Dynamic monitoring and early warning method for water and soil loss in water conservancy project

Abstract

The invention relates to the technical field of water and soil loss dynamic monitoring and early warning methods in water conservancy projects, and discloses a water and soil loss dynamic monitoring and early warning method, which comprises the following steps: synchronously acquiring multi-dimensional data of terrain, rainfall, soil humidity and vegetation coverage through unmanned aerial vehicle remote sensing and a ground sensor network; registering and correcting the multi-source heterogeneous data based on a space-time fusion algorithm; the improved SWAT model is combined with a machine learning algorithm to dynamically simulate a water and soil loss process; and triggering graded early warning according to a preset risk threshold and generating prevention and control suggestions. The system comprises a data acquisition module, a data fusion processing module, a water and soil loss simulation prediction module and an early warning response module. According to the scheme, the real-time performance, the spatial resolution and the early warning accuracy of water and soil loss monitoring are remarkably improved, and efficient decision support is provided for ecological protection of water conservancy projects.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of a water and soil loss dynamic monitoring and early warning method in water conservancy projects. BACKGROUND

[0002] In order to realize scientific prevention and control of the water and soil loss process, dynamic monitoring and early warning technology has increasingly become a core link of intelligent management of water conservancy projects. Such technology usually relies on a wireless sensor network arranged in a watershed or a slope area, acquires multiple source environmental parameters such as rainfall, soil moisture content, surface runoff and displacement in real time, constructs a water and soil loss risk assessment model, and thus supports decision response.

[0003] The water and soil loss dynamic monitoring method based on a wireless sensor network aims to continuously acquire information about changes in the state of a terrain surface through distributed sensing nodes, and gather data to a central platform for analysis and processing. The basic principle is to use sensors to perform high-frequency sampling on key hydrological and geological indexes, identify hot erosion areas in combination with a time-space correlation algorithm, and trigger a hierarchical early warning mechanism according to threshold rules. However, this technical route faces severe challenges in typical water conservancy scenes such as complex mountainous areas, gorges or dense vegetation.

[0004] The wireless sensor network generally adopts a static topology structure or a fixed clustering strategy for data transmission, and does not fully consider the shielding effect of terrain undulations on wireless signal propagation. In areas with sharp changes in elevation or obstacles, communication links between nodes are prone to failure due to line-of-sight interruption, causing local data to be unable to be returned and forming a monitoring blind area. At the same time, traditional relay node selection relies on a preset position or a simple distance metric, and is difficult to adapt to terrain micro-variations caused by dynamic processes such as rainfall erosion, soil displacement, etc., resulting in continuous deterioration of network connectivity. In addition, since water and soil loss events have the characteristics of strong suddenness and rapid evolution, data transmission interruption will directly lead to the lack of key process information, seriously affecting the timeliness and accuracy of the early warning model. SUMMARY

[0005] In order to solve the above problems, the application provides a water and soil loss dynamic monitoring and early warning method in water conservancy projects, comprising: deploying a multi-level heterogeneous sensing network to acquire ground sensing data, the ground sensing data including soil moisture content, surface runoff speed, rainfall, slope angle and micro-meteorological parameters; acquiring multi-source remote sensing image data, the multi-source remote sensing image data including optical satellite images, synthetic aperture radar images and unmanned aerial vehicle aerial images; performing adaptive time-space alignment and feature-level fusion on the ground sensing data and the multi-source remote sensing image data to generate a water and soil loss comprehensive sensing data set; Based on the comprehensive soil and water loss perception dataset, soil and water loss risk factors are extracted and a dynamic risk model is constructed. The soil and water loss risk factors include runoff erosivity index, soil erodibility index, vegetation protection index and topography driving index. The dynamic risk model is used to predict the spatiotemporal evolution trend of soil erosion, and a graded early warning signal is generated based on a preset threshold. The tiered early warning signal is pushed to the management terminal, and feedback control instructions are triggered to adjust the monitoring strategy or engineering measures.

[0006] Preferably, a multi-level heterogeneous sensing network is deployed to acquire ground sensing data, including: The ground sensing data is collected through the monitoring terminal layer, which includes distributed ground sensing units. Each ground sensing unit integrates a soil moisture sensor, a surface runoff velocity meter, a rainfall meter, a slope inclinometer, and a micro-weather station. The edge computing node layer receives and processes data from the monitoring terminal layer, and the edge computing node layer performs local data caching, preliminary filtering and anomaly detection. A communication relay layer is deployed in the communication obstruction area. The communication relay layer uses frequency hopping spread spectrum technology to dynamically select channels in the 400 MHz to 900 MHz frequency band to achieve reliable multi-hop transmission.

