Intelligent early warning method for monitoring water and soil loss in a drainage basin based on deep learning

Abstract

The application discloses a method for monitoring and early warning of water and soil loss in a drainage basin based on deep learning, which comprises the following steps: S1, a sensor array is arranged in advance in a sensitive area of a to-be-monitored drainage basin, and precipitation data, surface runoff data, suspended matter concentration data and unmanned aerial vehicle multispectral remote sensing data are synchronously collected; S2, fused data streams are input into an improved liquid neural network to generate state characteristic data streams; S3, state characteristic trajectory data are calculated through a neural ordinary differential equation; S4, multi-scale state stability analysis is implemented to identify an abnormal state inflection point; S5, a dynamic risk space mapping atlas is generated; and S6, the risk level of a risk area is determined, and corresponding water and soil conservation measures are executed. The application improves the accuracy of water and soil loss risk prediction and the timeliness of response, and is suitable for monitoring and protection of the ecological environment of a drainage basin.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to an intelligent early warning method for watershed soil and water loss monitoring based on deep learning. Background Technology

[0002] Research on watershed soil and water loss monitoring and early warning technologies is receiving increasing attention. Currently, mainstream watershed soil and water loss monitoring technologies mainly include macroscopic monitoring schemes based on remote sensing technology and microscopic real-time monitoring schemes relying on field sensor networks. Remote sensing monitoring technology uses multispectral remote sensing satellites or UAVs to periodically acquire surface image data, thereby analyzing changes in vegetation cover and exposed soil area, and assessing regional soil and water loss risk trends. Field sensor network monitoring technology, by setting up precipitation monitoring instruments, surface runoff monitoring instruments, and soil erosion sensors within the target area, collects key environmental indicator data in real time, achieving highly real-time local monitoring.

[0003] However, existing remote sensing monitoring methods typically suffer from insufficient timeliness and limited spatial resolution, making it difficult to accurately and promptly identify sudden or exacerbating events of soil erosion. While real-time monitoring schemes based on traditional sensor networks have improved monitoring frequency and accuracy, they often lack efficient analysis methods and intelligent early warning strategies, making it difficult to effectively uncover potential risk characteristics behind the data and proactively identify potential soil erosion risks in advance. Furthermore, traditional data processing methods generally employ static analysis models, lacking effective dynamic processing and prediction capabilities for continuously acquired real-time data streams, making it difficult for the system to quickly and accurately provide risk warnings in practical applications.

[0004] In recent years, with the development of artificial intelligence technology, dynamic deep learning technology, represented by liquid neural networks and neural ordinary differential equations, has been gradually applied to the field of environmental monitoring. It has shown significant advantages in processing continuous dynamic data and real-time risk prediction. However, the research and application of this technology in the field of watershed soil and water loss monitoring is still in the initial exploration stage and lacks systematic and standardized solutions. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent early warning method for watershed soil erosion monitoring based on deep learning. This invention has the advantages of high real-time performance, good prediction accuracy, and the ability to proactively identify the risk of sudden soil erosion in advance.

[0006] The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to embodiments of the present invention includes the following steps: S1. Pre-deploy sensor arrays in sensitive areas of the watershed to be monitored, and simultaneously collect precipitation data, surface runoff data, suspended solids concentration data and UAV multispectral remote sensing data. Perform multidimensional correlation and fusion based on the spatial coordinates and data types of the collected data to obtain a fused data stream. S2. Input the fused data stream into the improved liquid neural network, process the fused data stream through dynamic state space transformation, and generate a state feature data stream characterizing the risk of soil erosion in the watershed; S3. Based on the state feature data stream, a continuous time dynamic model for the state features of watershed soil and water loss risk is constructed through the constant differential equation. The state feature trajectory at multiple preset time scales in the future is calculated using the current state feature data stream as the initial value, and state feature trajectory data with adjustable time scale is obtained. S4. Perform multi-scale state stability analysis on the state characteristic trajectory data, identify abnormal state inflection points in the state characteristic trajectory data that reflect sudden soil erosion events, and generate a real-time abnormal state data set. S5. Establish a dynamic risk spatial mapping map based on the real-time abnormal state data set, and determine the spatial distribution of risk areas for sudden soil erosion events within the watershed; S6. Determine the risk level of each risk area in the watershed based on the dynamic risk space mapping map, and generate and automatically execute the corresponding soil and water conservation measures instructions based on the preset risk level and response measure correspondence.

[0007] Optionally, S1 specifically includes: S11. Determine the spatial layout of the sensor array based on the topography and vegetation distribution of the watershed, and arrange the precipitation sensor, surface runoff sensor and suspended solids concentration sensor in each sensitive area in a spatial grid manner. S12. Within each spatial grid, the precipitation sensor, surface runoff sensor, and suspended solids concentration sensor are located according to the grid center coordinates, and a unique spatial identifier is set for each sensor. S13. Plan the flight path of the UAV multispectral remote sensing system within the spatial deployment area of ​​the sensor array, set several waypoints for the flight path, and set the spatial coordinates and shooting angle of each waypoint. S14. The sensor array and the UAV multispectral remote sensing system are started simultaneously. The sensor array synchronously collects precipitation data, surface runoff data and suspended matter concentration data. The UAV multispectral remote sensing system synchronously captures vegetation coverage data and bare land change data according to the set flight path and waypoints. S15. After data acquisition is completed, the precipitation data, surface runoff data, suspended solids concentration data and UAV multispectral remote sensing data are matched with the spatial identifiers of the corresponding sensors and the spatial coordinates of the UAV waypoints, respectively. S16. Based on the spatial identifiers of the sensor array and the spatial coordinates of the UAV waypoints, and taking the coordinates of the center of the spatial grid as a reference, calculate the three-dimensional Euclidean distance from the spatial identifiers of each sensor and the spatial coordinates of the UAV waypoints to the coordinates of the center of the spatial grid. Filter the data set that meets the fusion conditions according to a preset spatial association distance threshold. Generate spatial position feature vectors for the data set using a multi-dimensional feature vector mapping method based on the coordinates of the center of the spatial grid. Calculate the spatial association strength between each spatial position feature vector using cosine similarity. Perform spatial weighted fusion based on the spatial association strength to generate a fused data stream.

