A soil and water loss monitoring method and system based on intelligent sensors

Abstract

The application discloses a kind of based on intelligent sensor's water and soil loss monitoring method and system, comprising the following steps: S1, data acquisition and time alignment are carried out;S2, image data is enhanced, edge is identified using Canny operator, regional feature map is extracted through DeepLabV3+ model, and monitoring feature sequence is constructed;S3, abnormal characteristic value is identified by Hampel filter, and mutation fragment is screened;S4, risk score sequence is constructed, trend inflection point is extracted using Kneedle algorithm, and loss trend sequence is generated;S5, the spatial range of water and soil loss is demarcated using Otsu method, and area is estimated;S6, water and soil loss risk grade is judged, and monitoring report is generated.The application fuses multisource perception and visual time series analysis technology, has the advantages of high precision, strong continuity and quantifiable, and improves the automation level of water and soil loss monitoring.

Description

Technical Field

[0001] This invention relates to the field of soil and water loss monitoring technology, and in particular to a method and system for soil and water loss monitoring based on intelligent sensors. Background Technology

[0002] With the increasing demands for ecological protection and natural resource monitoring, soil erosion has become a key factor affecting regional surface stability and sustainable agricultural development. Traditional methods of soil erosion monitoring rely heavily on manual patrols and fixed-point observations, which are not only time-consuming and labor-intensive but also difficult to achieve continuous monitoring and trend assessment over large areas. In recent years, the development of intelligent sensors and image acquisition equipment has provided the foundation for dynamic perception of surface information. Combined with remote images and on-site sensor data, preliminary observations of soil and water change processes can be achieved.

[0003] In existing technologies, some studies have begun to introduce machine learning models for intelligent identification of soil erosion; however, several shortcomings remain. Firstly, most methods rely solely on single-modal data, such as visual segmentation using only surface images, neglecting parameters closely related to soil erosion, such as rainfall intensity, soil moisture, and surface runoff from sensor data. This results in limited analytical dimensions and significant prediction bias. Secondly, soil erosion exhibits significant temporal evolution characteristics, and existing methods lack models for multi-time-step changes, failing to effectively identify trend changes and key abrupt changes during the erosion process, leading to insufficiently fine monitoring granularity.

[0004] Furthermore, while some methods attempt to extract image features using deep learning models, they often neglect precise spatial mapping and area estimation, making it difficult to support risk assessment and management decisions for specific areas. In terms of risk level identification, existing models generally rely on static thresholds or expert experience settings, lacking adaptive, multi-factor fusion intelligent judgment mechanisms, resulting in weak generalization ability and poor interpretability of the output results.

[0005] Therefore, how to provide a method and system for monitoring soil erosion based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for monitoring soil erosion based on intelligent sensors. This invention integrates intelligent sensor sensing technology, computer vision processing technology, and time-series data analysis methods, introduces edge recognition and deep semantic segmentation algorithms, and combines anomaly detection, trend inflection point recognition, and regression fitting operations. It has the advantages of strong monitoring continuity, high spatial positioning accuracy, strong trend recognition capability, and objective and reliable risk assessment results, and can effectively improve the automation level and practical value of soil erosion monitoring.

[0007] A method for monitoring soil erosion based on a smart sensor according to an embodiment of the present invention includes the following steps: S1. Collect smart sensor data and surface image data within a fixed time period at a preset frequency, and perform time alignment operation; S2. After enhancing the surface image data, the Canny operator is used to perform edge recognition. At the same time, the DeepLabV3+ model is used to extract the regional feature map of the surface image data, and the recognition results are combined with the smart sensor data to construct a monitoring feature sequence. S3. Perform sliding window analysis on the monitored feature sequences, identify abnormal feature values ​​through Hampel filter to calculate feature change rate, and use Z-Score normalization method to screen mutation fragments; S4. Based on the characteristic change rate and mutation fragments, and combined with historical monitoring thresholds, a risk score sequence is constructed. The Kneedle algorithm is used to extract the trend inflection point of the risk score sequence, and multi-segment regression fitting is performed to generate a loss trend sequence. S5. Perform time alignment operation between the soil loss trend sequence and the surface image data, perform visual inspection operation on the time alignment result, use the Otsu method to mark the spatial range of soil erosion and estimate the area, and generate image detection results. S6. Construct a fusion feature vector based on the image detection results and the erosion trend sequence, and determine the soil erosion risk level by combining the preset grade intervals, and generate a monitoring report.

