A method and system for monitoring soil and water loss in karst areas

Abstract

The present invention discloses a method and system for monitoring soil and water loss in karst areas. The monitoring method obtains soil and water loss risk prediction results for a target area based on a deployed acoustic sensor array, deploys the acoustic sensor array at key monitoring points in the target area to be monitored, and simultaneously collects acoustic wave signals from the surface and underground environments. The acoustic wave signals detected by the acoustic sensor array are collected and converted into frequency domain feature data. The frequency domain feature data is input into a pre-trained acoustic pattern recognition model to identify acoustic features related to soil and water loss. Based on the output results of the acoustic pattern recognition model, the type, degree and development trend of soil and water loss in the monitored area are judged to generate soil and water loss risk prediction results. The present invention realizes real-time and continuous monitoring of the soil and water loss process in the karst area through acoustic monitoring means, improves monitoring accuracy and early warning timeliness, and provides a scientific basis for soil and water conservation management.

Description

Technical Field

[0001] The present invention belongs to the technical field of soil and water loss monitoring, and in particular relates to a method and system for monitoring soil and water loss in karst areas. Background Art

[0002] At present, the main problems in soil and water loss monitoring in karst areas include: traditional ground monitoring stations are sparse and cannot fully cover complex terrain; remote sensing images have limited penetration in densely vegetated areas; the correlation between groundwater processes and surface erosion is difficult to quantify; and sudden erosion caused by intermittent heavy rainfall is difficult to capture in a timely manner.

[0003] Furthermore, existing monitoring methods struggle to simultaneously address both surface and subsurface environmental monitoring needs, lacking the ability to dynamically monitor soil erosion processes in the unique geological structure of karst. Traditional monitoring methods often rely on manual on-site surveys or remote sensing image analysis at fixed intervals, making it difficult to continuously monitor soil erosion processes and provide early warnings. This leads to significant monitoring lags and prevents timely and effective decision-making support for prevention and control measures. Summary of the Invention

[0004] To address the above issues, the present invention proposes a method and system for monitoring soil and water loss in karst areas. By deploying an acoustic sensor array to simultaneously collect sound wave signals from the surface and underground environments, and combining it with a deep learning algorithm to construct an acoustic pattern recognition model, the method and system achieve real-time dynamic monitoring and early warning of soil and water loss processes in karst areas.

[0005] The specific technical solutions are as follows:

[0006] The present invention provides a method for monitoring soil and water loss in karst areas. The method obtains soil and water loss risk prediction results for a target area based on a deployed acoustic sensor array. The method comprises the following steps:

[0007] Step S1: deploying an acoustic sensor array at key monitoring points in a target area to be monitored. The acoustic sensor array includes a plurality of acoustic sensors for simultaneously collecting acoustic wave signals from the surface and underground environments.

[0008] The key monitoring points in the target area are divided into surface monitoring points and underground monitoring points. The surface monitoring points include at least karst depressions and karst grooves; the underground monitoring points include karst caves, cave entrances and underground river entrances.

[0009] Step S2: collecting the acoustic wave signals detected by the acoustic sensor array and converting the acoustic wave signals into frequency domain feature data.

[0010] Step S3: input the frequency domain feature data into a pre-trained acoustic pattern recognition model, which is based on a deep learning algorithm and is used to identify acoustic features related to soil erosion.

[0011] Step S4: Based on the output results of the acoustic pattern recognition model, the type, degree and development trend of soil and water loss in the monitored area are determined to generate a soil and water loss risk prediction result.

[0012] Furthermore, the process of extracting the frequency domain feature data includes:

[0013] The original sound wave signal is preprocessed, and the preprocessing includes denoising, filtering and signal enhancement; the preprocessed signal is subjected to time-frequency analysis to extract time domain features and frequency domain features.

[0014] Calculate the acoustic features of the sound wave signal, including Mel-frequency cepstral coefficients (MFCCs), spectral flux, spectral centroid, and spectral bandwidth; and extract the statistical features of the sound wave signal, including mean, variance, skewness, kurtosis, and energy distribution.

[0015] Furthermore, the acoustic features are calculated using short-time Fourier transform (STFT):

[0016] ;in, is the original sound wave signal, Indicates in In the analysis window, the position in the window is The sample value of It is the sample index within the window, i.e. the number of FFT points, ranging from 0 to N-1; is the time frame index, indicating the analysis windows; is the frame shift, which indicates the sample number offset between adjacent windows; k is the frequency index, which indicates the frequency point in the short-time Fourier transform (STFT); is a window function; Represents the short-time Fourier transform coefficient of the kth frequency point on the mth time frame;

[0017] The calculation formula of Mel frequency cepstral coefficient MFCC is:

[0018] ;in, represents the number of Mel filter banks, represents the cepstral coefficient index;

[0019] The calculation formula of spectral centroid SC is: ;

[0020] The calculation formula of spectral flux SF is: ;

[0021] Sound wave energy distribution The calculation formula is: ;in,

[0022] represents the frequency band index, and Indicates the The lower and upper frequency indices of the frequency band.

[0023] Furthermore, the acoustic pattern recognition model is a hybrid model of a deep convolutional neural network and a long short-term memory network, and its structure includes: an input layer, a convolution layer, a pooling layer, a long short-term memory network layer, a fully connected layer, and a softmax layer;

[0024] The input layer receives acoustic feature vectors, which are multidimensional arrays containing time domain features, frequency domain features, acoustic features, and statistical features. The time domain features include the time series data of the original signal after preprocessing; the frequency domain features are the spectrum data obtained by short-time Fourier transform; the acoustic features include Mel-frequency cepstral coefficients (MFCC), spectral flux (SF), spectral centroid (SC), and spectral bandwidth (SBW); and the statistical features include the mean, variance, skewness, kurtosis, and energy distribution of each frequency band of the signal.

[0025] The convolutional layer extracts local patterns of acoustic features, the pooling layer reduces feature dimensions and improves model robustness, the long short-term memory network layer captures the temporal dependencies of acoustic features, and the fully connected layer and softmax layer output the probability distribution of soil erosion type and degree;

[0026] The mathematical expression of the convolutional layer operation is: ,in: is the output of the previous layer; is the convolution kernel weight; is the bias term; is the rectified linear unit activation function; Represents the convolution operation;

[0027] Input gate of the long short-term memory network LSTM: ;

[0028] Forget Gate: , output gate: ;

[0029] Candidate memory cells: ;

[0030] Memory unit update: ;

[0031] Hide status update: ;

[0032] in: is the input of the current time step, is the hidden state of the previous time step, is the memory cell state at the previous time step, , , , is the weight matrix, , , , is the bias vector, is the hyperbolic tangent activation function, is the Hadamard product, using element-by-element multiplication; Represents the sigmoid activation function;

[0033] The expression of the output layer is: ;in, is the output of the LSTM layer; is the output layer weight matrix; is the output layer bias vector; is the normalized exponential function.

[0034] Furthermore, the acoustic pattern recognition model is trained using a cross entropy loss function: ;in, is the total number of soil erosion types, is the true label, Predict probabilities for the model.