[0007] Preferably, the ground-based sensor data and the multi-source remote sensing image data are adaptively aligned in time and space and fused at the feature level to generate a comprehensive soil and water loss perception dataset, including: Establish a unified spatiotemporal grid with a grid cell size of 50 meters × 50 meters and a time step of 1 hour; The ground sensing data is mapped to the nearest grid cell according to geographical location and aligned to the hour using cubic spline interpolation. The multi-source remote sensing image data is resampled to the unified spatiotemporal grid and fused into a single temporal data using a time-weighted average method; Soil moisture content, surface runoff velocity, and rainfall intensity are used as dynamic input features, while vegetation cover, surface roughness, and slope are used as static background features to construct a multidimensional feature tensor. Each feature in the multidimensional feature tensor is normalized, and the normalization parameter is determined based on the minimum and maximum values ​​of historical three years of data.

[0008] Preferably, the method of extracting soil erosion risk factors and constructing a dynamic risk model based on the comprehensive soil erosion perception dataset includes: The runoff erosivity index is calculated by multiplying and summing the runoff kinetic energy and runoff volume per unit time. The soil erodibility index is determined based on soil texture, organic matter content, and aggregate stability. The vegetation protection index is calculated based on the weighted sum of vegetation cover and leaf area index. The topographic driving index was calculated using the topographic factor method in the general soil loss equation. The runoff erosivity index, soil erodibility index, vegetation protection index, and topographic driving index are weighted and summed to generate a comprehensive risk value. The weighting coefficients are determined by inverting historical disaster samples based on the regional hydrogeological characteristics.

[0009] Preferably, the dynamic risk model is used to predict the spatiotemporal evolution trend of soil erosion, and a graded early warning signal is generated based on a preset threshold, including: The comprehensive risk value sequence of the past 24 hours is input into a spatiotemporal convolutional long short-term memory network to predict the risk value of the next 12 hours. The spatiotemporal convolutional long short-term memory network comprises three layers of spatiotemporal convolutional modules and two layers of long short-term memory units. Residual analysis is performed between the predicted results and the measured risk values. If the predicted residual exceeds 15% for three consecutive hours, the online fine-tuning mechanism of the model is triggered. Blue, yellow, orange, or red warning signals are generated based on the range of the comprehensive risk value. Each warning level corresponds to a different risk threshold and response measures.

[0010] Preferably, the graded early warning signal is pushed to the management terminal, and a feedback control command is triggered to adjust the monitoring strategy or engineering measures, including: The warning signal is pushed to the mobile application and the command center screen via message queue telemetry transmission protocol; Under a blue alert, increase the frequency of drone inspections; Under a yellow alert, the backup sensing unit is activated and the data reporting frequency is increased to once per minute; Under an orange alert, remotely controlled, retractable sand-blocking fences are deployed, and the local storage acceleration mode of edge nodes is activated. Under a red alert, non-essential power supplies are cut off, with only the BeiDou short message communication module remaining operational to continuously report location and status.

[0011] Preferably, the soil moisture sensor adopts the frequency domain reflectance principle, with a sampling frequency of once per minute, a measurement range of 0% to 50%, and an accuracy of ±0.5%; the surface runoff velocity meter is based on the ultrasonic Doppler effect, with a sampling frequency of once every 10 seconds and a range of 0 to 3 meters per second; the rainfall meter has a tipping bucket structure and a resolution of 0.2 millimeters; the slope inclinometer uses a three-axis MEMS accelerometer and gyroscope fusion calculation, with an update cycle of 1 second; the micro-weather station collects ambient temperature, relative humidity, wind speed, and wind direction, with a data update frequency of once every 30 seconds; all sensor data are appended with a timestamp provided by the BeiDou time synchronization module, with a time synchronization error of less than 10 milliseconds.

[0012] Preferably, the optical satellite imagery has a spatial resolution of 10 meters and a revisit period of 5 days; the synthetic aperture radar imagery uses the C-band, has a spatial resolution of 20 meters, and a revisit period of 6 days; the UAV aerial imagery has a ground sampling distance of 5 centimeters and a flight altitude of 300 meters; all remote sensing images undergo radiometric correction, geometric correction, and atmospheric correction, and the coordinate system is converted to the National Geodetic 2000 coordinate system.