[0008] Optionally, S2 specifically includes: S21. Based on spatial grid units, perform spatiotemporal coupling partitioning on the fused data stream to construct a spatiotemporal coupled data frame sequence that simultaneously possesses temporal and spatial correlation characteristics; S22. The spatiotemporal coupled data frame sequence is sampled at a fixed step size according to the first preset time interval, the second preset time interval, and the third preset time interval, respectively, to obtain data subsequences that reflect short-term fluctuations, medium-term changes, and long-term trends. The three data subsequences are spatially matched according to a unified spatial grid coordinate. The sampling cross-over overlap area between the short-term fluctuation data subsequence and the medium-term change data subsequence, and between the medium-term change data subsequence and the long-term trend data subsequence are calculated in turn. Data frames with consistent spatial grid coordinates in the cross-over overlap area are extracted, time series aligned, and then cross-merged to generate multi-scale data features. S23. The multi-scale data features are respectively input into an improved liquid neural network with multiple parallel dynamic reservoirs. The differentiated initial states and dynamic connection methods of neurons in each reservoir are used to independently process the short-term fluctuation data features, medium-term change data features and long-term trend data features, and generate state evolution trajectories corresponding to short-term fluctuations, medium-term changes and long-term trends respectively. S24. Based on the state evolution trajectory of the reserve pool, extract and record the geometric features formed by each state evolution trajectory in the state space to obtain the state space geometric feature vectors corresponding to short-term fluctuations, medium-term changes and long-term trends. S25. Perform cross-scale fusion processing on the state space geometric feature vectors at different time scales to construct a fused state space feature vector that simultaneously contains information on short-term fluctuations, medium-term changes, and long-term trends, serving as a state feature data stream characterizing the risk of soil erosion in the watershed.

[0009] Optionally, the improved liquid neural network specifically includes an input layer, a scale allocation layer, a parallel dynamic reservoir layer, and a feature output layer: The input layer is used to receive the multi-scale data features and transmit them to the scale allocation layer; The scale allocation layer is used to identify short-term fluctuation data features, medium-term change data features and long-term trend data features in the multi-scale data features, and input them into the parallel dynamic reserve pools of the corresponding scales respectively. The parallel dynamic reserve pool layer includes a short-term reserve pool, a medium-term reserve pool, and a long-term reserve pool. Each reserve pool processes the short-term fluctuation data features, medium-term change data features, and long-term trend data features respectively. The neuron state values ​​in each reserve pool are set with differentiated leakage rate constants, state update step sizes, neuron connection weights, and bias terms according to the corresponding scale features, and the state evolution trajectory of the corresponding scale is generated through a dynamic state update mechanism. The feature output layer is used to extract and record the geometric features formed by the state evolution trajectories of neurons in the short-term, medium-term, and long-term reserve pools in the state space, generate state space geometric feature vectors corresponding to short-term fluctuations, medium-term changes, and long-term trends, and perform cross-scale fusion processing on the state space geometric feature vectors to construct a fused state space feature vector that simultaneously contains information on short-term fluctuations, medium-term changes, and long-term trends, which serves as a state feature data stream characterizing the risk of soil erosion in the watershed.

[0010] Optionally, each reserve pool in the parallel dynamic reserve pool layer executes independently, specifically including: The short-term reserve pool, designed for short-term fluctuation data characteristics, sets the leakage rate constant to the first leakage rate constant, sets the state update step size to the first state update step size, initializes the neuron connection weights to dense random connection weights, and initializes the bias term to the first bias term. The intermediate reserve pool sets the leakage rate constant to the second leakage rate constant, the state update step size to the second state update step size, initializes the neuron connection weights to sparse random connection weights, and initializes the bias term to the second bias term for the characteristics of intermediate change data. The long-term reserve pool, based on the characteristics of long-term trend data, sets the leakage rate constant to the third leakage rate constant, sets the state update step size to the third state update step size, initializes the neuron connection weights to sparse and clustered random connection weights, and initializes the bias term to the third bias term. The neuron state values ​​in the short-term, medium-term, and long-term reserve pools are updated using a dynamic state update mechanism. Through a dynamic state update mechanism, state evolution trajectories reflecting short-term fluctuations, medium-term changes, and long-term trend data characteristics are generated respectively.

[0011] Optionally, S3 specifically includes: S31. Determine the state feature vector of each spatial grid cell in the state feature data stream as the initial state vector of the continuous-time dynamic model at the starting time point; S32. Pre-set a first preset time scale, a second preset time scale and a third preset time scale, and divide the prediction time range into multiple continuous and non-overlapping time intervals based on the preset time scales to obtain multiple continuous time prediction intervals with different prediction lengths. S33. Starting from the initial state vector, construct a continuous-time dynamic model with the state change rate as the core in each continuous-time prediction interval, and describe the continuous-time evolution characteristics of the state feature vector through the constant differential equation. S34. Within each continuous time prediction interval, the neural ordinary differential equation is solved by a high-order adaptive step-size numerical integration method. The state value of the state feature vector of each spatial grid cell at each integration time within the continuous time prediction interval is calculated, and the continuous state evolution trajectory of the state feature vector within each continuous time prediction interval is obtained. S35. Connect the state evolution trajectories in each continuous time prediction interval in sequence to construct a state feature trajectory data sequence and perform numerical smoothing. S36. Based on the first preset time scale, the second preset time scale and the third preset time scale, extract the state feature data corresponding to each preset time scale from the smoothed state feature trajectory data sequence to form state feature trajectory data with adjustable time scale.

[0012] Optionally, S33 specifically includes: S331. Using the initial state vector as input, determine the initial state of the state feature vector within each continuous time prediction interval; S332. Define the continuous change of the state feature vector within the continuous time prediction interval as the derivative of the state feature vector with respect to time, and determine the expression form of the state change rate. S333. Construct a continuous-time neural ordinary differential equation with the state feature vector as the variable, define the rate of state change as a nonlinear dynamic function of the time derivative of the state feature vector, and introduce a historical integral memory term and a real-time differential adjustment term for the state feature vector in the continuous-time prediction interval. S334. Based on the state feature trajectory data within the historical continuous time prediction interval, the adaptive gradient descent algorithm is used to iteratively optimize the parameter values ​​of the weight matrix, input weight matrix, bias vector, historical integral memory term weight matrix, and real-time differential adjustment term weight matrix. S335. Substitute the optimized weight matrix, input weight matrix, bias vector, historical integral memory term weight matrix, and real-time difference adjustment term weight matrix into the continuous-time neural ordinary differential equation to obtain a continuous-time dynamic model that can reflect the dynamic evolution characteristics of the state feature vector as it changes over continuous time.