[0035] Furthermore, the training steps of the acoustic pattern recognition model are:

[0036] Step S31, collecting acoustic wave samples generated by different types of soil erosion processes, including acoustic wave samples under surface runoff, rainfall erosion, rockfall, underground dissolution, and underground river flow scenarios;

[0037] Step S32: labeling the acoustic wave samples and establishing a corresponding relationship between the acoustic characteristics and the type and degree of soil erosion;

[0038] Step S33, using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction technology to extract the discriminative features of the sound wave sample;

[0039] The expression of the principal component analysis PCA is: ;in, is the original acoustic feature matrix, is the principal component loading matrix, is the feature matrix after dimensionality reduction;

[0040] Calculated from the eigenvector: ;in, is the covariance matrix of the original features, is the eigenvalue diagonal matrix;

[0041] Calculate the conditional probability in high-dimensional space using t-distributed random neighbor embedding: ;

[0042] Conditional probability in low-dimensional space: ;

[0043] The optimization goal is to minimize the KL divergence: ;

[0044] Step S34, using batch gradient descent method to optimize model parameters, the learning rate is initially set to 0.001, and a learning rate decay strategy is used;

[0045] Cluster analysis uses the K-means algorithm: ;in, For samples Belong to cluster indicator variables; Cluster the center of is the number of clusters, is the sample size;

[0046] Step S35: Use early stopping to avoid overfitting and stop training when the performance of the validation set no longer improves;

[0047] Furthermore, the method also includes a sound wave propagation model calibration step: establishing a sound wave propagation model for a karst area, taking into account factors such as topography, rock material, vegetation cover, and meteorological conditions, and conducting tests using known sound sources to obtain the propagation characteristics of sound waves in a karst environment; based on the test results, calibrating the sound wave propagation model parameters, and using the calibrated propagation model to improve the accuracy of sound source positioning and signal analysis;

[0048] The sound wave propagation model is based on the wave equation: ;in, is the sound pressure, is the speed of sound, is the Laplace operator;

[0049] Under complex terrain conditions, the ray tracing method is used to calculate the sound wave propagation path: ;in, is the sound wave ray path, is the path length parameter, is the refractive index of the acoustic wave, which is related to the sound velocity of the medium. , is the reference sound speed;

[0050] Sound wave attenuation model considering terrain and atmospheric conditions: ;in, is the sound pressure level at the receiving point; is the sound pressure level of the sound source; is the distance from the sound source to the receiving point; is the atmospheric absorption coefficient; is the attenuation caused by terrain; Attenuation caused by atmospheric conditions.

[0051] Furthermore, in step S4, based on the output of the acoustic pattern recognition model, a soil and water loss risk prediction result is generated, and a soil and water loss risk prediction score is generated. Expressed as: ; For the Weight coefficient of soil erosion class; For the The probability of soil erosion occurring For the severity of soil erosion;

[0052] Set multi-level warning thresholds based on risk scores, divided into four warning levels: mild, moderate, severe, and catastrophic;

[0053] When the risk score exceeds the warning threshold, the system automatically triggers the corresponding level of warning signal;

[0054] For different warning levels, corresponding emergency response measures are formulated, including increasing monitoring frequency, notifying relevant departments and activating emergency plans.

[0055] Furthermore, the method further includes the step of calculating the amount of soil loss based on the acoustic characteristics:

[0056] Calculating soil loss using acoustic signal characteristics monitored by acoustic sensors , whose expression is:

[0057] ;in, is the amount of soil loss, in tons / hectare; kappa is the calibration coefficient related to soil type; and is the start and end time of monitoring; is the frequency-dependent soil loss transfer function; is the acoustic signal intensity in the time-frequency domain.

[0058] The conversion function Determined according to different soil types and particle size distribution:

[0059]

[0060] in, is the benchmark conversion factor; For the Weight coefficient of soil particle size; For the Characteristic frequency of soil particle size; is the frequency distribution width; M is the number of soil particle size classifications.

[0061] The acoustic signal strength It is positively correlated with the scouring intensity during the soil erosion process and is calculated by the power spectral density of the short-time Fourier transform: ;in, For the On the time frame The short-time Fourier transform coefficients of the frequency points are Corresponding to the frequency index , Corresponding to the timeframe index ;

[0062] The system calculates the amount of soil loss With preset threshold Based on the comparison results, the soil and water loss risk prediction score is adjusted. When the risk score increases ,in is the adjustment coefficient, This is the largest amount of soil loss in history.

[0063] Based on the same inventive concept, the present invention provides a karst area soil erosion monitoring system for executing the monitoring method of the present invention, wherein the monitoring system comprises: a signal acquisition unit, a signal processing unit, a pattern recognition unit, and a risk assessment unit connected in sequence;

[0064] Furthermore, the signal acquisition unit is used to receive the sound wave signal collected by the acoustic sensor array and convert the sound wave signal into frequency domain feature data;

[0065] The acoustic sensor array includes a plurality of acoustic sensors for collecting acoustic wave signals in a target area. The acoustic sensor array is arranged at surface monitoring points and underground monitoring points. The surface monitoring points include at least karst depressions and karst grooves, and the underground monitoring points include karst caves, cave entrances, and underground river entrances.

[0066] Furthermore, the signal processing unit is used to perform preprocessing, time-frequency analysis and feature extraction on the frequency domain feature data to obtain a multidimensional feature vector including time domain features, frequency domain features, acoustic features and statistical features;

[0067] Furthermore, the pattern recognition unit includes a pre-trained acoustic pattern recognition model for receiving the multi-dimensional feature vector and outputting recognition results of the type, degree and development trend of soil and water loss;

[0068] Furthermore, the risk assessment unit is used to calculate a soil and water loss risk score based on the output result of the acoustic pattern recognition model, determine the risk level according to a preset early warning threshold, and generate a soil and water loss risk prediction result.

[0069] Furthermore, the risk assessment unit also includes: an early warning release module, which is used to automatically generate and release early warning information according to the risk score, and transmit the soil and water loss risk prediction result to the monitoring center.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] This method deploys acoustic sensor arrays at key monitoring points in karst areas, simultaneously collecting acoustic signals from both the surface and underground environments. Combined with an acoustic pattern recognition model based on a deep learning algorithm, it enables real-time monitoring and early warning of soil erosion. This method overcomes the limitations of traditional monitoring methods in karst areas and can effectively identify different types of soil erosion, including surface runoff, rainfall erosion, rockfall, underground dissolution, and underground river flow, and predict their development trends. Through the calibration of the acoustic wave propagation model and a multi-level early warning mechanism, monitoring accuracy and early warning timeliness are improved, providing a scientific basis and technical support for soil and water conservation in karst areas. It has strong practicality and promotional value. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 The present invention is a flow chart of a method for monitoring soil and water loss in karst areas.

[0073] Figure 2 The figure is a schematic diagram of the composition of a karst area soil and water loss monitoring system of the present invention. DETAILED DESCRIPTION

[0074] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0075] Example 1

[0076] like Figure 1 FIG. 1 is a method for monitoring soil and water loss in karst areas according to the present invention. The method obtains soil and water loss risk prediction results for a target area based on a deployed acoustic sensor array. The method comprises the following steps:

[0077] Step S1: deploying an acoustic sensor array at key monitoring points in a target area to be monitored. The acoustic sensor array includes a plurality of acoustic sensors for simultaneously collecting acoustic wave signals from the surface and underground environments.

[0078] The key monitoring points in the target area are divided into surface monitoring points and underground monitoring points. The surface monitoring points include at least karst depressions and karst grooves; the underground monitoring points include karst caves, cave entrances and underground river entrances.

[0079] Taking a typical karst region in a certain province as an example, an acoustic sensor array was deployed across a 25-square-kilometer monitoring area. Specifically, three high-sensitivity microphone sensors were deployed at each of five surface monitoring points (including three karst depressions and two karst troughs), with a sampling frequency of 48 kHz. Two waterproof acoustic sensors were deployed at each of three underground monitoring points (including one inside a karst cave, one at the cave entrance, and one at the entrance to an underground river), with a sampling frequency of 44.1 kHz. The sensors were connected to a data collection center via a wireless network, enabling real-time data transmission.

[0080] Step S2: collecting the acoustic wave signals detected by the acoustic sensor array and converting the acoustic wave signals into frequency domain feature data.