[0008] Optionally, the smart sensor data includes soil moisture, rainfall intensity, and surface runoff velocity collected by multiple smart sensors deployed in the monitoring area, and the surface image data represents a set of time-series images of the surface condition covering the monitoring area acquired by an image acquisition device.

[0009] Optionally, S2 specifically includes: S21. Perform brightness normalization and contrast enhancement operations on each frame of the time-aligned surface image data. S22. Based on the enhanced surface image data, the Canny operator is used to perform multi-scale gradient calculation and non-maximum suppression operations to extract continuous edge responses in the surface image and form edge recognition results, specifically including: Gaussian filtering is performed on the enhanced surface image data, and convolution noise reduction is applied to the image pixels using a two-dimensional Gaussian kernel function with a fixed window size. The horizontal and vertical Sobel operators are applied respectively to calculate the horizontal and vertical gradient values ​​of each pixel in the noise reduction result, forming a gradient response matrix. The gradient magnitude and direction of each pixel are calculated based on the gradient response matrix. Non-maximum suppression is performed on all gradient magnitudes. The response intensity in the local neighborhood of each pixel is determined along the gradient direction corresponding to that pixel. The maximum response intensity in the gradient direction is retained, and all positions except the maximum response intensity are set to zero. Perform a dual-threshold edge classification operation on the suppression results. Set a high threshold and a low threshold. Pixels with a response intensity higher than the high threshold are marked as strong edges, and pixels located between the two thresholds and connected to strong edges are marked as weak edges. Pixels with a response intensity lower than the low threshold are removed. The pixel set consisting of strong edges and weak edges is used as the edge recognition result; S23. Perform dilated convolution operation on the enhanced surface image data using the DeepLabV3+ model to extract semantic response features at different spatial scales, and generate a regional feature map consistent with the spatial structure of the surface image data in the decoding path of the model. S24. Perform spatial alignment and feature stitching operations on the region feature map and edge recognition results to form fused image features; S25. Perform joint mapping processing on the fused image features and the corresponding smart sensor data according to the time step to construct the monitoring feature sequence.

[0010] Optionally, S23 specifically includes: S231. Input the enhanced surface image data into the encoding path of the DeepLabV3+ model, and extract the initial features through multi-layer convolution operations; S232. Perform dilated convolution on the enhanced surface image data in the encoding path to construct multi-scale receptive fields with different dilation rates and extract semantic features with spatial context dependencies in the surface image. S233. Concatenate the initial features and semantic features, and input the concatenation result into the decoding path of the DeepLabV3+ model. Perform upsampling and convolution operations in sequence to generate region feature maps.

[0011] Optionally, S3 specifically includes: S31. Construct a fixed-length sliding window for the monitoring feature sequence according to the time step, and extract the feature subsequence of the corresponding time step in each sliding window; S32. Use the Hampel filter to detect outliers at each time step in the feature subsequence, calculate the median and perform deviation analysis to identify outlier feature values. S33. Perform difference calculation between the abnormal feature values ​​and the feature values ​​of adjacent time steps to obtain the feature change rate of the corresponding time step and generate the feature change sequence; S34. Perform standardization on the feature change sequence, and use the Z-Score method to calculate the mean and standard deviation of the feature change rate at each time step to obtain a standard score, specifically including: Perform a mean calculation operation on all feature change rates in the feature change sequence, and calculate the square of the difference between the feature change rate and the mean at each time step; Sum all the squared results, average them, and then perform a square root operation to obtain the standard deviation. The standard score is obtained by subtracting the mean of the feature change rate from the feature change rate corresponding to each time step and dividing by the standard deviation. S35. Based on each standard score and the set mutation threshold, filter the time intervals in which the feature change rate continuously exceeds the mutation threshold and mark them as mutation fragments.