[0013] Optionally, S4 specifically includes: S41. Using state feature trajectory data as input, perform sequence segmentation processing according to the first preset time scale, the second preset time scale and the third preset time scale respectively to generate a multi-scale segmented state sequence. S42. For each scale of segmented state sequence, calculate the statistical mean and standard deviation of the state feature vector values ​​segment by segment, and determine the quantitative index of state stability by the degree to which the state feature vector values ​​deviate from the statistical mean. S43. Based on the changes in the state stability quantification index, calculate the difference change between adjacent state feature vectors in the state sequence of each scale segment, and determine the significance threshold of the difference change. S44. Based on the determined significance threshold, compare the adjacent difference changes in the segmented state sequence of each scale one by one. When the difference change exceeds the significance threshold, mark the corresponding state feature vector as a candidate abnormal state inflection point. S45. Perform cross-scale verification on the candidate abnormal state inflection points under the first preset time scale, the second preset time scale, and the third preset time scale respectively, determine the positions that are marked as abnormal state inflection points under multiple scales, and record them as the final abnormal state inflection points. S46. Based on the determined inflection points of abnormal states, sort and deduplicate the state feature vectors according to their positions on the original time axis to generate a real-time abnormal state data set.

[0014] Optionally, S5 specifically includes: S51. Map each abnormal state feature vector in the real-time abnormal state data set to the corresponding spatial grid cell to determine the spatial coordinate position of each abnormal state feature vector. S52. Construct a spatial anomaly weight matrix based on the anomaly degree quantification index according to the anomaly degree of each anomaly state feature vector and the spatial coordinate position of the corresponding spatial grid cell. S53. Using the spatial anomaly weight matrix as input, the spatial interpolation algorithm is used to calculate the anomaly propagation impact value of adjacent spatial grid cells that have not experienced anomalies step by step, and an anomaly propagation threshold is defined. S54. Based on the relationship between the impact value of abnormal propagation and the threshold value of abnormal propagation, determine the specific location of the spatial grid unit that is significantly affected by the abnormality, and determine the boundary range of the risk area of ​​sudden soil erosion event; S55. Based on the spatial grid boundary range of the risk area, calculate the cumulative sum of the abnormal propagation impact values ​​of all spatial grid units in the area to form a dynamic risk spatial distribution intensity map. S56. Update the dynamic risk spatial distribution intensity map in real time, and determine the specific spatial distribution location of the risk area of ​​sudden soil erosion event within the watershed space based on the updated risk spatial distribution intensity map.

[0015] Optionally, S6 specifically includes: S61. Based on the risk intensity and spatial correlation characteristics of each spatial grid unit in the dynamic risk spatial mapping map, generate risk level clustering relationships between spatial grid units; S62. Based on spatial risk level clustering relationships, identify and dynamically adjust the spatial extent of watershed risk areas, and automatically delineate initial risk areas, potential expansion areas, and risk propagation paths; S63. Based on the spatial range of the risk area, and combined with the continuity and changing trend of the spatial distribution of risk intensity, the risk level of the spatial grid unit is updated in real time to form a real-time dynamic risk level classification result. S64. Based on the real-time dynamic risk level classification results, each risk area is divided into stable area, diffusion area and drastic change area in real time, and a corresponding risk evolution model is automatically established. S65. Establish in advance a command system for soil and water conservation measures corresponding to stable zones, diffusion zones and drastic change zones respectively. The stable zone corresponds to the command for strengthening the monitoring of soil and water conservation measures, the diffusion zone corresponds to the command for early reinforcement of soil and water conservation facilities, and the drastic change zone corresponds to the command for emergency response mechanism and UAV image acquisition. S66. Based on the real-time dynamic risk level and corresponding risk evolution model of each risk area, automatically select and trigger the water and soil conservation measures instruction system for the corresponding area, and automatically execute the corresponding water and soil conservation measures in real time.

[0016] The beneficial effects of this invention are: (1) By deploying sensor arrays in sensitive areas of the watershed and integrating UAV multispectral remote sensing data, the present invention can acquire precipitation data, surface runoff data and suspended matter concentration data in real time, which significantly improves the real-time performance and integration of monitoring data, effectively improves the spatial accuracy and data timeliness of watershed soil erosion risk monitoring, and enhances the response speed to sudden or aggravated soil erosion events.

[0017] (2) This invention uses an improved liquid neural network combined with the real-time dynamic analysis of the neural ordinary differential equation to fuse data streams, which can achieve efficient identification and continuous time prediction of watershed soil erosion risk status characteristics, significantly improve the accuracy and real-time performance of soil erosion risk prediction, and show better adaptability in the dynamic monitoring and risk warning scenarios of watershed ecological environment.

[0018] (3) In terms of real-time risk status analysis, this invention effectively solves the problems of lack of proactive early warning capability and delayed risk identification in the existing technology through multi-scale state stability analysis and real-time abnormal state inflection point identification. It breaks through the limitation of traditional methods that cannot accurately identify risk change trends in advance, realizes proactive and accurate identification of risk status, and thus effectively improves the intelligent early warning capability of watershed soil erosion.

[0019] (4) By constructing a dynamic risk space mapping map and dividing the watershed risk areas in real time, this invention establishes a regional risk evolution model and dynamically implements corresponding soil and water conservation measures, which significantly improves the pertinence and efficiency of early warning measures and enhances the response capability and governance effect of watershed ecological environment protection. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall process of the intelligent early warning method for watershed soil and water loss monitoring based on deep learning proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 A deep learning-based intelligent early warning method for watershed soil erosion monitoring includes the following steps: S1. Pre-deploy sensor arrays in sensitive areas of the watershed to be monitored, and simultaneously collect precipitation data, surface runoff data, suspended solids concentration data and UAV multispectral remote sensing data. Perform multidimensional correlation and fusion based on the spatial coordinates and data types of the collected data to obtain a fused data stream. S2. Input the fused data stream into the improved liquid neural network, process the fused data stream through dynamic state space transformation, and generate a state feature data stream characterizing the risk of soil erosion in the watershed; S3. Based on the state feature data stream, a continuous time dynamic model for the state features of watershed soil and water loss risk is constructed through the constant differential equation. The state feature trajectory at multiple preset time scales in the future is calculated using the current state feature data stream as the initial value, and state feature trajectory data with adjustable time scale is obtained. S4. Perform multi-scale state stability analysis on the state characteristic trajectory data, identify abnormal state inflection points in the state characteristic trajectory data that reflect sudden soil erosion events, and generate a real-time abnormal state data set. S5. Establish a dynamic risk spatial mapping map based on the real-time abnormal state data set, and determine the spatial distribution of risk areas for sudden soil erosion events within the watershed; S6. Determine the risk level of each risk area in the watershed based on the dynamic risk space mapping map, and generate and automatically execute the corresponding soil and water conservation measures instructions based on the preset risk level and response measure correspondence.