[0081] Acoustic signal acquisition was conducted during a rainstorm, lasting four hours and accumulating 85 mm of rainfall. The acoustic sensor array recorded approximately 14,400 seconds of acoustic data during the rainstorm. The raw acoustic signals were preprocessed using a bandpass filter (cutoff frequency 100 Hz-15 kHz) to remove ambient noise, an adaptive noise reduction algorithm to suppress wind noise, and dynamic range compression to enhance the effective signal. After preprocessing, the continuous acoustic signal was segmented into 10-second analysis windows with a 50% overlap, resulting in approximately 2,880 signal segments. Time-frequency analysis of each signal segment was performed using a short-time Fourier transform (STFT) with a window length of 2,048 points and a frame shift of 512 points to obtain frequency domain feature data.

[0082] Step S3: input the frequency domain feature data into a pre-trained acoustic pattern recognition model, which is based on a deep learning algorithm and is used to identify acoustic features related to soil erosion.

[0083] The extracted frequency domain features include 13th-order Mel-frequency cepstral coefficients (MFCCs), spectral flux (SF), spectral centroid (SC), and spectral bandwidth (SBW), as well as statistical features (mean, variance, skewness, kurtosis, and energy distribution in 10 frequency bands). For a 10-second signal segment, a 68-dimensional feature vector was extracted. These feature vectors were then fed into a pre-trained acoustic pattern recognition model for analysis. The model employed a CNN-LSTM hybrid architecture. The CNN layer consisted of three convolutional layers (with kernel sizes of 3×3, 5×5, and 7×7, and 32, 64, and 128 filters, respectively), each followed by a max pooling layer. The LSTM layer consisted of two bidirectional LSTM layers, each with 128 units. The fully connected layer contained 256 neurons, with a dropout rate of 0.5 to prevent overfitting. The output layer used a softmax function to map the probabilities of five soil erosion types (surface runoff, rainfall erosion, rockfall, underground dissolution, and underground river flow).

[0084] The model was trained using 30,000 labeled sound wave samples from 10 different karst regions. Five-fold cross-validation was used to evaluate model performance, ultimately achieving a classification accuracy of 92.7% on the test set. Model parameters were updated using the Adam optimizer, with an initial learning rate of 0.001, decayed by 10% every 20 epochs, and a batch size of 64 for a total of 100 epochs. To improve the model's generalization in real-world settings, data augmentation techniques were employed during training, including the addition of Gaussian white noise with varying signal-to-noise ratios, random time stretching, and frequency shifting.

[0085] Step S4: Based on the output results of the acoustic pattern recognition model, the type, degree and development trend of soil and water loss in the monitored area are determined to generate a soil and water loss risk prediction result.

[0086] For this rainstorm, acoustic pattern recognition model analysis revealed strong surface runoff signals detected at depression monitoring points with a confidence level of 0.95; moderate rainfall erosion signals detected at karst gullies with a confidence level of 0.87; a small-scale rockfall signal detected at a karst cave entrance with a confidence level of 0.83; and a significant increase in underground river flow detected at the entrance of an underground river with a confidence level of 0.91. Based on these identification results, a risk scoring formula was used to calculate the soil and water loss risk of this rainstorm.

[0087] For surface runoff: , , For rainfall washoff: , For rockfalls: For underground dissolution: , For underground river flows: ; According to the risk scoring formula: Risk=0.552.

[0088] Based on preset warning thresholds (minor: 0.0-0.3; moderate: 0.3-0.5; severe: 0.5-0.7; catastrophic: >0.7), the soil erosion risk score for this rainstorm was 0.552, placing it at the severe warning level. The system automatically generated a warning message and sent it to the local soil and water conservation authorities, recommending the following emergency measures: increase monitoring frequency to every 30 minutes; dispatch personnel to conduct on-site inspections at the rockfall site; inspect the drainage systems in communities downstream of the underground river; and prepare temporary drainage facilities, sandbags, and other protective supplies.

[0089] During the calibration process for the acoustic wave propagation model, tests were conducted using a simulated sound source at a known location (including a 20Hz-20kHz frequency sweep signal). The sound pressure level attenuation was measured at different distances (10m, 50m, 100m, and 200m). The attenuation coefficients in the acoustic wave propagation model were adjusted based on local topographic data and rock material parameters (primarily limestone with a sound velocity of approximately 6000m / s). The test results showed that under conditions of 70% humidity and 25°C, the actual attenuation coefficient for surface sound wave propagation was approximately 1.2 times the theoretical value, while the attenuation coefficient for sound wave propagation in underground caves was approximately 0.8 times the theoretical value. Based on these test data, the acoustic wave propagation model parameters were calibrated, improving the sound source localization accuracy, reducing the average localization error from ±15m to ±5m.

[0090] In this monitoring example, the system operated continuously for seven days, covering the complete monitoring cycle before, during, and after the rainstorm. Analysis of the temporal changes in acoustic wave characteristics during the rainstorm revealed a time lag of approximately 15 minutes between surface runoff intensity and rainfall intensity, and a time lag of approximately 40 minutes between changes in underground river flow and surface runoff intensity. These temporal relationships provide a basis for developing more accurate early warning times. After the seven-day monitoring period, technicians conducted an on-site verification and found that the system's predicted areas of soil erosion matched actual conditions 89% of the time, demonstrating the effectiveness and accuracy of this monitoring method.

[0091] The process of extracting the frequency domain feature data includes:

[0092] The original sound wave signal is preprocessed, and the preprocessing includes denoising, filtering and signal enhancement; the preprocessed signal is subjected to time-frequency analysis to extract time domain features and frequency domain features.

[0093] Calculate the acoustic features of the sound wave signal, including Mel-frequency cepstral coefficients (MFCCs), spectral flux, spectral centroid, and spectral bandwidth; and extract the statistical features of the sound wave signal, including mean, variance, skewness, kurtosis, and energy distribution.

[0094] Acoustic features are calculated using short-time Fourier transform (STFT):

[0095] ;in, is the original sound wave signal, Indicates in In the analysis window, the position in the window is The sample value of It is the sample index within the window, i.e. the number of FFT points, ranging from 0 to N-1; is the time frame index, indicating the analysis windows; is the frame shift, which indicates the sample number offset between adjacent windows; k is the frequency index, which indicates the frequency point in the short-time Fourier transform (STFT); is a window function; Represents the short-time Fourier transform coefficient of the kth frequency point on the mth time frame;

[0096] The calculation formula of Mel frequency cepstral coefficient MFCC is:

[0097] ;in, represents the number of Mel filter banks, represents the cepstral coefficient index;

[0098] The calculation formula of spectral centroid SC is: ;

[0099] The calculation formula of spectral flux SF is: ;

[0100] Sound wave energy distribution The calculation formula is: ;in,

[0101] represents the frequency band index, and Indicates the The lower and upper frequency indices of the frequency band.

[0102] The acoustic pattern recognition model is a hybrid model of a deep convolutional neural network and a long short-term memory network, and its structure includes: an input layer, a convolution layer, a pooling layer, a long short-term memory network layer, a fully connected layer, and a softmax layer;

[0103] The input layer receives acoustic feature vectors, which are multidimensional arrays containing time domain features, frequency domain features, acoustic features, and statistical features. The time domain features include the time series data of the original signal after preprocessing; the frequency domain features are the spectrum data obtained by short-time Fourier transform; the acoustic features include Mel-frequency cepstral coefficients (MFCC), spectral flux (SF), spectral centroid (SC), and spectral bandwidth (SBW); and the statistical features include the mean, variance, skewness, kurtosis, and energy distribution of each frequency band of the signal.