[0013] Preferably, the dynamic risk model operates with a 72-hour sliding time window, updating the comprehensive risk value of each grid cell hourly; the weighting coefficients satisfy... + + + =1, where , , The weights correspond to the runoff erosivity index, soil erodibility index, vegetation protection index, and topographic driving index, respectively.

[0014] Preferably, the blue warning corresponds to a comprehensive risk value greater than 0.3 and less than or equal to 0.5; the yellow warning corresponds to a risk value greater than 0.5 and less than or equal to 0.7; the orange warning corresponds to a risk value greater than 0.7 and less than or equal to 0.9; and the red warning corresponds to a risk value greater than 0.9. The warning thresholds are dynamically adjusted according to the season, watershed characteristics, and project importance, and the adjustment rules are provided by an expert knowledge base.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1. By constructing a multi-level heterogeneous sensing network consisting of a monitoring terminal layer, an edge computing node layer, and a communication relay layer, and addressing the signal obstruction problem in complex terrain, frequency-hopping spread spectrum multi-hop transmission technology and solar-powered self-organizing network relay nodes are employed, combined with a spatial kriging interpolation compensation mechanism. This achieves full-area coverage and continuous acquisition of core parameters such as soil moisture content, surface runoff velocity, and rainfall. This completely solves the problems of communication link failure and data transmission interruption inherent in traditional static sensor networks, ensuring no data loss throughout the entire process of soil erosion and guaranteeing the integrity and continuity of the monitoring system.

[0017] 2. An innovative adaptive spatiotemporal alignment and feature-level fusion strategy is adopted to accurately register ground-based dynamic sensing data with multi-source remote sensing static / macroscopic data. Through unified spatiotemporal grids, cubic spline interpolation, time-weighted averaging, and normalization, a multi-dimensional feature tensor is constructed, eliminating dimensional differences and spatiotemporal misalignments among the heterogeneous data from multiple sources. The fused integrated sensing dataset is stored in NetCDF format, supporting efficient reading and parallel computing. Compared to a single data source, this significantly improves data integrity and accuracy, providing comprehensive and reliable foundational support for risk modeling.

[0018] 3. Based on four core risk factors—runoff erosivity, soil erodibility, vegetation protection, and topographical factors—a 72-hour sliding time window dynamic risk model is constructed. This model updates the comprehensive risk value of grid cells hourly, overcoming the limitations of traditional static models that cannot adapt to the sudden and rapid evolution of soil erosion. A spatiotemporal convolutional LSTM network is used to predict risk trends for the next 12 hours. Residual analysis and online fine-tuning mechanisms ensure that prediction errors are controlled within a reasonable range. Compared to existing technologies, the early warning response lag time is significantly shortened, achieving a shift from passive response to proactive prediction.

[0019] 4. Establish a four-level early warning system (blue-yellow-orange-red), formulate differentiated and implementable response measures for different risk levels, and use the MQTT protocol to push early warning signals to mobile terminals and command centers with low latency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the dynamic monitoring and early warning method for soil erosion in water conservancy projects proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of adaptive fusion of multi-source remote sensing and ground sensing data in this invention; Figure 3 This is a flowchart illustrating the logical process of extracting and dynamically modeling soil and water loss risk factors in this invention. Figure 4 This is a logical flowchart of the water and soil erosion trend prediction and graded early warning based on spatiotemporal evolution characteristics in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the monitoring terminal, edge computing node and cloud platform in this invention; Figure 6 This is a closed-loop collaborative logic framework diagram of the early warning response and feedback control mechanism in this invention. Detailed Implementation

[0021] Please refer to Figures 1 to 6 In complex terrain conditions, traditional wireless sensor networks deployed in water conservancy engineering areas often experience communication link interruptions due to obstruction by mountains, vegetation, or man-made structures, creating data transmission blind spots. These blind spots prevent the real-time collection and transmission of key parameters related to soil erosion, such as soil moisture content, surface runoff velocity, and slope erosion, leading to a delayed response of the monitoring system to sudden soil erosion events and hindering accurate early warning and emergency control. To address this technical problem, this invention proposes a dynamic monitoring and early warning method for soil erosion in water conservancy engineering. This method constructs a multi-level heterogeneous sensing network, integrates multi-source remote sensing and ground sensor data, establishes a risk dynamic modeling mechanism based on spatiotemporal evolution characteristics, and introduces an edge-cloud collaborative computing architecture to achieve continuous, comprehensive, and high-precision perception and hierarchical early warning of the soil erosion process.