[0023] In this embodiment, S1 specifically includes: S11. Determine the spatial layout of the sensor array based on the topography and vegetation distribution of the watershed, and arrange the precipitation sensor, surface runoff sensor and suspended solids concentration sensor in each sensitive area in a spatial grid manner. S12. Within each spatial grid, the precipitation sensor, surface runoff sensor, and suspended solids concentration sensor are located according to the grid center coordinates, and a unique spatial identifier is set for each sensor. S13. Plan the flight path of the UAV multispectral remote sensing system within the spatial deployment area of ​​the sensor array, set several waypoints for the flight path, and set the spatial coordinates and shooting angle of each waypoint. S14. The sensor array and the UAV multispectral remote sensing system are started simultaneously. The sensor array synchronously collects precipitation data, surface runoff data and suspended matter concentration data. The UAV multispectral remote sensing system synchronously captures vegetation coverage data and bare land change data according to the set flight path and waypoints. S15. After data acquisition is completed, the precipitation data, surface runoff data, suspended solids concentration data and UAV multispectral remote sensing data are matched with the spatial identifiers of the corresponding sensors and the spatial coordinates of the UAV waypoints, respectively. S16. Based on the spatial identifiers of the sensor array and the spatial coordinates of the UAV waypoints, and taking the coordinates of the center of the spatial grid as a reference, calculate the three-dimensional Euclidean distance from the spatial identifiers of each sensor and the spatial coordinates of the UAV waypoints to the coordinates of the center of the spatial grid. Filter the data set that meets the fusion conditions according to a preset spatial association distance threshold. Generate spatial position feature vectors for the data set using a multi-dimensional feature vector mapping method based on the coordinates of the center of the spatial grid. Calculate the spatial association strength between each spatial position feature vector using cosine similarity. Perform spatial weighted fusion based on the spatial association strength to generate a fused data stream.

[0024] In this embodiment, S2 specifically includes: S21. Based on spatial grid units, perform spatiotemporal coupling partitioning on the fused data stream to construct a spatiotemporal coupled data frame sequence that simultaneously possesses temporal and spatial correlation characteristics; S22. The spatiotemporal coupled data frame sequence is sampled at a fixed step size according to the first preset time interval, the second preset time interval, and the third preset time interval, respectively, to obtain data subsequences that reflect short-term fluctuations, medium-term changes, and long-term trends. The three data subsequences are spatially matched according to a unified spatial grid coordinate. The sampling cross-over overlap area between the short-term fluctuation data subsequence and the medium-term change data subsequence, and between the medium-term change data subsequence and the long-term trend data subsequence are calculated in turn. Data frames with consistent spatial grid coordinates in the cross-over overlap area are extracted, time series aligned, and then cross-merged to generate multi-scale data features. S23. The multi-scale data features are respectively input into an improved liquid neural network with multiple parallel dynamic reservoirs. The differentiated initial states and dynamic connection methods of neurons in each reservoir are used to independently process the short-term fluctuation data features, medium-term change data features and long-term trend data features, and generate state evolution trajectories corresponding to short-term fluctuations, medium-term changes and long-term trends respectively. S24. Based on the state evolution trajectory of the reserve pool, extract and record the geometric features formed by each state evolution trajectory in the state space to obtain the state space geometric feature vectors corresponding to short-term fluctuations, medium-term changes and long-term trends. S25. Perform cross-scale fusion processing on the state space geometric feature vectors at different time scales to construct a fused state space feature vector that simultaneously contains information on short-term fluctuations, medium-term changes, and long-term trends, serving as a state feature data stream characterizing the risk of soil erosion in the watershed.

[0025] In this embodiment, the improved liquid neural network specifically includes an input layer, a scale allocation layer, a parallel dynamic reservoir layer, and a feature output layer: The input layer is used to receive the multi-scale data features and transmit them to the scale allocation layer; The scale allocation layer is used to identify short-term fluctuation data features, medium-term change data features and long-term trend data features in the multi-scale data features, and input them into the parallel dynamic reserve pools of the corresponding scales respectively. The parallel dynamic reserve pool layer includes a short-term reserve pool, a medium-term reserve pool, and a long-term reserve pool. Each reserve pool processes the short-term fluctuation data features, medium-term change data features, and long-term trend data features respectively. The neuron state values ​​in each reserve pool are set with differentiated leakage rate constants, state update step sizes, neuron connection weights, and bias terms according to the corresponding scale features, and the state evolution trajectory of the corresponding scale is generated through a dynamic state update mechanism. The feature output layer is used to extract and record the geometric features formed by the state evolution trajectories of neurons in the short-term, medium-term, and long-term reserve pools in the state space, generate state space geometric feature vectors corresponding to short-term fluctuations, medium-term changes, and long-term trends, and perform cross-scale fusion processing on the state space geometric feature vectors to construct a fused state space feature vector that simultaneously contains information on short-term fluctuations, medium-term changes, and long-term trends, which serves as a state feature data stream characterizing the risk of soil erosion in the watershed.

[0026] In this embodiment, each reserve pool in the parallel dynamic reserve pool layer executes independently, specifically including: The short-term reserve pool, designed for short-term fluctuation data characteristics, sets the leakage rate constant to the first leakage rate constant, sets the state update step size to the first state update step size, initializes the neuron connection weights to dense random connection weights, and initializes the bias term to the first bias term. The intermediate reserve pool sets the leakage rate constant to the second leakage rate constant, the state update step size to the second state update step size, initializes the neuron connection weights to sparse random connection weights, and initializes the bias term to the second bias term for the characteristics of intermediate change data. The long-term reserve pool, based on the characteristics of long-term trend data, sets the leakage rate constant to the third leakage rate constant, sets the state update step size to the third state update step size, initializes the neuron connection weights to sparse and clustered random connection weights, and initializes the bias term to the third bias term. The neuron state values ​​in the short-term, medium-term, and long-term reserve pools are updated using the following dynamic state update mechanisms:

[0027] ; in, For the first A neuron at time 1 The state value, For state update step size, This represents the leakage rate constant for the corresponding scale. For the first in the reserve pool The first neuron to the second The connection weights of each neuron For the first The input feature to the first The input connection weights of each neuron. For the first The input features at time... eigenvalues, For the first Bias terms for each neuron, It is the hyperbolic tangent activation function; The scale-aware adaptive dynamic feedback term for: ; in, For the first The adaptive adjustment coefficients of each neuron for the input scale. For the first The specific historical retrieval interval of the scale feature targeted by each neuron The dimension of the input features; The formula introduces a scale-aware adaptive dynamic feedback term during the neuron state update process. It dynamically adjusts the state update amount by utilizing the change amplitude of input feature values ​​within the historical backtracking interval. The leakage rate constant and the adaptive adjustment coefficient work together to achieve differentiated responses to short-term, medium-term, and long-term data features. This effectively improves the sensitivity and adaptability of the neuron state value update process, enabling the network to more accurately capture and reflect changes in input features at different time scales, thereby improving the accuracy of state feature trajectory generation.