[0104] The convolutional layer extracts local patterns of acoustic features, the pooling layer reduces feature dimensions and improves model robustness, the long short-term memory network layer captures the temporal dependencies of acoustic features, and the fully connected layer and softmax layer output the probability distribution of soil erosion type and degree;

[0105] The mathematical expression of the convolutional layer operation is: ,in: is the output of the previous layer; is the convolution kernel weight; is the bias term; is the rectified linear unit activation function; Represents the convolution operation;

[0106] Input gate of the long short-term memory network LSTM: ;

[0107] Forget Gate: , output gate: ;

[0108] Candidate memory cells: ;

[0109] Memory unit update: ;

[0110] Hide status update: ;

[0111] in: is the input of the current time step, is the hidden state of the previous time step, is the memory cell state at the previous time step, , , , is the weight matrix, , , , is the bias vector, is the hyperbolic tangent activation function, is the Hadamard product, using element-by-element multiplication; Represents the sigmoid activation function;

[0112] The expression of the output layer is: ;in, is the output of the LSTM layer; is the output layer weight matrix; is the output layer bias vector; is the normalized exponential function.

[0113] The acoustic pattern recognition model is trained using the cross entropy loss function: ;in, is the total number of soil erosion types, is the true label, Predict probabilities for the model.

[0114] The training steps of the acoustic pattern recognition model are:

[0115] Step S31, collecting acoustic wave samples generated by different types of soil erosion processes, including acoustic wave samples under surface runoff, rainfall erosion, rockfall, underground dissolution, and underground river flow scenarios;

[0116] Step S32: labeling the acoustic wave samples and establishing a corresponding relationship between the acoustic characteristics and the type and degree of soil erosion;

[0117] Step S33, using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction technology to extract the discriminative features of the sound wave sample;

[0118] The expression of the principal component analysis PCA is: ;in, is the original acoustic feature matrix, is the principal component loading matrix, is the feature matrix after dimensionality reduction;

[0119] Calculated from the eigenvector: ;in, is the covariance matrix of the original features, is the eigenvalue diagonal matrix;

[0120] Calculate the conditional probability in high-dimensional space using t-distributed random neighbor embedding: ;

[0121] Conditional probability in low-dimensional space: ;

[0122] The optimization goal is to minimize the KL divergence: ;

[0123] Step S34, using batch gradient descent method to optimize model parameters, the learning rate is initially set to 0.001, and a learning rate decay strategy is used;

[0124] Cluster analysis uses the K-means algorithm: ;in, For samples Belong to cluster indicator variables; Cluster the center of is the number of clusters, is the sample size;

[0125] Step S35: Use early stopping to avoid overfitting and stop training when the performance of the validation set no longer improves;

[0126] The method further includes a sound wave propagation model calibration step: establishing a sound wave propagation model for a karst area, taking into account factors affecting topography, rock material, vegetation cover, and meteorological conditions, and conducting tests using known sound sources to obtain the propagation characteristics of sound waves in the karst environment; calibrating the sound wave propagation model parameters based on the test results, and using the calibrated propagation model to improve the accuracy of sound source positioning and signal analysis;

[0127] The sound wave propagation model is based on the wave equation: ;in, is the sound pressure, is the speed of sound,

[0128] is the Laplace operator;

[0129] Under complex terrain conditions, the ray tracing method is used to calculate the sound wave propagation path: ;in, is the sound wave ray path, is the path length parameter, is the refractive index of the acoustic wave, which is related to the sound velocity of the medium. , is the reference sound speed;

[0130] Sound wave attenuation model considering terrain and atmospheric conditions: ;in, is the sound pressure level at the receiving point; is the sound pressure level of the sound source; is the distance from the sound source to the receiving point; is the atmospheric absorption coefficient; is the attenuation caused by terrain; Attenuation caused by atmospheric conditions.

[0131] This paper collected acoustic wave samples of five types of soil erosion. The sample data came from 30 monitoring points in three typical karst areas of Liupanshui City, Bijie City and Qiannan Prefecture in Guizhou Province, over a period of two years, covering different seasons and weather conditions. Specifically, 12,500 surface runoff samples, each 15 seconds long, were collected after rainfall events of varying intensities (light, moderate, heavy, and torrential rain). 10,800 rainfall erosion samples, each 10 seconds long, covered a range of rainfall intensities, from drizzle to torrential rain. 3,600 rockfall samples, each 5 seconds long, included small (<0.5 m³), ​​medium (0.5-5 m³), ​​and large (>5 m³) rockfall events. 5,200 underground dissolution samples, each 30 seconds long, were obtained through a combination of specially designed water dissolution experiments and field monitoring. 8,900 underground river flow samples, each 20 seconds long, were collected at underground river entrances across different seasons and flow conditions. These samples were reviewed and labeled by professional acoustic engineers and soil and water conservation experts to ensure labeling quality.

[0132] Acoustic sample labeling uses a three-level labeling system. On-site staff will make preliminary labeling based on the actual situation at the time of collection, recording information such as the type, time, location and visual estimation of soil erosion intensity; acoustic experts will listen to and perform spectral analysis of the acoustic wave signals to verify and refine the preliminary labeling, and accurately mark the starting and ending points of the effective signals on the timeline; soil and water conservation experts will conduct a comprehensive assessment of the samples based on the hydrological and meteorological data collected during the same period and the results of field investigations to determine the precise type of soil erosion and the quantitative severity.

[0133] For example, for a rockfall acoustic wave sample, the final annotation includes basic information (acquisition time: 14:23:45, July 15, 2022; acquisition location: northeast corner of Maolan Nature Reserve, Libo County, Qiannan Prefecture; device ID: KS-M07; weather conditions: thunderstorm; ambient noise: moderate), acoustic characteristics (peak intensity: 82dB; main frequency band: 120-800Hz; duration: 2.3 seconds; signal-to-noise ratio: 18dB), and soil erosion information (type: rockfall; volume: approximately 2.8m³; collapsed material: limestone; associated phenomenon: small landslide; severity coefficient: 0.75). This multi-level, multi-dimensional annotation system ensures the high quality and reliability of training data.

[0134] The feature dimensionality reduction process employed a two-stage strategy, using principal component analysis (PCA) to reduce the original 68-dimensional acoustic features to 30 dimensions, preserving approximately 95% of the variance. The t-SNE algorithm was then applied to further reduce the features to 3 dimensions for visualization and clustering. During the PCA dimensionality reduction process, the first 10 principal components were found to be primarily associated with time-frequency features (accounting for 62% of the total variance), while principal components 11 to 20 were primarily associated with statistical features (accounting for 24% of the total variance), and principal components 21 to 30 were primarily associated with spectral features within the acoustic signature (accounting for 9% of the total variance). This indicates that time-frequency features play a dominant role in identifying soil erosion types, consistent with the physical properties of acoustic waves generated during soil erosion. Through t-SNE visualization, it is clearly observed that the five types of soil erosion form relatively separated clusters in the feature space. Among them, the surface runoff type and the underground river flow type have some overlap, which is consistent with the similarity of their physical mechanisms; while the rock collapse type samples are the most dispersed in the feature space, reflecting the huge differences in the acoustic characteristics of collapse events of different scales.

[0135] Model parameter optimization used batch gradient descent with momentum, with a batch size of 128, an initial learning rate of 0.001, and a momentum coefficient of 0.9. To prevent gradient explosion, gradient clipping was used to limit the gradient norm to less than 5. Cosine annealing was used as the learning rate decay strategy, halving the learning rate every 50 epochs. Multiple regularization techniques were applied during model training: the L2 regularization coefficient was set to 0.0005, and dropout rates were set at different levels for different layers (0.2 for convolutional layers, 0.5 for LSTM layers, and 0.6 for fully connected layers). To address sample imbalance (e.g., the number of rockfall samples was significantly lower than that of other types), a weighted cross-entropy loss function was used, with weights inversely proportional to the number of samples in each category. In the K-means clustering analysis, the silhouette coefficient was used to determine the optimal number of clusters, K=5, which corresponds to the five predefined soil erosion types. The clustering results show that the intra-class closeness and inter-class separation of different types of samples have reached an ideal level, and the clustering accuracy has reached 87.3%, which further verifies the effectiveness of the extracted features.