[0022] The method includes the following steps: S1, deploy a multi-level heterogeneous sensing network to acquire ground sensing data; S2, acquire multi-source remote sensing image data; S3, perform adaptive spatiotemporal alignment and feature-level fusion on the ground sensor data and the multi-source remote sensing image data to generate a comprehensive soil and water loss perception dataset. S4. Based on the comprehensive soil and water loss perception dataset, extract soil and water loss risk factors and construct a dynamic risk model; S5. Based on the dynamic risk model, predict the spatiotemporal evolution trend of soil erosion and generate graded early warning signals according to preset thresholds. S6, push the graded early warning signal to the management terminal and trigger feedback control instructions to adjust the monitoring strategy or engineering measures.

[0023] In step S1, a multi-level heterogeneous sensing network is deployed to acquire ground sensing data. This multi-level heterogeneous sensing network consists of a monitoring terminal layer, an edge computing node layer, and a communication relay layer. The monitoring terminal layer includes several distributed ground sensing units, each integrating a soil moisture sensor, a surface runoff velocity meter, a rainfall meter, a slope inclinometer, and a micro-weather station. The soil moisture sensor uses the frequency domain reflectance principle, outputting volumetric water content values ​​at a sampling frequency of once per minute, with a measurement range of 0% to 50% and an accuracy of ±0.5%. The surface runoff velocity meter, based on the ultrasonic Doppler effect, is installed at the bottom of the slope runoff channel to measure the surface velocity of water flow in real time, with a sampling frequency of once every 10 seconds and a range of 0 to 3 meters per second.

[0024] The rain gauge is a tipping bucket structure with a resolution of 0.2 mm, recording cumulative rainfall and instantaneous rainfall intensity. The slope inclinometer uses a three-axis MEMS accelerometer and gyroscope fusion calculation to output the three-dimensional slope inclinometer, with an update cycle of 1 second. The micro-weather station collects ambient temperature, relative humidity, wind speed, and wind direction, with a data update frequency of once every 30 seconds. All sensor data are appended with high-precision timestamps, and the time reference is uniformly provided by the BeiDou time synchronization module, with a time synchronization error of less than 10 milliseconds.

[0025] The monitoring terminal layer communicates with the edge computing node layer via a low-power wide-area network (LPWAN) protocol. In areas with severe signal obstruction, a communication relay layer is added. These relay nodes are solar-powered and possess self-organizing network capabilities. They dynamically select the optimal channel within the 400 MHz to 900 MHz frequency band using frequency hopping spread spectrum technology, ensuring reliable data packet transmission across multi-hop paths. Each edge computing node has a coverage radius of no more than 2 kilometers and connects to no fewer than 20 ground sensing units. The edge computing nodes have built-in embedded processors and solid-state storage, performing local data caching, preliminary filtering, and anomaly detection. When a ground sensing unit fails to respond to heartbeat packets three times consecutively, the edge computing node determines that its communication has been interrupted and initiates a data interpolation compensation mechanism for neighboring nodes. It uses spatial kriging interpolation to estimate the soil moisture content and runoff velocity at missing points. The interpolation weights are determined by Euclidean distance and terrain shading factors. The terrain shading factors are pre-calculated using a digital elevation model and stored in the edge node's local database.

[0026] In step S2, multi-source remote sensing image data is acquired. This multi-source remote sensing image data includes optical satellite imagery, synthetic aperture radar (SAR) imagery, and UAV aerial imagery. Optical satellite imagery is sourced from the Gaofen series or Sentinel-2 satellites, with a spatial resolution of 10 meters and a revisit period of 5 days. It is used to extract land cover type, vegetation cover, and the proportion of bare soil area. SAR imagery is sourced from the Sentinel-1 satellite, using the C-band, with a spatial resolution of 20 meters and a revisit period of 6 days. It possesses all-weather observation capabilities and is used to invert surface roughness, soil moisture, and surface deformation information. UAV aerial imagery is acquired by a fixed-wing UAV equipped with a multispectral camera, flying at an altitude of 300 meters and a ground sampling distance of 5 centimeters. Routine inspections are performed weekly, increasing to daily during heavy rain warnings. This is used to identify gully erosion, gully development, and signs of localized landslides. All remote sensing images undergo radiometric, geometric, and atmospheric correction. The coordinate system is converted to the National Geodetic 2000 coordinate system, and the time label is accurate to the image center time.