[0028] Through a dynamic state update mechanism, state evolution trajectories reflecting short-term fluctuations, medium-term changes, and long-term trend data characteristics are generated respectively.

[0029] In this embodiment, S3 specifically includes: S31. Determine the state feature vector of each spatial grid cell in the state feature data stream as the initial state vector of the continuous-time dynamic model at the starting time point; S32. Pre-set a first preset time scale, a second preset time scale and a third preset time scale, and divide the prediction time range into multiple continuous and non-overlapping time intervals based on the preset time scales to obtain multiple continuous time prediction intervals with different prediction lengths. S33. Starting from the initial state vector, construct a continuous-time dynamic model with the state change rate as the core in each continuous-time prediction interval, and describe the continuous-time evolution characteristics of the state feature vector through the constant differential equation. S34. Within each continuous time prediction interval, the neural ordinary differential equation is solved by a high-order adaptive step-size numerical integration method. The state value of the state feature vector of each spatial grid cell at each integration time within the continuous time prediction interval is calculated, and the continuous state evolution trajectory of the state feature vector within each continuous time prediction interval is obtained. S35. Connect the state evolution trajectories in each continuous time prediction interval in sequence to construct a state feature trajectory data sequence and perform numerical smoothing. S36. Based on the first preset time scale, the second preset time scale and the third preset time scale, extract the state feature data corresponding to each preset time scale from the smoothed state feature trajectory data sequence to form state feature trajectory data with adjustable time scale.

[0030] In this embodiment, S33 specifically includes: S331. Using the initial state vector as input, determine the initial state of the state feature vector within each continuous time prediction interval; S332. Define the continuous change of the state feature vector within the continuous time prediction interval as the derivative of the state feature vector with respect to time, and determine the expression form of the state change rate. S333. Construct a continuous-time neural ordinary differential equation with the state eigenvector as the variable, define the rate of state change as a nonlinear dynamic function of the time derivative of the state eigenvector, and introduce a historical integral memory term and a real-time differential adjustment term for the state eigenvector within the continuous-time prediction interval: ; in, Represents any time The state feature vector has a dimension equal to the state feature dimension, which is the same as the dimension of the initial state vector. Represents any time The input external feature vector has the same dimension as the state feature vector. This is the weight matrix for the interactions between the state feature vectors, with dimensions equal to the product of the state feature dimensions. The weight matrix is ​​the input of the external feature vector to the state feature vector, and its dimension is the state feature dimension multiplied by the state feature dimension. This is the bias vector of the state feature vector, with the same dimension as the state feature vector. This is the weight matrix for the historical integral memory terms, with dimensions equal to the product of the state feature dimensions. The length of the historical time window for the integral memory term. This is the weight matrix for the real-time differential adjustment term, with dimensions equal to the product of the state feature dimensions. The time interval for the real-time differential adjustment term, the function It is the hyperbolic tangent activation function; The formula defines the derivative of the state feature vector as a nonlinear dynamic function and introduces a historical integral memory term and a real-time difference adjustment term. It dynamically adjusts the evolution rate of the state feature vector by accumulating the integral of historical states and the difference magnitude of real-time state changes. This achieves a dual sensitive response of the state feature vector to historical trends and current real-time changes, thereby effectively improving the stability and accuracy of the state feature vector prediction over continuous time and enhancing the model's ability to accurately capture the dynamic trend of soil erosion risk.

[0031] S334. Based on the state feature trajectory data within the historical continuous time prediction interval, the adaptive gradient descent algorithm is used to iteratively optimize the parameter values ​​of the weight matrix, input weight matrix, bias vector, historical integral memory term weight matrix, and real-time differential adjustment term weight matrix. S335. Substitute the optimized weight matrix, input weight matrix, bias vector, historical integral memory term weight matrix, and real-time difference adjustment term weight matrix into the continuous-time neural ordinary differential equation to obtain a continuous-time dynamic model that can reflect the dynamic evolution characteristics of the state feature vector as it changes over continuous time.

[0032] In this embodiment, S4 specifically includes: S41. Using state feature trajectory data as input, perform sequence segmentation processing according to the first preset time scale, the second preset time scale and the third preset time scale respectively to generate a multi-scale segmented state sequence. S42. For each scale of segmented state sequence, calculate the statistical mean and standard deviation of the state feature vector values ​​segment by segment, and determine the quantitative index of state stability by the degree to which the state feature vector values ​​deviate from the statistical mean. S43. Based on the changes in the state stability quantification index, calculate the difference change between adjacent state feature vectors in the state sequence of each scale segment, and determine the significance threshold of the difference change. S44. Based on the determined significance threshold, compare the adjacent difference changes in the segmented state sequence of each scale one by one. When the difference change exceeds the significance threshold, mark the corresponding state feature vector as a candidate abnormal state inflection point. S45. Perform cross-scale verification on the candidate abnormal state inflection points under the first preset time scale, the second preset time scale, and the third preset time scale respectively, determine the positions that are marked as abnormal state inflection points under multiple scales, and record them as the final abnormal state inflection points. S46. Based on the determined inflection points of abnormal states, sort and deduplicate the state feature vectors according to their positions on the original time axis to generate a real-time abnormal state data set.

[0033] In this embodiment, S5 specifically includes: S51. Map each abnormal state feature vector in the real-time abnormal state data set to the corresponding spatial grid cell to determine the spatial coordinate position of each abnormal state feature vector. S52. Construct a spatial anomaly weight matrix based on the anomaly degree quantification index according to the anomaly degree of each anomaly state feature vector and the spatial coordinate position of the corresponding spatial grid cell. S53. Using the spatial anomaly weight matrix as input, the spatial interpolation algorithm is used to calculate the anomaly propagation impact value of adjacent spatial grid cells that have not experienced anomalies step by step, and an anomaly propagation threshold is defined. S54. Based on the relationship between the impact value of abnormal propagation and the threshold value of abnormal propagation, determine the specific location of the spatial grid unit that is significantly affected by the abnormality, and determine the boundary range of the risk area of ​​sudden soil erosion event; S55. Based on the spatial grid boundary range of the risk area, calculate the cumulative sum of the abnormal propagation impact values ​​of all spatial grid units in the area to form a dynamic risk spatial distribution intensity map. S56. Update the dynamic risk spatial distribution intensity map in real time, and determine the specific spatial distribution location of the risk area of ​​sudden soil erosion event within the watershed space based on the updated risk spatial distribution intensity map.