[0136] The early stopping method was implemented as follows: after every five epochs of training, model performance was evaluated on the validation set. Training was stopped if the F1 score on the validation set did not improve significantly (improvement of less than 0.1%) after 10 consecutive evaluations. In this study, model training typically converged after 120-150 epochs. To comprehensively evaluate model performance, in addition to accuracy, other metrics such as precision, recall, F1 score, and confusion matrix were calculated. The average precision for the five types of soil erosion was 89.6%, the average recall was 86.2%, and the average F1 score was 87.9%. Analysis of the confusion matrix revealed a misclassification rate of approximately 8% between surface runoff and subsurface river flow, consistent with the feature overlap observed in the t-SNE visualization. The rockfall type had the highest recognition accuracy (93.5%), attributed to its unique acoustic signature.

[0137] The model training of the present invention also takes into account the improvement of generalization and anti-interference capabilities. To enhance the model's adaptability to different monitoring environments, a variety of background noises were introduced into the training data, including wind noise, biological noise (such as bird calls and insect chirping), and human noise (such as traffic and construction), with the signal-to-noise ratio ranging from -5dB to 20dB. Through data augmentation technology, the original training sample size was expanded by 3 times, enhancing the generalization ability of the model. To verify the performance of the model in a new environment, test samples were collected in the Puzhehei Karst area of ​​Yunnan Province, outside Guizhou Province. The model recognition accuracy rate reached 83.7%, demonstrating its good generalization ability. In addition, the present invention also implements a continuous model update mechanism, collecting new sample data every quarter for model fine-tuning, enabling the system to adapt to factors such as environmental changes and equipment aging, and maintain high recognition performance. In two years of actual application, through quarterly updates, the average recognition accuracy of the model at the same monitoring point increased from the initial 89.3% to 94.1%, demonstrating the effectiveness of continuous learning.

[0138] In terms of hardware implementation of model training, the present invention adopts a distributed computing architecture and uses 4 servers equipped with NVIDIA Tesla V100 GPUs for parallel training. A single complete training takes about 12 hours. In order to meet the needs of edge computing, a pruned and quantized version of the model has also been developed. The number of parameters has been reduced by 78%, and the inference speed has been increased by 3.2 times. It can be deployed on field equipment with limited computing resources. At monitoring points in the wild where there is no electricity supply, the long-term autonomous operation of the system is achieved by combining solar power supply and low-power computing modules. The average delay in the model inference stage is only 236 milliseconds, which meets the needs of real-time monitoring. Through these technical means, the present invention realizes the optimization of the entire process from training to deployment, providing an efficient and reliable artificial intelligence solution for soil and water loss monitoring in karst areas.

[0139] In step S4, based on the output of the acoustic pattern recognition model, a soil and water loss risk prediction result is generated, and a soil and water loss risk prediction score is generated. Expressed as: ; For the Weight coefficient of soil erosion class; For the The probability of soil erosion occurring For the severity of soil erosion;

[0140] Set multi-level warning thresholds based on risk scores, divided into four warning levels: mild, moderate, severe, and catastrophic;

[0141] When the risk score exceeds the warning threshold, the system automatically triggers the corresponding level of warning signal;

[0142] For different warning levels, corresponding emergency response measures are formulated, including increasing monitoring frequency, notifying relevant departments and activating emergency plans.

[0143] In the present invention, soil erosion types are classified in detail according to acoustic characteristics and are divided into five categories: surface runoff type, rainfall erosion type, rock collapse type, underground dissolution type and underground river flow type. Surface runoff type soil erosion is characterized by a low-frequency (20-200Hz) continuous acoustic wave signal with uniform spectral energy distribution, and the amplitude increases with the increase of runoff intensity; rainfall erosion type soil erosion is characterized by a medium-frequency (200-1000Hz) irregular pulse acoustic wave signal with obvious raindrop impact acoustic spectrum characteristics, and the spectral energy varies with rainfall intensity; rock collapse type soil erosion is characterized by a broadband (100-5000Hz) sudden high-energy acoustic wave signal with obvious impact and attenuation process, and the spectral energy is mainly concentrated in the low-frequency band; underground dissolution type soil erosion is characterized by a low-frequency (50-500Hz) continuous weak signal with special water-rock interaction spectral characteristics; underground river flow type soil erosion is characterized by a medium-low frequency (100-800Hz) continuous acoustic wave signal with typical water turbulence acoustic characteristics, and the spectral center of mass moves toward high frequency with increasing flow velocity.

[0144] Based on the above soil and water loss types, the present invention sets detailed risk scoring standards and warning thresholds. Specifically, the severity , The quantitative standards are as follows: For surface runoff type, when the flow velocity is less than 0.5m / s, The value is 0.3; when the flow velocity is between 0.5-1.5m / s, Take the value as 0.6; when the flow velocity is greater than 1.5m / s, The value is 0.9. Rainfall erosion type, when the rainfall intensity is less than 30mm / h, The value is 0.4; when the rainfall intensity is between 30-60mm / h, The value is 0.7; when the rainfall intensity is greater than 60mm / h, The value is 0.95. Rock collapse type, according to the collapse volume, when it is less than 1m³, Take the value as 0.5; when the volume is between 1-10m³, Take the value as 0.8; when the volume is greater than 10m³, Take the value as 1.0. Underground dissolution type, according to the dissolution rate, when the rate is low (less than 5mm / year), The value is 0.3; at medium rate (5-20mm / year), The value is 0.6; at high rates (greater than 20 mm / year), The value is 0.8. Underground river flow type, according to the flow change rate, when the increase is less than 50%, The value is 0.4; when the growth is between 50-100%, The value is 0.7; when the growth is greater than 100%, The value is 0.9.

[0145] Weight coefficient The determination is based on the historical data statistics of soil and water loss in karst areas and expert experience evaluation, taking into account the degree of harm and frequency of occurrence of different types of soil and water loss. =0.30, rainfall washout type =0.25, rock fall type =0.20, underground dissolution type =0.10, underground river flow type = 0.15. The sum of these weighting coefficients equals 1, ensuring the normalization of the risk score.

[0146] The specific division of warning thresholds and corresponding emergency response measures are as follows: When the risk score is between 0.0-0.3, it is judged as a mild warning level. Emergency response measures include maintaining regular monitoring frequency, recording monitoring data, and no special intervention is required; when the risk score is between 0.3-0.5, it is judged as a moderate warning level. Emergency response measures include increasing the monitoring frequency to once every hour, notifying local soil and water conservation departments to keep an eye on it, and preparing necessary protective materials; when the risk score is between 0.5-0.7, it is judged as a severe warning level. Emergency response measures include increasing the monitoring frequency to once every 30 minutes, dispatching technical personnel to high-risk areas for on-site inspections, notifying relevant communities and enterprises to make protective preparations, and activating the first-level emergency plan; when the risk score is greater than 0.7, it is judged as a disaster warning level. Emergency response measures include increasing the monitoring frequency to once every 10 minutes, immediately notifying local governments and soil and water conservation, emergency management and other departments, considering personnel evacuation and traffic control measures, and fully activating emergency plans and rescue systems.