[0012] Optionally, the process of identifying anomalous feature values ​​using a Hampel filter specifically includes: Sort all feature values ​​in the feature subsequence and select the feature value in the middle position as the feature median of the current sliding window. The feature values ​​at each time step in the feature subsequence are compared with the feature median, and the absolute value of the difference is taken to form a set of deviation values. Sort the set of deviation values ​​and select the deviation value in the middle position as the median absolute deviation value. Multiply the median absolute deviation value by the preset anomaly coefficient factor to obtain the anomaly reference value. Use the median of the features minus the anomaly reference value as the lower limit of the interval, and the median of the features plus the anomaly reference value as the upper limit of the interval to construct the anomaly judgment interval. Map the deviation values ​​corresponding to each time step to the anomaly judgment interval, and mark the feature values ​​corresponding to the deviation values ​​that are not in the anomaly judgment interval as abnormal feature values, otherwise mark them as non-abnormal feature values.

[0013] Optionally, S4 specifically includes: S41. Perform time alignment operation on the feature change rate and the mutation fragment, and extract the corresponding feature change rate in each mutation fragment according to the time step to form the mutation subsequence; S42. Perform a moving weighted average processing on each variation subsequence, set weight coefficients based on existing historical monitoring thresholds, and generate a risk score sequence based on the weighted average results of each time step, specifically including: Set a fixed-length sliding window for the changing subsequence and perform a sliding traversal operation on the changing subsequence step by step; Extract the feature change rate within the current sliding window, and set weight coefficients based on the difference between each feature change rate and the existing historical monitoring thresholds; The risk score value at the current sliding window center time step is obtained by performing a weighted summation operation on the feature change rate within the sliding window according to the corresponding weight coefficients. Arrange all risk scores in chronological order to construct a risk score sequence; S43. Perform first-order difference operation on the risk score sequence to construct the rate of change sequence, and use the Kneedle algorithm to identify trend inflection points; S44. Divide the risk score sequence into multiple score segments based on the trend inflection point, and perform least squares fitting operation to extract the fitting slope and trend direction of each score segment to construct the churn trend sequence.

[0014] Optionally, S43 specifically includes: S431. Perform a first-order difference operation on the risk score sequence at each time step, subtract the risk score values ​​of adjacent time steps to form a rate of change sequence; S432. Using the Kneedle algorithm, a risk scoring curve is constructed based on the rate of change sequence. The curvature value of each time step is calculated in the risk scoring curve with a fixed sliding interval to construct a curvature sequence. S433. Perform a maximum value search operation on the curvature sequence and extract the maximum points in the curvature sequence as candidate inflection points; S434. Perform a change direction judgment operation on all candidate inflection points, and retain candidate inflection points whose risk score value changes in direction in adjacent time steps greater than a preset direction threshold as trend inflection points.

[0015] Optionally, S5 specifically includes: S51. Perform time alignment operation on the loss trend sequence and the surface image data to establish a one-to-one mapping relationship between each time step in the loss trend sequence and the corresponding surface image data. S52. Perform visual detection operations frame by frame on the time-aligned surface image data, and determine the detection interest area in the surface image based on the corresponding loss trend information in the loss trend sequence. S53. Within the detection area of ​​interest, perform grayscale processing on the surface image data, and use the Otsu method to perform threshold segmentation on the grayscale result to generate a binary segmentation result for the soil erosion area, specifically including: Extract the corresponding surface image portion from the region of interest and convert it into a grayscale image. Count the number of times each brightness value appears in the grayscale image to form grayscale distribution data. Based on the grayscale distribution data, the Otsu method is used to calculate the brightness threshold, specifically including: Based on the grayscale distribution data, each brightness value is selected as a candidate boundary value in turn; Based on the candidate threshold value, the pixels in the grayscale image are divided into two categories: one is the set of pixels whose brightness value is not greater than the candidate threshold value, and the other is the set of pixels whose brightness value is greater than the candidate threshold value. Calculate the pixel proportion and corresponding average brightness value of the two pixel sets in the grayscale image, and calculate the inter-class variance between the two pixel sets based on the pixel proportion and average brightness value; The candidate boundary value with the largest inter-class variance is selected as the brightness boundary value; All pixels in the grayscale image are classified according to the brightness threshold. Pixels with brightness values ​​higher than the threshold are set as foreground and pixels with brightness values ​​lower than the threshold are set as background, generating a black and white binary image as the binary segmentation result. S54. Perform connected component analysis on the binary segmentation results to extract continuous pixel regions and determine the spatial range of soil erosion. S55. Perform statistical calculations on the number of pixels within the spatial range to obtain the corresponding soil erosion area, and use the spatial range and the soil erosion area together to form the image detection result.