[0034] In this embodiment, S6 specifically includes: S61. Based on the risk intensity and spatial correlation characteristics of each spatial grid unit in the dynamic risk spatial mapping map, generate risk level clustering relationships between spatial grid units; S62. Based on spatial risk level clustering relationships, identify and dynamically adjust the spatial extent of watershed risk areas, and automatically delineate initial risk areas, potential expansion areas, and risk propagation paths; S63. Based on the spatial range of the risk area, and combined with the continuity and changing trend of the spatial distribution of risk intensity, the risk level of the spatial grid unit is updated in real time to form a real-time dynamic risk level classification result. S64. Based on the real-time dynamic risk level classification results, each risk area is divided into stable area, diffusion area and drastic change area in real time, and a corresponding risk evolution model is automatically established. S65. Establish in advance a command system for soil and water conservation measures corresponding to stable zones, diffusion zones and drastic change zones respectively. The stable zone corresponds to the command for strengthening the monitoring of soil and water conservation measures, the diffusion zone corresponds to the command for early reinforcement of soil and water conservation facilities, and the drastic change zone corresponds to the command for emergency response mechanism and UAV image acquisition. S66. Based on the real-time dynamic risk level and corresponding risk evolution model of each risk area, automatically select and trigger the water and soil conservation measures instruction system for the corresponding area, and automatically execute the corresponding water and soil conservation measures in real time.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a watershed soil erosion monitoring and early warning scenario in a certain region. Real-time soil erosion risk monitoring and early warning analysis were conducted on a key watershed within the region. This watershed covers a vast area with complex terrain and diverse vegetation cover. In recent years, frequent torrential rains have exacerbated the soil erosion risk, severely impacting the local ecological environment. In this application scenario, traditional methods mainly rely on regular manual inspections and satellite remote sensing image analysis for monitoring. The inspection cycle is generally 7 to 10 days. Remote sensing image analysis has low spatial resolution and an image update frequency of 10 days, making it difficult to achieve high-frequency, real-time, and accurate risk identification and early warning. This results in an inability to respond promptly to sudden or exacerbated soil erosion events, leading to a significant lag in monitoring effectiveness.

[0036] In practical applications, the watershed is first divided into spatial grids based on the topographic features and vegetation distribution of different areas, with each grid unit covering 100 square meters. High-precision sensor arrays, including precipitation sensors, surface runoff sensors, and suspended solids concentration sensors, are deployed within the watershed. Each sensor is precisely located according to the center of the spatial grid, and a unique spatial identifier is assigned to each sensor unit within the array. Furthermore, flight routes for a multispectral remote sensing system using unmanned aerial vehicles (UAVs) are planned within the sensor array's coverage area. Each UAV flies at an altitude of 120 meters, and spatial waypoints are set along the flight route. The coordinates and shooting angle information of each waypoint are clearly recorded, enabling high-frequency monitoring of vegetation cover and bare land changes (per hour).

[0037] During data acquisition, the sensor array and the UAV system are activated synchronously to collect real-time data on precipitation, surface runoff, suspended solids concentration, and vegetation cover in each grid cell. The acquired data is then precisely matched with the spatial identifiers of the corresponding sensors and the spatial coordinates of the UAV waypoints. Using the grid cell spatial coordinates as a reference, the three-dimensional Euclidean distance from each sensor and UAV waypoint coordinate to the grid center is calculated. Based on the cosine similarity algorithm, the spatial data is accurately fused to generate a fused data stream.

[0038] Subsequently, this embodiment utilizes an improved liquid neural network to further analyze the fused data stream. First, the fused data stream is spatiotemporally coupled and partitioned to construct a data frame sequence that simultaneously contains time-series features and spatial correlation features. The sequence is then sampled at multiple scales at time intervals of 5 minutes, 30 minutes, and 1 hour to obtain data subsequences representing short-term fluctuations, medium-term changes, and long-term trends. Then, the data features at different scales are input into short-term, medium-term, and long-term reserve pools, respectively. By setting different scale leakage rates and state update step sizes, and employing a scale-aware adaptive dynamic feedback term for dynamic state updates, accurate generation of state feature trajectory data is achieved.

[0039] In the risk prediction stage, the system utilizes neural ordinary differential equations to perform continuous-time modeling and predictive analysis of state characteristic trajectory data. By constructing a continuous-time dynamic model with the state change rate as the core, historical integral memory terms and real-time difference adjustment terms are introduced to improve the model's predictive stability. After 28 rounds of iterative optimization, the weight matrices and bias vectors are optimized, and the model's prediction error on the validation set is reduced to 0.029, achieving good prediction accuracy.

[0040] In practical monitoring applications, the system identifies anomalous state inflection points based on the generated state characteristic trajectory data through multi-scale state stability analysis. During a heavy rainfall event, the system successfully identified 17 potential anomalous state inflection points within two hours and quickly generated a real-time anomalous state data set. Subsequently, based on this anomalous data set, a dynamic risk spatial mapping map was established. The impact value of anomalous propagation was calculated using a spatial interpolation algorithm, accurately delineating the boundaries of two key risk areas and potential diffusion paths, thus achieving real-time dynamic risk level classification.

[0041] Finally, based on the risk level clustering relationships, this embodiment identifies stable zones, diffusion zones, and drastic change zones, and establishes corresponding regional risk evolution models. In the first three hours of the continuous heavy rainfall, the system automatically executed 19 enhanced monitoring commands for stable zones, 7 commands for advance reinforcement measures in diffusion zones, and 2 commands for emergency response and UAV-based image acquisition in drastic change zones. The system's timely response ensured that no significant soil erosion occurred in the drastic change zones or diffusion zones, and vegetation coverage and soil erosion concentration indicators were significantly better than previous monitoring data.

[0042] The table below shows a comparison between the predicted and actual monitoring data of soil erosion indicators during a rainstorm event in this embodiment.

[0043] Table 1 Comparison of Predicted and Measured Soil Erosion Indicators During Rainstorms

[0044] As can be seen from the data in Table 1 above, the system of the present invention has high prediction accuracy for suspended solids concentration, surface runoff, and vegetation coverage. The average error between the measured and predicted values ​​is within 5%. Specifically, the prediction error for suspended solids concentration in grid cell A012 is 3.89%, the prediction error for surface runoff is 2.52%, and the prediction error for vegetation coverage is 0.67%. The prediction results demonstrate the reliability and stability of the system in responding to sudden soil erosion events.