[0027] In step S3, adaptive spatiotemporal alignment and feature-level fusion are performed on the ground-sensing data and the multi-source remote sensing image data to generate a comprehensive soil and water loss perception dataset. First, a unified spatiotemporal grid is established, with a grid cell size of 50 meters × 50 meters and a time step of 1 hour. For the ground-sensing data, it is mapped to the nearest grid cell based on geographical location, and non-integer time data is aligned to the hour using cubic spline interpolation. For the remote sensing image data, it is resampled to the unified grid using bilinear interpolation, and multi-day images are fused into a single temporal data using a time-weighted average method. Subsequently, feature-level fusion is performed: soil moisture content, surface runoff velocity, and rainfall intensity are used as dynamic input features, and vegetation cover, surface roughness, and slope are used as static background features to construct a multi-dimensional feature tensor. Each dimension of this tensor corresponds to a physical quantity, and each grid cell corresponds to a feature vector at each time step. To eliminate dimensional differences, all features are normalized using the following normalization formula:

[0028] in, These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature in the historical three-year data set, respectively. These are the normalized feature values. The normalization parameters are stored in the central database of the cloud platform and updated regularly. The fused comprehensive soil and water loss sensing dataset is stored in NetCDF format, including spatial, temporal, and feature dimensions, supporting efficient reading and parallel computation.

[0029] In step S4, soil erosion risk factors are extracted based on the comprehensive soil erosion perception dataset, and a dynamic risk model is constructed. The risk factors include runoff erosivity index, soil erodibility index, vegetation protection index, and topographic driving index. The runoff erosivity index is calculated by multiplying runoff kinetic energy and runoff volume per unit time, and its expression is:

[0030] in, For the first Surface runoff velocity at each time step This represents the corresponding flow rate per unit width. For time step, This represents the duration of the current rainfall event in steps. The soil erodibility index is determined by a combination of soil texture, organic matter content, and aggregate stability, mapped to a dimensionless value between 0 and 1 using a laboratory calibration curve. The vegetation protection index is defined as a weighted sum of vegetation cover and leaf area index, with weighting coefficients preset according to the regional ecological type. The topographic driving index is determined by both slope and slope length, calculated using the topographic factor method in the general soil loss equation. The above four risk factors are calculated independently in each grid cell, and a comprehensive risk value is generated through weighted summation.

[0031] in, The soil erodibility index. The vegetation protection index, Terrain-driven index, , , For the weighting coefficients, satisfying + + + =1, the weight value is determined by inversion of historical disaster samples based on the regional hydrogeological characteristics and stored in the model parameter library. The dynamic risk model runs in a sliding time window mode with a time window length of 72 hours, and updates the comprehensive risk value of each grid cell every hour.

[0032] In step S5, the spatiotemporal evolution trend of soil erosion is predicted based on the dynamic risk model, and a graded early warning signal is generated according to a preset threshold. The prediction employs a spatiotemporal convolutional long short-term memory network. The network input is the comprehensive risk value sequence of the past 24 hours, and the output is the predicted risk value for the next 12 hours. The network structure includes three layers of spatiotemporal convolutional modules, each with a kernel size of 3×3, a time step of 1, and channel numbers of 16, 32, and 64 respectively, followed by two layers of long short-term memory units, with a total of 128 hidden units. The model is trained on a cloud platform, and the inference process is executed by edge computing nodes to reduce cloud load. Residual analysis is performed between the prediction results and the measured risk values. If the prediction residual exceeds 15% for three consecutive hours, an online fine-tuning mechanism is triggered to update the network weights using the latest data.

[0033] The tiered early warning signals are divided into four levels: blue, yellow, orange, and red. A blue warning corresponds to a comprehensive risk value greater than 0.3 and less than or equal to 0.5, indicating a slight risk of soil erosion, requiring increased patrols. A yellow warning corresponds to a risk value greater than 0.5 and less than or equal to 0.7, indicating a moderate risk, suggesting the initiation of temporary protective measures. An orange warning corresponds to a risk value greater than 0.7 and less than or equal to 0.9, indicating a high risk, requiring immediate deployment of sand-trapping dams or covering nets. A red warning corresponds to a risk value greater than 0.9, indicating an extremely high risk, potentially triggering landslides or debris flows, necessitating the evacuation of personnel and the closure of relevant water conservancy facilities. The warning thresholds can be dynamically adjusted based on season, watershed characteristics, and project importance; the adjustment rules are provided by an expert knowledge base.