[0016] A soil erosion monitoring system based on intelligent sensors according to an embodiment of the present invention includes: The data acquisition module is used to acquire smart sensor data and surface image data within a fixed time period at a preset frequency, and to perform time alignment operations. The feature recognition module is used to enhance the surface image data, perform edge recognition operation using the Canny operator, extract regional feature maps of the surface image data through the DeepLabV3+ model, and construct monitoring feature sequences by combining the recognition results with smart sensor data. The mutation screening module is used to perform sliding window analysis on the monitored feature sequences, identify abnormal feature values ​​through Hampel filter to calculate the feature change rate, and screen mutation fragments using the Z-Score normalization method. The trend analysis module is used to construct a risk score sequence based on the characteristic change rate and mutation fragments, combined with historical monitoring thresholds. It uses the Kneedle algorithm to extract the trend inflection point of the risk score sequence and performs multi-segment regression fitting to generate a loss trend sequence. The visual detection module is used to perform time alignment operations between the soil loss trend sequence and surface image data, perform visual detection operations on the time alignment results, use the Otsu method to identify the spatial range of soil erosion and estimate the area, and generate image detection results. The risk assessment module is used to construct a fused feature vector based on image detection results and erosion trend sequence, and to determine the risk level of soil erosion by combining it with preset grade intervals, and generate a monitoring report.

[0017] The beneficial effects of this invention are: First, this invention constructs a monitoring system covering spatial features and temporal evolution by jointly modeling multi-source sensor data and surface image data. This effectively solves the limitations of traditional soil erosion monitoring methods that rely solely on image information and ignore temporal dynamic changes and quantitative sensor data. It achieves synchronous acquisition and alignment processing of key elements such as soil moisture, rainfall intensity, and surface runoff velocity, thereby improving the comprehensiveness and real-time nature of the data foundation.

[0018] Secondly, addressing the issue of complex visual features and significant multi-scale changes in soil erosion areas in images, this invention introduces the Canny edge detection and DeepLabV3+ semantic segmentation model, fusing image boundary and regional semantic information to improve the accuracy of spatial range extraction and area estimation of soil erosion areas. Furthermore, through the trend analysis path constructed using the Hampel filter, Z-Score method, and Kneedle algorithm, it is able to identify abrupt changes and trend inflection points during the monitoring process, and accurately extract trend features by combining a multi-segment fitting strategy.

[0019] Finally, in the risk level assessment stage, this invention constructs a fusion feature vector to couple and analyze image spatial information with trend evolution indicators, and judges the risk level according to preset classification rules to generate a structured monitoring report, thereby improving the interpretability, traceability and practical guidance significance of soil erosion monitoring results. 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 of a soil and water loss monitoring method based on intelligent sensors proposed in this invention; Figure 2 This is a core flowchart of a soil erosion monitoring method based on intelligent sensors proposed in this invention. Figure 3 This is a modular structure diagram of a soil and water loss monitoring system based on intelligent sensors 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 1-2 A method for monitoring soil erosion based on intelligent sensors includes the following steps: S1. Collect smart sensor data and surface image data within a fixed time period at a preset frequency, and perform time alignment operation; S2. After enhancing the surface image data, the Canny operator is used to perform edge recognition. At the same time, the DeepLabV3+ model is used to extract the regional feature map of the surface image data, and the recognition results are combined with the smart sensor data to construct a monitoring feature sequence. S3. Perform sliding window analysis on the monitored feature sequences, identify abnormal feature values ​​through Hampel filter to calculate feature change rate, and use Z-Score normalization method to screen mutation fragments; S4. Based on the characteristic change rate and mutation fragments, and combined with historical monitoring thresholds, a risk score sequence is constructed. The Kneedle algorithm is used to extract the trend inflection point of the risk score sequence, and multi-segment regression fitting is performed to generate a loss trend sequence. S5. Perform time alignment operation between the soil loss trend sequence and the surface image data, perform visual inspection operation on the time alignment result, use the Otsu method to mark the spatial range of soil erosion and estimate the area, and generate image detection results. S6. Construct a fusion feature vector based on the image detection results and the erosion trend sequence, and determine the soil erosion risk level by combining the preset grade intervals, and generate a monitoring report.

