowed areas is used to fill data gaps. The SRGAN super-resolution model is used to improve the SAR image resolution from 5m to 1m, and combined with the YOLOv8-tiny model (confidence threshold set at 0.65), three areas with an area <80m² along the provincial highway are accurately identified. 2 Small-scale landslide hazards.

[0039] During the early warning and linkage phase, based on a differentiated threshold system, the early warning threshold for ground subsidence in residential areas is set at 20mm / 6 months, along high-speed rail lines (provincial highway branches within the region) at 10mm / 6 months, and around river channels at 5mm / month for landslide displacement rate. When a landslide displacement rate of 6.2mm / month is detected on a river valley slope, a point-level early warning is immediately triggered, and the secondary disaster prediction sub-model is linked simultaneously. The estimated volume of the landslide (calculated based on LiDARDEM), the regional slope, and the cumulative rainfall in the past 7 days are input to assess the runoff path and impact range of the debris flow caused by the landslide and clarify the potential risks to two downstream villages.

[0040] Through monthly field surveys conducted in conjunction with the local Natural Resources Bureau and Geological Exploration Institute, and six months of continuous data comparison and verification, this solution effectively addresses core issues such as low coherence of SAR data in high-vegetation areas, data distortion in steep terrain, missed detection of small-scale disasters, and crude early warning systems. It enables accurate identification, dynamic monitoring, and coordinated early warning of secondary disasters for different types of disasters, providing comprehensive and reliable technical support for regional geological disaster prevention and control.

[0041] Therefore, this invention provides a geological disaster monitoring system and method based on remote sensing data, which deeply integrates remote sensing, Internet of Things and human data from multiple sources. It solves the problems of insufficient data fusion, poor model generalization, weak adaptability to complex scenarios, crude early warning and lack of linkage with secondary disasters in existing technologies. It achieves accurate adaptation to different disaster types and dynamic model calibration, effectively identifies small-scale disasters, outputs point-level quantitative early warning, improves the comprehensiveness, accuracy and scientific nature of monitoring and early warning, and provides efficient decision support for geological disaster prevention and control.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A geological disaster monitoring method based on remote sensing data, characterized in that, Includes the following steps: S1. Acquire multi-source data of the monitoring area, including remote sensing image data, SAR image data, LiDAR data, IoT sensor data and human geography data, and perform hierarchical preprocessing on the multi-source data through edge computing nodes; S2. Based on a three-level fusion architecture of pixel-level, feature-level, and decision-level, the preprocessed multi-source data is fused to obtain a fused dataset. S3. Perform feature extraction on the fused dataset to obtain a comprehensive feature set containing spectral features, texture features, elevation features and deformation features; S4. Call the preset model pool, which includes a landslide-specific prediction model, a ground subsidence-specific prediction model, and a debris flow-specific prediction model. Identify the disaster type in the monitoring area through a disaster type classifier, and adaptively schedule the corresponding special prediction model. Combine historical disaster data and comprehensive feature set to predict the impact range, disaster severity, and development trend of geological disasters. S5. Based on the error comparison between measured data and prediction results, trigger the model dynamic fine-tuning mechanism to incrementally update the currently scheduled dedicated prediction model; S6. The prediction results are adapted to complex scenarios by adopting the LiDARDEM-assisted SAR geometric correction algorithm and the combined algorithm of SAR super-resolution reconstruction and small target detection. S7. Construct a differentiated threshold system, and perform point-level quantitative evaluation on the prediction results after scenario adaptation processing based on the differentiated threshold system. When the evaluation result exceeds the corresponding threshold, trigger a disaster warning and link the secondary disaster prediction sub-model to generate a disaster chain risk assessment result.

2. The geological disaster monitoring method based on remote sensing data according to claim 1, characterized in that, In S1, the IoT sensor data includes rainfall data, soil moisture data, and groundwater level data; the human geography data includes building height data, road grade data, and underground pipeline distribution data. The hierarchical preprocessing includes: vegetation removal from LiDAR data, denoising of SAR image data, and preliminary denoising of remote sensing image data through edge computing nodes; the remaining preprocessing operations are completed in the cloud. The cloud preprocessing includes radiometric calibration, atmospheric correction, and geometric correction of remote sensing image data; filtering, phase unwrapping, and orbital error removal of SAR image data; and ground point extraction and digital elevation model construction of LiDAR data.

3. The geological disaster monitoring method based on remote sensing data according to claim 1, characterized in that, In S2, the fusion process of the three-level fusion architecture includes: Pixel-level fusion: Digital elevation models generated from LiDAR data are used to assist phase unwrapping of SAR image data and eliminate low-coherence interference in vegetated areas; Feature-level fusion: The attention mechanism is used to perform weighted fusion of the key features corresponding to each data. The key features include the spectral features of remote sensing image data, the deformation features of SAR image data, and the elevation features of LiDAR data. Decision-level fusion: The disaster type determination results output by the disaster type classifier are used to integrate decision information from various data dimensions and optimize the effectiveness of the fused dataset.