[0147] In long-term monitoring practice, this invention has also established an adaptive adjustment mechanism for warning thresholds. By comparing warning results with actual soil and water loss conditions, warning threshold parameters are regularly updated. During rainy seasons, the warning threshold is appropriately lowered to enhance system sensitivity; during dry seasons, the threshold is appropriately raised to reduce false alarm rates. Furthermore, customized warning thresholds are established for each monitored area based on the geological characteristics and soil and water loss history of each karst region, further improving the accuracy and practicality of warnings.

[0148] For example, for areas with loose geological structure and severe soil erosion in history, the threshold of the severe warning level can be appropriately lowered to 0.45-0.65; while for areas with stable geology and complete soil and water conservation measures, the threshold can be appropriately raised to 0.55-0.75.

[0149] To verify the effectiveness of the early warning system, the present invention was tested and applied for one year in three typical karst watersheds in a county in a certain province.

[0150] Results show that the system's generated warnings match actual soil and water loss incidents 82% of the time, with a 91% match for severe and catastrophic warning levels. Warnings are issued an average of 40 minutes in advance, providing valuable decision-making time for local soil and water conservation management and disaster prevention and mitigation. Over the course of a year, the system issued 182 warnings, including 98 minor, 56 moderate, 22 severe, and 6 catastrophic warnings. These warnings effectively guided local soil and water conservation efforts and averted numerous potentially serious soil and water loss disasters.

[0151] The method also includes a step of calculating the amount of soil loss based on acoustic characteristics. By analyzing the intensity, frequency distribution and time domain variation characteristics of the sound wave signal, a quantitative relationship model between the acoustic characteristics and the amount of soil loss is established, thereby achieving an accurate assessment of the degree of soil erosion in the karst area.

[0152] Calculating soil loss using acoustic signal characteristics monitored by acoustic sensors , whose expression is:

[0153] ;in, is the amount of soil loss, in tons / hectare; kappa is the calibration coefficient related to soil type; and is the start and end time of monitoring; is the frequency-dependent soil loss transfer function; is the acoustic signal intensity in the time-frequency domain.

[0154] The conversion function Determined according to different soil types and particle size distribution:

[0155]

[0156] in, is the benchmark conversion factor; For the Weight coefficient of soil particle size; For the Characteristic frequency of soil particle size; is the frequency distribution width; M is the number of soil particle size classifications.

[0157] The acoustic signal strength It is positively correlated with the scouring intensity during the soil erosion process and is calculated by the power spectral density of the short-time Fourier transform: ;in, For the On the time frame The short-time Fourier transform coefficients of the frequency points are Corresponding to the frequency index , Corresponding to the timeframe index ;

[0158] The system calculates the amount of soil loss With preset threshold Based on the comparison results, the soil and water loss risk prediction score is adjusted. When the risk score increases ,in is the adjustment coefficient, This is the largest amount of soil loss in history.

[0159] The present invention also provides a method for assessing soil loss based on acoustic characteristics. This method uses the physical correlation between acoustic wave signal characteristics and the soil scouring process to establish a quantitative relationship model between acoustic parameters and soil loss. Studies have shown that the sound waves generated by soil particles of different particle sizes during the scouring process have different frequency characteristics: fine sand mainly produces sound waves of 800-1200Hz, coarse sand mainly produces sound waves of 400-800Hz, small-particle soil aggregates mainly produce sound waves of 200-400Hz, and large-particle soil aggregates mainly produce sound waves of 100-200Hz. Based on this characteristic, the present invention can identify the loss characteristics of soils of different components by analyzing the spectral distribution of the sound wave signal.

[0160] In practical applications, the system first establishes a standard soil sample library containing acoustic characteristic data of typical soil types in the study area. Through controlled experiments, the acoustic wave signals generated by different types of soil when water flows of different intensities are flushed, and the corresponding soil loss is recorded to establish a calibration curve. Determined by factors such as soil type, moisture content, and compaction, it is usually between 0.5 and 2.0. Frequency conversion function It reflects the corresponding relationship between acoustic wave signals of different frequencies and soil loss. The parameters in this function are obtained by fitting experimental data.

[0161] For example, during a four-hour period of heavy rainfall in one study area, the acoustic signal intensity at a monitoring point in a karst depression increased significantly in the 400-800 Hz frequency band, with the average power spectral density reaching eight times the normal level. Based on a pre-established model, the system calculated that soil loss at this monitoring point reached 4.3 tons per hectare, exceeding the warning threshold of 3.0 tons per hectare. Consequently, the risk assessment unit increased the risk score by 0.15, ultimately reaching 0.65, triggering a severe warning signal.

[0162] In the two years since the system was deployed, 32 soil loss assessments were compared with field measurements, achieving an accuracy rate of 88.4% and an average relative error of 12.3%. This result demonstrates the accuracy and practicality of the acoustic signature-based soil loss assessment method, providing an important quantitative basis for soil and water conservation efforts in karst areas.

[0163] Example 2

[0164] like Figure 2As shown, a schematic diagram of the composition of a karst area soil and water loss monitoring system of the present invention is used to execute the monitoring method of Example 1. The monitoring system includes: a signal acquisition unit, a signal processing unit, a pattern recognition unit and a risk assessment unit connected in sequence.

[0165] The signal acquisition unit is used to receive the sound wave signal collected by the acoustic sensor array and convert the sound wave signal into frequency domain feature data.

[0166] The acoustic sensor array includes multiple acoustic sensors for collecting sound wave signals in the target area. The acoustic sensor array is set at surface monitoring points and underground monitoring points. The surface monitoring points include at least karst depressions and dissolution grooves, and the underground monitoring points include karst caves, cave entrances, and underground river entrances.

[0167] The signal processing unit is used to perform preprocessing, time-frequency analysis and feature extraction on the frequency domain feature data to obtain a multidimensional feature vector including time domain features, frequency domain features, acoustic features and statistical features.

[0168] The pattern recognition unit includes a pre-trained acoustic pattern recognition model, which is used to receive the multi-dimensional feature vector, output recognition results of the type, degree and development trend of soil and water loss, and calculate the amount of soil loss.

[0169] The risk assessment unit is used to calculate the soil and water loss risk score based on the output results of the acoustic pattern recognition model, combine the soil loss amount assessment results, judge the risk level according to the preset early warning threshold, and generate a soil and water loss risk prediction result.

[0170] The risk assessment unit further comprises: an early warning issuing module, which is used to automatically generate and issue early warning information according to the risk score, and transmit the soil and water loss risk prediction result to a monitoring center.

[0171] The signal acquisition unit consists of a field-deployed acoustic sensor array and a data preprocessing module. The acoustic sensor array utilizes three types of acoustic sensors: an omnidirectional microphone for surface monitoring (frequency range: 20Hz-20kHz, sensitivity: -38dB, signal-to-noise ratio: >68dB), with a sampling rate of 48kHz; a low-frequency sensor for cave monitoring (frequency range: 5Hz-10kHz, sensitivity: -35dB, signal-to-noise ratio: >72dB), with a sampling rate of 44.1kHz; and a hydrophone for underwater monitoring (frequency range: 10Hz-15kHz, sensitivity: -160dB at 1V / μPa, maximum operating depth: 50m), with a sampling rate of 32kHz. All sensors utilize IP67 waterproof housings with built-in dust and moisture filters to withstand the high humidity of the karst region.