[0023] In this embodiment, the smart sensor data includes soil moisture, rainfall intensity, and surface runoff velocity collected by multiple smart sensors deployed in the monitoring area, and the surface image data represents a set of time-series images of the surface condition covering the monitoring area acquired by an image acquisition device.

[0024] In this embodiment, S2 specifically includes: S21. Perform brightness normalization and contrast enhancement operations on each frame of the time-aligned surface image data. S22. Based on the enhanced surface image data, the Canny operator is used to perform multi-scale gradient calculation and non-maximum suppression operations to extract continuous edge responses in the surface image and form edge recognition results, specifically including: Gaussian filtering is performed on the enhanced surface image data, and convolution noise reduction is applied to the image pixels using a two-dimensional Gaussian kernel function with a fixed window size. The horizontal and vertical Sobel operators are applied respectively to calculate the horizontal and vertical gradient values ​​of each pixel in the noise reduction result, forming a gradient response matrix. The gradient magnitude and direction of each pixel are calculated based on the gradient response matrix. Non-maximum suppression is performed on all gradient magnitudes. The response intensity in the local neighborhood of each pixel is determined along the gradient direction corresponding to that pixel. The maximum response intensity in the gradient direction is retained, and all positions except the maximum response intensity are set to zero. Perform a dual-threshold edge classification operation on the suppression results. Set a high threshold and a low threshold. Pixels with a response intensity higher than the high threshold are marked as strong edges, and pixels located between the two thresholds and connected to strong edges are marked as weak edges. Pixels with a response intensity lower than the low threshold are removed. The pixel set consisting of strong edges and weak edges is used as the edge recognition result; S23. Perform dilated convolution operation on the enhanced surface image data using the DeepLabV3+ model to extract semantic response features at different spatial scales, and generate a regional feature map consistent with the spatial structure of the surface image data in the decoding path of the model. S24. Perform spatial alignment and feature stitching operations on the region feature map and edge recognition results to form fused image features; S25. Perform joint mapping processing on the fused image features and the corresponding smart sensor data according to the time step to construct the monitoring feature sequence.

[0025] In this embodiment, S23 specifically includes: S231. Input the enhanced surface image data into the encoding path of the DeepLabV3+ model, and extract the initial features through multi-layer convolution operations; S232. Perform dilated convolution on the enhanced surface image data in the encoding path to construct multi-scale receptive fields with different dilation rates and extract semantic features with spatial context dependencies in the surface image. S233. Concatenate the initial features and semantic features, and input the concatenation result into the decoding path of the DeepLabV3+ model. Perform upsampling and convolution operations in sequence to generate region feature maps.

[0026] In this embodiment, S233 specifically includes: S2331. Perform feature concatenation operation on the channel dimension to construct a fused feature representation by combining the initial features and semantic features; S2332. Input the fused feature representation into the decoding path of the DeepLabV3+ model, and perform an upsampling operation on the fused feature representation using bilinear interpolation to restore the spatial size to one-quarter of the initial resolution of the encoding path; S2333. Perform two convolution operations on the upsampling results in sequence, and perform ReLU activation and batch normalization operations on the convolution results to extract semantic consistency and boundary continuity features respectively. S2334. Perform channel compression on the extraction results and upsample the compressed results to restore the spatial size to the same original resolution as the surface image data, generating high-resolution features as regional feature maps.

[0027] In this embodiment, S3 specifically includes: S31. Construct a fixed-length sliding window for the monitoring feature sequence according to the time step, and extract the feature subsequence of the corresponding time step in each sliding window; S32. Use the Hampel filter to detect outliers at each time step in the feature subsequence, calculate the median and perform deviation analysis to identify outlier feature values. S33. Perform difference calculation between the abnormal feature values ​​and the feature values ​​of adjacent time steps to obtain the feature change rate of the corresponding time step and generate the feature change sequence; S34. Perform standardization on the feature change sequence, and use the Z-Score method to calculate the mean and standard deviation of the feature change rate at each time step to obtain a standard score, specifically including: Perform a mean calculation operation on all feature change rates in the feature change sequence, and calculate the square of the difference between the feature change rate and the mean at each time step; Sum all the squared results, average them, and then perform a square root operation to obtain the standard deviation. The standard score is obtained by subtracting the mean of the feature change rate from the feature change rate corresponding to each time step and dividing by the standard deviation. S35. Based on each standard score and the set mutation threshold, filter the time intervals in which the feature change rate continuously exceeds the mutation threshold and mark them as mutation fragments.