4. The geological disaster monitoring method based on remote sensing data according to claim 1, characterized in that, In S4, the landslide-specific prediction model is an improved U-Net + attention mechanism model, the ground subsidence-specific prediction model is an LSTM + XGBoost fusion model, and the debris flow-specific prediction model is a CNN + hydrological model coupled model. The disaster type classifier extracts terrain slope features, rainfall correlation features, and land cover features from the comprehensive feature set to achieve automatic disaster type identification and model scheduling.

5. The geological disaster monitoring method based on remote sensing data according to claim 1, characterized in that, In S5, the model dynamic fine-tuning mechanism includes: At each preset interval, GNSS measured displacement data and manual survey data of the monitoring area are collected, and the error is calculated with the prediction results of the corresponding period. When the calculated error value exceeds the preset error threshold, federated learning is adopted to fine-tune the current dedicated prediction model in small batches using incrementally collected measured data, without retraining the entire model.

6. The geological disaster monitoring method based on remote sensing data according to claim 1, characterized in that, In S6, the LiDARDEM-assisted SAR geometric correction algorithm includes: generating a slope aspect map and distortion risk map of the monitoring area based on the LiDAR digital elevation model; using multi-track SAR data fusion to eliminate interference in overlapping areas; and filling data gaps in the shadow area by interpolating the deformation trends of adjacent non-shadow areas. The differentiated threshold system sets different warning thresholds according to the scene type of the monitoring area. The scene types include residential areas, high-speed rail lines, riverbanks, and mountainous areas.

7. A geological disaster monitoring system based on remote sensing data, characterized in that, The method for monitoring geological hazards based on remote sensing data according to any one of claims 1-6 includes: The multi-source data acquisition and preprocessing module is used to acquire remote sensing image data, SAR image data, LiDAR data, IoT sensor data and human geography data, and complete hierarchical preprocessing through edge computing nodes and cloud processing units; The three-level fusion module is used to fuse preprocessed multi-source data based on a three-level fusion architecture of pixel level, feature level, and decision level, and output a fused dataset. The feature extraction module is used to extract features from the fused dataset to generate a comprehensive feature set that includes spectral features, texture features, elevation features and deformation features; The adaptive prediction module includes a model pool, a disaster type classifier, and a prediction execution unit. The model pool stores prediction models for landslides, ground subsidence, and debris flows. The disaster type classifier is used to identify disaster types and schedule corresponding models. The prediction execution unit is used to combine historical data and a comprehensive feature set to output prediction results. The dynamic calibration module is used to collect measured data, calculate the error between the measured data and the prediction results, and trigger incremental fine-tuning of the model. The scene adaptation module includes a distortion correction subunit and a small target detection subunit. The distortion correction subunit adopts the LiDARDEM-assisted SAR geometric correction algorithm, and the small target detection subunit adopts the SAR super-resolution reconstruction and YOLOv8-tiny combined algorithm. The early warning linkage module is used to construct a differentiated threshold system, conduct point-level quantitative assessment, trigger disaster early warning, and link with the secondary disaster prediction sub-model to generate disaster chain risk assessment results.

8. The geological disaster monitoring system based on remote sensing data according to claim 7, characterized in that, The multi-source data acquisition and preprocessing module includes: The remote sensing data subunit is used to receive periodic scan data from medium-resolution satellites, key monitoring data from high-resolution satellites, synthetic aperture radar data, and airborne LiDAR data from UAVs. The Internet of Things (IoT) sub-unit is used to connect to rainfall sensors, soil moisture sensors, and groundwater level sensors to collect corresponding IoT sensor data. The humanities data interface is used to connect to the city's GIS system to obtain data on building height, road grade, and underground pipeline distribution. The edge processing unit is used to remove vegetation from LiDAR data, denoise SAR data, and perform preliminary noise reduction on remote sensing images. The cloud processing unit is used to perform radiometric calibration, atmospheric correction, and geometric correction of remote sensing images; SAR data filtering, phase unwrapping, orbital error removal; and LiDAR data ground point extraction and digital elevation model construction.

9. The geological disaster monitoring system based on remote sensing data according to claim 7, characterized in that, In the model pool of the adaptive prediction module, the landslide-specific prediction model is an improved U-Net + attention mechanism model with an embedded edge enhancement module; the ground subsidence-specific prediction model is a fusion model of LSTM time-series prediction model and XGBoost factor analysis model; the debris flow-specific prediction model is a coupled model of CNN landslide source area identification model and hydrological runoff calculation model; the disaster type classifier realizes automatic disaster type determination and model scheduling by extracting topographic, rainfall and surface cover correlation features from the comprehensive feature set.

10. The geological disaster monitoring system based on remote sensing data according to claim 7, characterized in that, The distortion correction subunit of the scene adaptation module generates a slope aspect map and a distortion risk map through a LiDAR digital elevation model, eliminates overlapping interference by fusing ascending and descending SAR data, and fills in the shadow area data by interpolating the deformation trend of adjacent areas. The secondary disaster prediction sub-model of the early warning linkage module includes a landslide dam prediction sub-model, which is used to estimate the rate of rise of the landslide dam water level and the risk of dam failure based on debris flow accumulation, river width and flow data.

Images