[0034] In step S6, the graded early warning signal is pushed to the management terminal, triggering feedback control commands to adjust monitoring strategies or engineering measures. The management terminal includes a mobile application and a command center screen. The early warning signal is pushed through a message queue telemetry transmission protocol to ensure low-latency delivery. Simultaneously, the system automatically generates feedback control commands: under a blue warning, increase the frequency of UAV inspections; under a yellow warning, activate backup sensing units and increase data reporting frequency to once per minute; under an orange warning, remotely control the deployment of the retractable sand-blocking fence and activate the local storage acceleration mode of the edge nodes; under a red warning, cut off unnecessary power supply, retaining only the BeiDou short message communication module to continuously report location and status. The execution status of all control commands is transmitted back from the edge nodes to the cloud platform, forming a closed-loop feedback. The cloud platform records the entire process data of each early warning event for subsequent model optimization and contingency plan revision.

[0035] The proposed method addresses monitoring blind spots caused by signal obstruction through a multi-level sensing network, enhances sensing accuracy through multi-source data fusion, achieves trend prediction using a dynamic risk model, and ensures timely response through tiered early warning and closed-loop control. The entire process operates within an edge-cloud collaborative architecture, balancing real-time performance with computational depth, and is suitable for water conservancy projects in complex terrain to control soil erosion.

Claims

1. A method for dynamic monitoring and early warning of soil erosion in water conservancy projects, characterized in that, include: A multi-level heterogeneous sensing network is deployed to acquire ground sensing data, including soil moisture content, surface runoff velocity, rainfall, slope angle, and micro-meteorological parameters. Acquire multi-source remote sensing image data, including optical satellite imagery, synthetic aperture radar imagery, and UAV aerial imagery; Adaptive spatiotemporal alignment and feature-level fusion are performed on the ground sensor data and the multi-source remote sensing image data to generate a comprehensive soil and water loss perception dataset. Based on the comprehensive soil and water loss perception dataset, soil and water loss risk factors are extracted and a dynamic risk model is constructed. The soil and water loss risk factors include runoff erosivity index, soil erodibility index, vegetation protection index and topography driving index. The dynamic risk model is used to predict the spatiotemporal evolution trend of soil erosion, and a graded early warning signal is generated based on a preset threshold. The tiered early warning signal is pushed to the management terminal, and feedback control instructions are triggered to adjust the monitoring strategy or engineering measures.

2. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 1, characterized in that, Deploying multi-level heterogeneous sensing networks to acquire ground-based sensing data, including: The ground sensing data is collected through the monitoring terminal layer, which includes distributed ground sensing units. Each ground sensing unit integrates a soil moisture sensor, a surface runoff velocity meter, a rainfall meter, a slope inclinometer, and a micro-weather station. The edge computing node layer receives and processes data from the monitoring terminal layer, and the edge computing node layer performs local data caching, preliminary filtering and anomaly detection. A communication relay layer is deployed in the communication obstruction area. The communication relay layer uses frequency hopping spread spectrum technology to dynamically select channels in the 400 MHz to 900 MHz frequency band to achieve reliable multi-hop transmission.

3. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 2, characterized in that, Adaptive spatiotemporal alignment and feature-level fusion are performed on the ground-sensing data and the multi-source remote sensing image data to generate a comprehensive soil and water loss perception dataset, including: Establish a unified spatiotemporal grid with a grid cell size of 50 meters × 50 meters and a time step of 1 hour; The ground sensing data is mapped to the nearest grid cell according to geographical location and aligned to the hour using cubic spline interpolation. The multi-source remote sensing image data is resampled to the unified spatiotemporal grid and fused into a single temporal data using a time-weighted average method; Soil moisture content, surface runoff velocity, and rainfall intensity are used as dynamic input features, while vegetation cover, surface roughness, and slope are used as static background features to construct a multidimensional feature tensor. Each feature in the multidimensional feature tensor is normalized, and the normalization parameter is determined based on the minimum and maximum values ​​of historical three years of data.

4. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 3, characterized in that, Based on the comprehensive soil and water loss perception dataset, soil and water loss risk factors are extracted and a dynamic risk model is constructed, including: The runoff erosivity index is calculated by multiplying and summing the runoff kinetic energy and runoff volume per unit time. The soil erodibility index is determined based on soil texture, organic matter content, and aggregate stability. The vegetation protection index is calculated based on the weighted sum of vegetation cover and leaf area index. The topographic driving index was calculated using the topographic factor method in the general soil loss equation. The runoff erosivity index, soil erodibility index, vegetation protection index, and topographic driving index are weighted and summed to generate a comprehensive risk value. The weighting coefficients are determined by inverting historical disaster samples based on the regional hydrogeological characteristics.

5. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 4, characterized in that, The dynamic risk model predicts the spatiotemporal evolution trend of soil erosion, and generates graded early warning signals based on preset thresholds, including: The comprehensive risk value sequence of the past 24 hours is input into a spatiotemporal convolutional long short-term memory network to predict the risk value for the next 12 hours. The spatiotemporal convolutional long short-term memory network comprises three layers of spatiotemporal convolutional modules and two layers of long short-term memory units. Residual analysis is performed between the predicted results and the measured risk values. If the predicted residual exceeds 15% for three consecutive hours, the online fine-tuning mechanism of the model is triggered. Blue, yellow, orange, or red warning signals are generated based on the range of the comprehensive risk value. Each warning level corresponds to a different risk threshold and response measures.

6. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 5, characterized in that, The tiered early warning signal is pushed to the management terminal, triggering feedback control commands to adjust the monitoring strategy or engineering measures, including: The warning signal is pushed to the mobile application and the command center screen via message queue telemetry transmission protocol; Under a blue alert, increase the frequency of drone inspections; Under a yellow alert, the backup sensor unit is activated and the data reporting frequency is increased to once per minute; Under an orange alert, remotely controlled, liftable sand-blocking fences are deployed, and the local storage acceleration mode of edge nodes is activated. Under a red alert, non-essential power supplies are cut off, with only the BeiDou short message communication module remaining operational to continuously report location and status.

7. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 2, characterized in that, The soil moisture sensor employs the frequency domain reflectance principle, sampling once per minute, with a measurement range of 0% to 50% and an accuracy of ±0.5%. The surface runoff velocity meter is based on the ultrasonic Doppler effect, sampling once every 10 seconds, with a range of 0 to 3 meters per second. The rainfall meter has a tipping bucket structure and a resolution of 0.2 millimeters. The slope inclinometer uses a fusion calculation of a three-axis MEMS accelerometer and a gyroscope, with an update cycle of 1 second. The micro-weather station collects ambient temperature, relative humidity, wind speed, and wind direction, with a data update frequency of once every 30 seconds. All sensor data are appended with a timestamp provided by the BeiDou time synchronization module, with a time synchronization error of less than 10 milliseconds.

8. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 3, characterized in that, The optical satellite imagery has a spatial resolution of 10 meters and a revisit period of 5 days; the synthetic aperture radar imagery uses the C-band, has a spatial resolution of 20 meters, and a revisit period of 6 days; the UAV aerial imagery has a ground sampling distance of 5 centimeters and a flight altitude of 300 meters; all remote sensing images have undergone radiometric correction, geometric correction, and atmospheric correction, and the coordinate system has been converted to the National Geodetic 2000 coordinate system.

9. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 4, characterized in that, The dynamic risk model operates with a 72-hour sliding time window, updating the comprehensive risk value of each grid cell hourly; the weighting coefficients satisfy... + + + =1, where , , The weights correspond to the runoff erosivity index, soil erodibility index, vegetation protection index, and topographic driving index, respectively.

10. The method for dynamic monitoring and early warning of soil erosion in water conservancy projects according to claim 5, characterized in that, The blue alert corresponds to a comprehensive risk value greater than 0.3 and less than or equal to 0.5; the yellow alert corresponds to a risk value greater than 0.5 and less than or equal to 0.7; the orange alert corresponds to a risk value greater than 0.7 and less than or equal to 0.9; and the red alert corresponds to a risk value greater than 0.

9. The warning threshold is dynamically adjusted based on the season, watershed characteristics, and project importance, with the adjustment rules provided by an expert knowledge base.

Images