[0028] In this embodiment, the process of identifying abnormal feature values ​​using a Hampel filter specifically includes: Sort all feature values ​​in the feature subsequence and select the feature value in the middle position as the feature median of the current sliding window. The feature values ​​at each time step in the feature subsequence are compared with the feature median, and the absolute value of the difference is taken to form a set of deviation values. Sort the set of deviation values ​​and select the deviation value in the middle position as the median absolute deviation value. Multiply the median absolute deviation value by the preset anomaly coefficient factor to obtain the anomaly reference value. Use the median of the features minus the anomaly reference value as the lower limit of the interval, and the median of the features plus the anomaly reference value as the upper limit of the interval to construct the anomaly judgment interval. Map the deviation values ​​corresponding to each time step to the anomaly judgment interval, and mark the feature values ​​corresponding to the deviation values ​​that are not in the anomaly judgment interval as abnormal feature values, otherwise mark them as non-abnormal feature values.

[0029] In this embodiment, S4 specifically includes: S41. Perform time alignment operation on the feature change rate and the mutation fragment, and extract the corresponding feature change rate in each mutation fragment according to the time step to form the mutation subsequence; S42. Perform a moving weighted average processing on each variation subsequence, set weight coefficients based on existing historical monitoring thresholds, and generate a risk score sequence based on the weighted average results of each time step, specifically including: Set a fixed-length sliding window for the changing subsequence and perform a sliding traversal operation on the changing subsequence step by step; Extract the feature change rate within the current sliding window, and set weight coefficients based on the difference between each feature change rate and the existing historical monitoring thresholds; The risk score value at the current sliding window center time step is obtained by performing a weighted summation operation on the feature change rate within the sliding window according to the corresponding weight coefficients. Arrange all risk scores in chronological order to construct a risk score sequence; S43. Perform first-order difference operation on the risk score sequence to construct the rate of change sequence, and use the Kneedle algorithm to identify trend inflection points; S44. Divide the risk score sequence into multiple score segments based on the trend inflection point, and perform least squares fitting operation to extract the fitting slope and trend direction of each score segment to construct the churn trend sequence.

[0030] In this embodiment, S43 specifically includes: S431. Perform a first-order difference operation on the risk score sequence at each time step, subtract the risk score values ​​of adjacent time steps to form a rate of change sequence; S432. Using the Kneedle algorithm, a risk scoring curve is constructed based on the rate of change sequence. The curvature value of each time step is calculated in the risk scoring curve with a fixed sliding interval to construct a curvature sequence. S433. Perform a maximum value search operation on the curvature sequence and extract the maximum points in the curvature sequence as candidate inflection points; S434. Perform a change direction judgment operation on all candidate inflection points, and retain candidate inflection points whose risk score value changes in direction in adjacent time steps greater than a preset direction threshold as trend inflection points.

[0031] In this embodiment, the process of constructing a loss trend sequence based on trend inflection points specifically includes: Based on the trend inflection point, the risk scoring sequence is divided into multiple scoring segments according to the time step, and each scoring segment is continuous on the time axis and has no overlap. Extract the time index and corresponding risk score value within each scoring segment to construct a sample point set, wherein the sample point is composed of a combination of time step and corresponding risk score value; Based on the sample point set, calculate the average time index and the average risk score, and construct the center coordinate point of the scoring segment; Using the least squares method, at each time step of the scoring segment, the difference between the time index and the average time index is calculated, and the difference between the corresponding risk score and the average risk score is also calculated. The two types of difference results are multiplied together, and all product results are summed to obtain the deviation sum value. Calculate the sum of squares of the differences between all time indices and the average time index, and divide the sum of the deviations by the sum of squares to generate the fit slope for the scoring segment; Determine the trend direction of the time step. If the fitting slope is positive, it is marked as an upward trend; if the fitting slope is negative, it is marked as a downward trend; if the fitting slope approaches zero, it is marked as a stable trend. Arrange the risk scores, fitting slopes, and trend directions of all scoring segments in chronological order to construct a churn trend sequence.