[0045] This embodiment, through detailed data collection, model optimization, and verification of practical application effects, fully demonstrates that the intelligent early warning method of the present invention can effectively overcome the problems of insufficient timeliness, low accuracy, and delayed prediction in traditional soil erosion monitoring technologies, and has obvious technical advantages and application value in the field of monitoring and early warning of sudden soil erosion risks.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent early warning method for watershed soil erosion monitoring, characterized in that, Includes the following steps: S1. Pre-deploy sensor arrays in sensitive areas of the watershed to be monitored, and simultaneously collect precipitation data, surface runoff data, suspended solids concentration data and UAV multispectral remote sensing data. Perform multidimensional correlation and fusion based on the spatial coordinates and data types of the collected data to obtain a fused data stream. S2. Input the fused data stream into the improved liquid neural network, process the fused data stream through dynamic state space transformation, and generate a state feature data stream characterizing the risk of soil erosion in the watershed, specifically including: S21. Based on spatial grid units, perform spatiotemporal coupling partitioning on the fused data stream to construct a spatiotemporal coupled data frame sequence that simultaneously possesses temporal and spatial correlation characteristics; S22. The spatiotemporal coupled data frame sequence is sampled at a fixed step size according to the first preset time interval, the second preset time interval, and the third preset time interval, respectively, to obtain data subsequences that reflect short-term fluctuations, medium-term changes, and long-term trends. The three data subsequences are spatially matched according to a unified spatial grid coordinate. The sampling cross-over overlap area between the short-term fluctuation data subsequence and the medium-term change data subsequence, and between the medium-term change data subsequence and the long-term trend data subsequence are calculated in sequence. Data frames with consistent spatial grid coordinates in the cross-over overlap area are extracted, time series aligned, and then cross-merged to generate multi-scale data features. S23. Input the short-term fluctuation data features, medium-term change data features and long-term trend data features in the multi-scale data features into the improved liquid neural network with multiple parallel dynamic reservoirs respectively. Utilize the differentiated initial states and dynamic connection methods of the neurons in each reservoir to process the short-term fluctuation data features, medium-term change data features and long-term trend data features independently, and generate state evolution trajectories corresponding to short-term fluctuations, medium-term changes and long-term trends respectively. S24. Based on the state evolution trajectory of the reserve pool, extract and record the geometric features formed by each state evolution trajectory in the state space to obtain the state space geometric feature vectors corresponding to short-term fluctuations, medium-term changes and long-term trends. S25. Perform cross-scale fusion processing on the state space geometric feature vectors at different time scales to construct a fused state space feature vector that simultaneously contains short-term fluctuations, medium-term changes and long-term trend information, as a state feature data stream characterizing the risk of soil erosion in the watershed. S3. Based on the state feature data stream, a continuous time dynamic model for the state features of watershed soil and water loss risk is constructed through the constant differential equation. The state feature trajectory at multiple preset time scales in the future is calculated using the current state feature data stream as the initial value, and state feature trajectory data with adjustable time scale is obtained. S4. Perform multi-scale state stability analysis on the state characteristic trajectory data to identify anomalous state inflection points reflecting sudden soil erosion events in the state characteristic trajectory data, and generate a real-time anomalous state data set, specifically including: S41. Using state feature trajectory data as input, perform sequence segmentation processing according to the first preset time scale, the second preset time scale and the third preset time scale respectively to generate a multi-scale segmented state sequence. S42. For each scale of segmented state sequence, calculate the statistical mean and standard deviation of the state feature vector values ​​segment by segment, and determine the quantitative index of state stability by the degree to which the state feature vector values ​​deviate from the statistical mean. S43. Based on the changes in the state stability quantification index, calculate the difference change between adjacent state feature vectors in the state sequence of each scale segment, and determine the significance threshold of the difference change. S44. Based on the determined significance threshold, compare the adjacent difference changes in the segmented state sequence of each scale one by one. When the difference change exceeds the significance threshold, mark the corresponding state feature vector as a candidate abnormal state inflection point. S45. Perform cross-scale verification on the candidate abnormal state inflection points under the first preset time scale, the second preset time scale, and the third preset time scale respectively, determine the positions that are marked as abnormal state inflection points under multiple scales, and record them as the final abnormal state inflection points. S46. Based on the determined inflection points of abnormal states, sort and deduplicate the state feature vectors according to their positions on the original time axis to generate a real-time abnormal state data set. S5. Establish a dynamic risk spatial mapping map based on the real-time abnormal state data set, and determine the spatial distribution of risk areas for sudden soil erosion events within the watershed; S6. Determine the risk level of each risk area in the watershed based on the dynamic risk space mapping map, and generate and automatically execute the corresponding soil and water conservation measures instructions based on the preset risk level and response measure correspondence.

2. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Determine the spatial layout of the sensor array based on the topography and vegetation distribution of the watershed, and arrange the precipitation sensor, surface runoff sensor and suspended solids concentration sensor in each sensitive area in a spatial grid manner. S12. Within each spatial grid, the precipitation sensor, surface runoff sensor, and suspended solids concentration sensor are located according to the grid center coordinates, and a unique spatial identifier is set for each sensor. S13. Plan the flight path of the UAV multispectral remote sensing system within the spatial deployment area of ​​the sensor array, set several waypoints for the flight path, and set the spatial coordinates and shooting angle of each waypoint. S14. The sensor array and the UAV multispectral remote sensing system are started simultaneously. The sensor array synchronously collects precipitation data, surface runoff data and suspended matter concentration data. The UAV multispectral remote sensing system synchronously captures vegetation coverage data and bare land change data according to the set flight path and waypoints. S15. After data acquisition is completed, the precipitation data, surface runoff data, suspended solids concentration data and UAV multispectral remote sensing data are matched with the spatial identifiers of the corresponding sensors and the spatial coordinates of the UAV waypoints, respectively. S16. Based on the spatial identifiers of the sensor array and the spatial coordinates of the UAV waypoints, and taking the coordinates of the center of the spatial grid as a reference, calculate the three-dimensional Euclidean distance from the spatial identifiers of each sensor and the spatial coordinates of the UAV waypoints to the coordinates of the center of the spatial grid. Filter the data set that meets the fusion conditions according to a preset spatial association distance threshold. Generate spatial position feature vectors for the data set using a multi-dimensional feature vector mapping method based on the coordinates of the center of the spatial grid. Calculate the spatial association strength between each spatial position feature vector using cosine similarity. Perform spatial weighted fusion based on the spatial association strength to generate a fused data stream.

3. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 1, characterized in that, The improved liquid neural network specifically includes an input layer, a scale allocation layer, a parallel dynamic reservoir layer, and a feature output layer: The input layer is used to receive the multi-scale data features and transmit them to the scale allocation layer; The scale allocation layer is used to identify short-term fluctuation data features, medium-term change data features and long-term trend data features in the multi-scale data features, and input them into the parallel dynamic reserve pools of the corresponding scales respectively. The parallel dynamic reserve pool layer includes a short-term reserve pool, a medium-term reserve pool, and a long-term reserve pool. Each reserve pool processes the short-term fluctuation data features, medium-term change data features, and long-term trend data features respectively. The neuron state values ​​in each reserve pool are set with differentiated leakage rate constants, state update step sizes, neuron connection weights, and bias terms according to the corresponding scale features, and the state evolution trajectory of the corresponding scale is generated through a dynamic state update mechanism. The feature output layer is used to extract and record the geometric features formed by the state evolution trajectories of neurons in the short-term, medium-term, and long-term reserve pools in the state space, generate state space geometric feature vectors corresponding to short-term fluctuations, medium-term changes, and long-term trends, and perform cross-scale fusion processing on the state space geometric feature vectors to construct a fused state space feature vector that simultaneously contains information on short-term fluctuations, medium-term changes, and long-term trends, which serves as a state feature data stream characterizing the risk of soil erosion in the watershed.

4. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 3, characterized in that, Each reserve pool in the parallel dynamic reserve pool layer executes independently, specifically including: The short-term reserve pool, designed for short-term fluctuation data characteristics, sets the leakage rate constant to the first leakage rate constant, sets the state update step size to the first state update step size, initializes the neuron connection weights to dense random connection weights, and initializes the bias term to the first bias term. The intermediate reserve pool sets the leakage rate constant to the second leakage rate constant, the state update step size to the second state update step size, initializes the neuron connection weights to sparse random connection weights, and initializes the bias term to the second bias term for the characteristics of intermediate change data. The long-term reserve pool, based on the characteristics of long-term trend data, sets the leakage rate constant to the third leakage rate constant, sets the state update step size to the third state update step size, initializes the neuron connection weights to sparse and clustered random connection weights, and initializes the bias term to the third bias term. The neuron state values ​​in the short-term, medium-term, and long-term reserve pools are updated using a dynamic state update mechanism. Through a dynamic state update mechanism, state evolution trajectories reflecting short-term fluctuations, medium-term changes, and long-term trend data characteristics are generated respectively.

5. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Determine the state feature vector of each spatial grid cell in the state feature data stream as the initial state vector of the continuous-time dynamic model at the starting time point; S32. Pre-set a first preset time scale, a second preset time scale and a third preset time scale, and divide the prediction time range into multiple continuous and non-overlapping time intervals based on the preset time scales to obtain multiple continuous time prediction intervals. S33. Starting from the initial state vector, construct a continuous-time dynamic model with the state change rate as the core in each continuous-time prediction interval, and describe the continuous-time evolution characteristics of the state feature vector through the constant differential equation. S34. Within each continuous time prediction interval, the neural ordinary differential equation is solved by a high-order adaptive step-size numerical integration method. The state value of the state feature vector of each spatial grid cell at each integration time within the continuous time prediction interval is calculated, and the continuous state evolution trajectory of the state feature vector within each continuous time prediction interval is obtained. S35. Connect the state evolution trajectories in each continuous time prediction interval in sequence to construct a state feature trajectory data sequence and perform numerical smoothing. S36. Based on the first preset time scale, the second preset time scale and the third preset time scale, extract the state feature data corresponding to each preset time scale from the smoothed state feature trajectory data sequence to form state feature trajectory data with adjustable time scale.

6. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 5, characterized in that, Specifically, S33 includes: S331. Using the initial state vector as input, determine the initial state of the state feature vector within each continuous time prediction interval; S332. Define the continuous change of the state feature vector within the continuous time prediction interval as the derivative of the state feature vector with respect to time, and determine the expression form of the state change rate. S333. Construct a continuous-time neural ordinary differential equation with the state feature vector as the variable, define the rate of state change as a nonlinear dynamic function of the time derivative of the state feature vector, and introduce a historical integral memory term and a real-time differential adjustment term for the state feature vector in the continuous-time prediction interval. S334. Based on the state characteristic trajectory data within the historical continuous time prediction interval, the adaptive gradient descent algorithm is used to iteratively optimize the parameter values ​​of the weight matrix, input weight matrix, bias vector, historical integral memory term weight matrix, and real-time differential adjustment term weight matrix. S335. Substitute the optimized weight matrix, input weight matrix, bias vector, historical integral memory term weight matrix, and real-time difference adjustment term weight matrix into the continuous-time neural ordinary differential equation to obtain a continuous-time dynamic model that can reflect the dynamic evolution characteristics of the state feature vector as it changes over continuous time.

7. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Map each abnormal state feature vector in the real-time abnormal state data set to the corresponding spatial grid cell to determine the spatial coordinate position of each abnormal state feature vector. S52. Construct a spatial anomaly weight matrix based on the anomaly degree quantification index according to the anomaly degree of each anomaly state feature vector and the spatial coordinate position of the corresponding spatial grid cell. S53. Using the spatial anomaly weight matrix as input, the spatial interpolation algorithm is used to calculate the anomaly propagation impact value of adjacent spatial grid cells that have not experienced anomalies step by step, and an anomaly propagation threshold is defined. S54. Based on the relationship between the impact value of abnormal propagation and the threshold value of abnormal propagation, determine the specific location of the spatial grid unit that is significantly affected by the abnormality, and determine the boundary range of the risk area of ​​sudden soil erosion event; S55. Based on the spatial grid boundary range of the risk area, calculate the cumulative sum of the abnormal propagation impact values ​​of all spatial grid units in the area to form a dynamic risk spatial distribution intensity map. S56. Update the dynamic risk spatial distribution intensity map in real time, and determine the specific spatial distribution location of the risk area of ​​sudden soil erosion event within the watershed space based on the updated risk spatial distribution intensity map.

8. The intelligent early warning method for watershed soil and water loss monitoring based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Based on the risk intensity and spatial correlation characteristics of each spatial grid unit in the dynamic risk spatial mapping map, generate risk level clustering relationships between spatial grid units; S62. Based on spatial risk level clustering relationships, identify and dynamically adjust the spatial extent of watershed risk areas, and automatically delineate initial risk areas, potential expansion areas, and risk propagation paths; S63. Based on the spatial range of the risk area, and combined with the continuity and changing trend of the spatial distribution of risk intensity, the risk level of the spatial grid unit is updated in real time to form a real-time dynamic risk level classification result. S64. Based on the real-time dynamic risk level classification results, each risk area is divided into stable area, diffusion area and drastic change area in real time, and a corresponding risk evolution model is automatically established. S65. Establish in advance a command system for soil and water conservation measures corresponding to stable zones, diffusion zones and drastic change zones respectively. The stable zone corresponds to the command for strengthening the monitoring of soil and water conservation measures, the diffusion zone corresponds to the command for early reinforcement of soil and water conservation facilities, and the drastic change zone corresponds to the command for emergency response mechanism and UAV image acquisition. S66. Based on the real-time dynamic risk level and corresponding risk evolution model of each risk area, automatically select and trigger the water and soil conservation measures instruction system for the corresponding area, and automatically execute the corresponding water and soil conservation measures in real time.

Images