[0172] Data transmission utilizes a multi-level architecture: sensors within monitoring points connect to an on-site edge computing unit via Bluetooth Low Energy (BLE 5.0). The edge computing unit then transmits data to a cloud server via 4G / 5G mobile networks or LoRa long-range IoT technology. In remote areas with poor network coverage, a satellite communication backup link ensures data transmission reliability. To reduce bandwidth requirements, the edge computing unit performs preliminary signal compression and feature extraction, transmitting only critical data, reducing data transmission by an average of 90%. The edge computing unit is based on an ARM Cortex-A72 processor, equipped with 2GB of RAM and 32GB of flash memory. It supports the TensorFlow Lite framework, enabling local execution of lightweight model inference. The unit's low-power design, coupled with a 100W solar panel and a 26,800mAh lithium battery pack, allows for 14 days of continuous operation without an external power source.

[0173] The signal processing unit, deployed on a cloud server and equipped with dual Intel Xeon processors and 32GB of memory, is responsible for computationally intensive signal processing tasks. It first denoises the received acoustic signal using a combined filtering strategy: a Wiener filter for stable ambient noise, a median filter for impulsive noise, and an adaptive wavelet threshold filter for random noise. The signal-to-noise ratio (SNR) of the denoised signal improves by an average of 12dB. The time-frequency analysis module uses a combination of short-time Fourier transform and wavelet packet transform to capture both the short-time and frequency characteristics of the signal. The feature extraction module utilizes a parallel computing architecture to simultaneously process multiple signal segments, achieving an eightfold increase in computational speed compared to a single-threaded implementation.

[0174] The pattern recognition unit implements a two-level recognition architecture: the first level identifies soil erosion types using a CNN-LSTM hybrid model to distinguish five basic soil erosion types; the second level assesses severity using a regression model that outputs a severity coefficient between 0 and 1. To improve system robustness, the unit also implements a model integration mechanism, running three independently trained models (CNN-LSTM, pure CNN, and gradient boosting tree) simultaneously, using a weighted voting strategy to determine the final recognition result. Model weights are dynamically adjusted based on each model's historical performance under different environmental conditions. The unit also implements incremental learning, updating the model monthly with newly collected annotated data to ensure continuous optimization of system performance. In actual application, the model accuracy has gradually increased from 88% upon initial deployment to 95% after one year.

[0175] The pattern recognition unit also includes a soil loss assessment module, which is used to calculate the soil loss based on the acoustic feature data. The calculation formula is:

[0176]

[0177] A mapping relationship between acoustic characteristics of different frequencies and the soil loss process was established, and a multi-channel acoustic parameter analysis method was used to identify the characteristic scour frequencies of soil particles of different sizes. The area of ​​soil erosion and the scope of impact were estimated through the spatial distribution analysis of acoustic energy combined with the topographic and geomorphological characteristics of the karst area. A spatiotemporal distribution map of soil loss was generated to provide a decision-making basis for risk assessment units.

[0178] The risk assessment unit is the system's decision-making hub, responsible for calculating soil erosion risk scores based on identification results and generating early warning information. This unit utilizes a combination of rule-based expert systems and statistical models, comprehensively considering acoustic identification results, historical data, terrain characteristics, and meteorological conditions. The risk assessment process incorporates spatiotemporal correlation analysis: if multiple nearby monitoring points simultaneously detect soil erosion signals, the risk score automatically increases. If the current identification results significantly deviate from historical patterns, the system initiates a cross-validation mechanism to prevent false alarms. The early warning dissemination module distributes early warning information through multiple channels: detailed technical reports to monitoring center administrators; text messages and app push notifications to field personnel; standardized early warning notifications to local government departments; and public-accessible warning information on public platforms. Early warning information includes elements such as risk level, impact area, development trends, recommended measures, and expiration date. The content is automatically adjusted based on the recipient to ensure effective and actionable communication.

[0179] In terms of system operation and maintenance, this embodiment has designed a comprehensive self-diagnosis and remote maintenance mechanism. Each hardware component is equipped with a health monitoring module that regularly checks battery power, storage space, network connectivity, and sensor status. When an anomaly is detected, the system automatically records a diagnostic log and sends an alert to maintenance personnel. For common faults, the system can execute automatic recovery procedures, such as restarting the communication module, switching to a backup power source, or starting a backup sensor. The remote maintenance interface allows technicians to access on-site devices through an encrypted connection to perform firmware updates, parameter adjustments, and troubleshooting, significantly reducing the need for on-site maintenance. Statistics show that the system's mean time between failures (MTBF) reaches 8,760 hours (one year), the mean time to recovery (MTTR) is less than 24 hours, and the annual availability exceeds 99.7%.

[0180] In practical application, this system was deployed at 25 monitoring points in a county in a certain province in 2021, covering approximately 50 square kilometers of a typical karst area. During its 18-month operation, the system issued 253 warnings, including 126 minor, 84 moderate, 37 severe, and 6 catastrophic. Compared with manual on-site survey results, the warning accuracy rate was 92.1%. Notably, during a torrential downpour in June 2022, the system issued a 45-minute advance warning of an impending large-scale rockfall, enabling local authorities to promptly close off the danger zone and avert potential casualties. The system also discovered three previously unknown areas of active underground cave development, providing valuable data for karst geological research. A comprehensive evaluation showed that after the system's deployment, the timeliness of responses to soil erosion incidents in the monitored area increased by 300%, emergency response efficiency increased by 150%, and related economic losses decreased by approximately 65%, fully demonstrating the practical application value and social benefits of this invention.

[0181] The system's scalable design allows for the flexible addition of new monitoring points and sensor types without changing the core architecture. The cloud platform supports parallel data processing for up to 1,000 monitoring points, with single-point data processing latency controlled to less than 100 milliseconds. The system also includes external data interfaces for integration with other monitoring systems, such as meteorological stations, hydrological stations, and earthquake monitoring networks, enabling data interoperability and comprehensive analysis. This open architecture lays the foundation for future expansion of system functionality and application scenarios, making the system suitable not only for soil and water loss monitoring but also for applications in geological disaster early warning, ecological and environmental protection, and smart city development.

[0182] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for monitoring soil and water loss in karst areas, wherein the method is based on a deployed acoustic sensor array to obtain soil and water loss risk prediction results for a target area, characterized in that: The monitoring method comprises the following steps: Step S1: deploying an acoustic sensor array at key monitoring points in a target area to be monitored, wherein the acoustic sensor array includes a plurality of acoustic sensors for simultaneously collecting acoustic wave signals from the surface and underground environments; The key monitoring points in the target area are divided into surface monitoring points and underground monitoring points. The surface monitoring points include at least karst depressions and karst grooves; the underground monitoring points include karst caves, cave entrances, and underground river entrances. Step S2, collecting the acoustic wave signals detected by the acoustic sensor array and converting the acoustic wave signals into frequency domain feature data; The process of extracting the frequency domain feature data includes: preprocessing the original sound wave signal, wherein the preprocessing includes denoising, filtering and signal enhancement; performing time-frequency analysis on the preprocessed signal to extract time domain features and frequency domain features; Calculate the acoustic features of the sound wave signal, including Mel-frequency cepstral coefficients (MFCCs), spectral flux, spectral centroid, and spectral bandwidth; and extract the statistical features of the sound wave signal, including mean, variance, skewness, kurtosis, and energy distribution. Step S3: inputting the frequency domain feature data into a pre-trained acoustic pattern recognition model, wherein the model is based on a deep learning algorithm and is used to identify acoustic features related to soil erosion; Step S4, based on the output results of the acoustic pattern recognition model, determining the type, degree and development trend of soil erosion in the monitored area, and generating a soil erosion risk prediction result; Calculating soil loss using acoustic signal characteristics monitored by acoustic sensors , whose expression is: ;in, is the amount of soil loss, in tons / hectare; is the soil type related calibration factor; and is the start and end time of monitoring; is the frequency-dependent soil loss transfer function; is the acoustic signal intensity in the time-frequency domain; The acoustic signal strength It is positively correlated with the scouring intensity during the soil erosion process and is calculated by the power spectral density of the short-time Fourier transform: ;in, For the On the time frame The short-time Fourier transform coefficients of the frequency points are Corresponding to the frequency index , Corresponding to the timeframe index ; According to the calculated soil loss With preset threshold Based on the comparison results, the soil and water loss risk prediction score is adjusted. When the risk score increases ,in is the adjustment coefficient, This is the largest amount of soil loss in history.