[0032] In this embodiment, S5 specifically includes: S51. Perform time alignment operation on the loss trend sequence and the surface image data to establish a one-to-one mapping relationship between each time step in the loss trend sequence and the corresponding surface image data. S52. Perform visual detection operations frame by frame on the time-aligned surface image data, and determine the detection interest area in the surface image based on the corresponding loss trend information in the loss trend sequence. S53. Within the detection area of ​​interest, perform grayscale processing on the surface image data, and use the Otsu method to perform threshold segmentation on the grayscale result to generate a binary segmentation result for the soil erosion area, specifically including: Extract the corresponding surface image portion from the region of interest and convert it into a grayscale image. Count the number of times each brightness value appears in the grayscale image to form grayscale distribution data. Based on the grayscale distribution data, the Otsu method is used to calculate the brightness threshold, specifically including: Based on the grayscale distribution data, each brightness value is selected as a candidate boundary value in turn; Based on the candidate threshold value, the pixels in the grayscale image are divided into two categories: one is the set of pixels whose brightness value is not greater than the candidate threshold value, and the other is the set of pixels whose brightness value is greater than the candidate threshold value. Calculate the pixel proportion and corresponding average brightness value of the two pixel sets in the grayscale image, and calculate the inter-class variance between the two pixel sets based on the pixel proportion and average brightness value; The candidate boundary value with the largest inter-class variance is selected as the brightness boundary value; All pixels in the grayscale image are classified according to the brightness threshold. Pixels with brightness values ​​higher than the threshold are set as foreground and pixels with brightness values ​​lower than the threshold are set as background, generating a black and white binary image as the binary segmentation result. S54. Perform connected component analysis on the binary segmentation results to extract continuous pixel regions and determine the spatial range of soil erosion. S55. Perform statistical calculations on the number of pixels within the spatial range to obtain the corresponding soil erosion area, and use the spatial range and the soil erosion area together to form the image detection result.

[0033] In this embodiment, S6 specifically includes: S61. Perform feature alignment operation on the image detection results and the loss trend sequence according to time steps to construct combined data; S62. Combine the spatial range and soil erosion area in the image detection results with the risk score, fitting slope and trend direction in the erosion trend sequence to generate a fusion feature vector for each time step. S63. Perform a risk level judgment operation on the fused feature vector and determine the corresponding risk level label based on the preset risk level interval; S64. Based on each fused feature vector and its corresponding risk level label, generate a soil and water loss monitoring report.

[0034] refer to Figure 3 A soil erosion monitoring system based on intelligent sensors, comprising: The data acquisition module is used to acquire smart sensor data and surface image data within a fixed time period at a preset frequency, and to perform time alignment operations. The feature recognition module is used to enhance the surface image data, perform edge recognition operation using the Canny operator, extract regional feature maps of the surface image data through the DeepLabV3+ model, and construct monitoring feature sequences by combining the recognition results with smart sensor data. The mutation screening module is used to perform sliding window analysis on the monitored feature sequences, identify abnormal feature values ​​through Hampel filter to calculate the feature change rate, and screen mutation fragments using the Z-Score normalization method. The trend analysis module is used to construct a risk score sequence based on the characteristic change rate and mutation fragments, combined with historical monitoring thresholds. It uses the Kneedle algorithm to extract the trend inflection point of the risk score sequence and performs multi-segment regression fitting to generate a loss trend sequence. The visual detection module is used to perform time alignment operations between the soil loss trend sequence and surface image data, perform visual detection operations on the time alignment results, use the Otsu method to identify the spatial range of soil erosion and estimate the area, and generate image detection results. The risk assessment module is used to construct a fused feature vector based on image detection results and erosion trend sequence, and to determine the risk level of soil erosion by combining it with preset grade intervals, and generate a monitoring report.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical hilly terrain area, with the goal of conducting all-weather intelligent monitoring and risk assessment of soil erosion in a local watershed. This area has significant topographic relief, loose topsoil, and low vegetation cover, making it prone to significant surface erosion and sediment loss under heavy or continuous rainfall conditions. Traditional monitoring methods rely on manual patrols and fixed-point measurements, which are not only inefficient and time-consuming but also difficult to obtain continuous and effective monitoring data under adverse weather conditions. Therefore, an intelligent method with high timeliness, high accuracy, and automatic judgment capabilities is urgently needed.