2. The monitoring method according to claim 1, wherein the acoustic characteristics are calculated using short-time Fourier transform (STFT): ;in, is the original sound wave signal, Indicates in In the analysis window, the position in the window is The sample value of It is the sample index within the window, i.e. the number of FFT points, ranging from 0 to N-1; is the time frame index, indicating the analysis windows; is the frame shift, which indicates the sample number offset between adjacent windows; k is the frequency index, which indicates the frequency point in the short-time Fourier transform (STFT); is a window function; Represents the short-time Fourier transform coefficient of the kth frequency point on the mth time frame; The calculation formula of Mel frequency cepstral coefficient MFCC is: ;in, represents the number of Mel filter banks, represents the cepstral coefficient index; The calculation formula of spectral centroid SC is: ; The calculation formula of spectral flux SF is: ; Sound wave energy distribution The calculation formula is: ;in, represents the frequency band index, and Indicates the The lower and upper frequency indices of the frequency band.

3. The monitoring method according to claim 2, characterized in that: The acoustic pattern recognition model is a hybrid model of a deep convolutional neural network and a long short-term memory network, and its structure includes: an input layer, a convolution layer, a pooling layer, a long short-term memory network layer, a fully connected layer, and a softmax layer; The input layer receives acoustic feature vectors, which are multidimensional arrays containing time domain features, frequency domain features, acoustic features, and statistical features. The time domain features include the time series data of the original signal after preprocessing; the frequency domain features are the spectrum data obtained by short-time Fourier transform; the acoustic features include Mel-frequency cepstral coefficients (MFCC), spectral flux (SF), spectral centroid (SC), and spectral bandwidth (SBW); and the statistical features include the mean, variance, skewness, kurtosis, and energy distribution of each frequency band of the signal. The convolutional layer extracts local patterns of acoustic features, the pooling layer reduces feature dimensions and improves model robustness, the long short-term memory network layer captures the temporal dependencies of acoustic features, and the fully connected layer and softmax layer output the probability distribution of soil erosion type and degree; The mathematical expression of the convolutional layer operation is: ,in: is the output of the previous layer; is the convolution kernel weight; is the bias term; is the rectified linear unit activation function; Represents the convolution operation; Input gate of the long short-term memory network LSTM: ; Forget Gate: , output gate: ; Candidate memory cells: ; Memory unit update: ; Hide status update: ; in: is the input of the current time step, is the hidden state of the previous time step, is the memory cell state at the previous time step, is the weight matrix, is the bias vector, is the hyperbolic tangent activation function, is the Hadamard product, using element-by-element multiplication; Represents the sigmoid activation function; The expression of the output layer is: ;in, is the output of the LSTM layer; is the output layer weight matrix; is the output layer bias vector; is the normalized exponential function.

4. The monitoring method according to claim 3, characterized in that: The acoustic pattern recognition model is trained using the cross entropy loss function: ;in, is the total number of soil erosion types, is the true label, Predict probabilities for the model.

5. The monitoring method according to claim 4, characterized in that: The training steps of the acoustic pattern recognition model are: Step S31: Collect acoustic wave samples generated by different types of soil erosion processes, including acoustic wave samples from surface runoff, rainfall erosion, rockfall, underground dissolution, and underground river flow scenarios; Step S32: labeling the acoustic wave samples and establishing a correspondence between the acoustic features and the type and degree of soil erosion; Step S33: extracting the discriminative features of the sound wave samples using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction techniques; The expression of the principal component analysis PCA is: ;in, is the original acoustic feature matrix, is the principal component loading matrix, is the feature matrix after dimensionality reduction; Calculated from the eigenvector: ;in, is the covariance matrix of the original features, is the eigenvalue diagonal matrix; Calculate the conditional probability in high-dimensional space using t-distributed random neighbor embedding: ; Conditional probability in low-dimensional space: ; The optimization goal is to minimize the KL divergence: ; Step S34, using batch gradient descent method to optimize model parameters, the learning rate is initially set to 0.001, and a learning rate decay strategy is used; Cluster analysis uses the K-means algorithm: ;in, For samples Belong to cluster indicator variables; Cluster the center of is the number of clusters, is the sample size; In step S35, early stopping is used to avoid overfitting, and training is stopped when the performance of the validation set no longer improves.

6. The monitoring method according to claim 5, characterized in that: The method further includes a sound wave propagation model calibration step: establishing a sound wave propagation model for a karst area, taking into account factors affecting topography, rock material, vegetation cover, and meteorological conditions, and conducting tests using known sound sources to obtain the propagation characteristics of sound waves in the karst environment; calibrating the sound wave propagation model parameters based on the test results, and using the calibrated propagation model to improve the accuracy of sound source positioning and signal analysis; The sound wave propagation model is based on the wave equation: ;in, is the sound pressure, is the speed of sound, is the Laplace operator; Under complex terrain conditions, the ray tracing method is used to calculate the sound wave propagation path: ;in, is the sound wave ray path, is the path length parameter, is the refractive index of the acoustic wave, which is related to the sound velocity of the medium. , is the reference sound speed; Sound wave attenuation model considering terrain and atmospheric conditions: ;in, is the sound pressure level at the receiving point; is the sound pressure level of the sound source; is the distance from the sound source to the receiving point; is the atmospheric absorption coefficient; is the attenuation caused by terrain; Attenuation caused by atmospheric conditions.

7. The monitoring method according to claim 6, characterized in that: In step S4, based on the output of the acoustic pattern recognition model, a soil and water loss risk prediction result is generated, and a soil and water loss risk prediction score is generated. Expressed as: ; For the Weight coefficient of soil erosion class; For the The probability of soil erosion occurring For the severity of soil erosion; According to the risk score, the warning levels are divided into four levels: mild, moderate, severe and catastrophic; When the risk score exceeds the warning threshold, the system automatically triggers a warning signal of the corresponding level.

8. A karst area soil and water loss monitoring system, used to implement the monitoring method according to any one of claims 1 to 7, characterized in that: The monitoring system comprises: a signal acquisition unit, a signal processing unit, a pattern recognition unit and a risk assessment unit connected in sequence; The signal acquisition unit is used to receive the sound wave signal collected by the acoustic sensor array and convert the sound wave signal into frequency domain feature data; The acoustic sensor array includes a plurality of acoustic sensors for collecting acoustic wave signals in a target area. The acoustic sensor array is arranged at surface monitoring points and underground monitoring points. The surface monitoring points include at least karst depressions and karst grooves, and the underground monitoring points include karst caves, cave entrances, and underground river entrances. The signal processing unit is used to perform preprocessing, time-frequency analysis and feature extraction on the frequency domain feature data to obtain a multidimensional feature vector including time domain features, frequency domain features, acoustic features and statistical features; The pattern recognition unit includes a pre-trained acoustic pattern recognition model for receiving the multi-dimensional feature vector and outputting recognition results of the type, degree and development trend of soil and water loss; The risk assessment unit is used to calculate the soil and water loss risk score based on the output result of the acoustic pattern recognition model, determine the risk level according to a preset early warning threshold, and generate a soil and water loss risk prediction result.

9. The monitoring system according to claim 8, characterized in that The risk assessment unit further comprises: an early warning issuing module, which is used to automatically generate and issue early warning information according to the risk score, and transmit the soil and water loss risk prediction result to a monitoring center.

Images