[0036] In the actual deployment of this invention, multiple intelligent sensor nodes were deployed within the monitoring area to collect basic data such as soil moisture, rainfall intensity, and surface runoff velocity. Simultaneously, multiple image acquisition devices were deployed on different slopes to periodically record surface image sequences. All devices achieved time synchronization and data aggregation through an edge gateway, forming a structured raw monitoring data set. The system collects full data every 10 minutes, with an image resolution of 1920×1080 and a sensor sampling accuracy of ±0.01 units.

[0037] In the data processing stage, the surface image is first normalized and enhanced, edge recognition is performed using the Canny operator, and regional feature map extraction is completed using the DeepLabV3+ model, thereby accurately identifying surface erosion traces and exposed areas.

[0038] The system then fuses image features with sensor data to construct a multi-source monitoring feature sequence. The system tracks state changes within each time period using a sliding window mechanism, identifies outliers using a Hampel filter, filters out periods of rapid and continuous change using the Z-Score method, and finally uses the Kneedle algorithm to identify trend inflection points, ultimately constructing a loss trend sequence.

[0039] During the risk assessment phase, the system automatically fuses trend data with image area estimation results, determines the risk level based on the set risk level range, and generates a structured report.

[0040] During a continuous 7-day observation period, the system processed over 5,000 frames of image data and over 100,000 sensor data entries. To verify the advantages of this invention over traditional manual inspection methods in terms of response speed, spatial recognition accuracy, and trend judgment accuracy, a comparative analysis was conducted between the results of manual recording and the output of the monitoring system of this invention.

[0041] The specific comparison indicators and results are shown in the table below: Table 1. Performance Comparison Data Between the Monitoring System of the Present Invention and Manual Inspection Method

[0042] As can be seen from the data in Table 1, under conditions of high rainfall and significantly increased surface runoff velocity, the present invention can effectively identify large areas of surface soil erosion. The spatial range estimation results are generally higher than those of manual inspection records. In terms of trend identification, the system can accurately locate multiple trend inflection points and identify risk trends of continuous growth or decline.

[0043] Compared to manual methods, the response delay of this invention is significantly reduced, with an average response time controlled within 10 minutes. Furthermore, the risk levels assessed are generally higher than those determined by human experience, demonstrating stronger risk foresight and prevention value. In particular, on the 3rd, 5th, and 7th days of continuously high rainfall intensity, the system assessed the risk as "high-risk," which was highly consistent with the subsequent actual localized erosion damage, verifying the effectiveness and stability of this invention in practical applications.

[0044] In summary, this invention significantly outperforms traditional methods in multi-source data fusion, continuous monitoring, trend identification, and risk assessment, demonstrating high practical application value and potential for intelligent management. If needed, it can be further integrated with edge computing deployment schemes to achieve low-latency, localized monitoring, providing efficient intelligent decision support for surface ecological protection and agricultural water and soil management.

[0045] 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 method for monitoring soil erosion based on intelligent sensors, characterized in that, Includes the following steps: S1. Collect smart sensor data and surface image data within a fixed time period at a preset frequency, and perform time alignment operation; S2. After enhancing the surface image data, the Canny operator is used to perform edge recognition. At the same time, the DeepLabV3+ model is used to extract the regional feature map of the surface image data, and the recognition results are combined with the smart sensor data to construct a monitoring feature sequence. S3. Perform sliding window analysis on the monitored feature sequences, identify abnormal feature values ​​through Hampel filter to calculate feature change rate, and use Z-Score normalization method to screen mutation fragments; S4. Based on the characteristic change rate and mutation fragments, and combined with historical monitoring thresholds, a risk score sequence is constructed. The Kneedle algorithm is used to extract the trend inflection points of the risk score sequence, and multi-segment regression fitting is performed to generate a churn trend sequence, specifically including: S41. Perform time alignment operation on the feature change rate and the mutation fragment, and extract the corresponding feature change rate in each mutation fragment according to the time step to form the mutation subsequence; S42. Perform a moving weighted average processing on each variation subsequence, set weight coefficients based on existing historical monitoring thresholds, and generate a risk score sequence based on the weighted average results of each time step, specifically including: Set a fixed-length sliding window for the changing subsequence and perform a sliding traversal operation on the changing subsequence step by step; Extract the feature change rate within the current sliding window, and set weight coefficients based on the difference between each feature change rate and the existing historical monitoring thresholds; The risk score value at the current sliding window center time step is obtained by performing a weighted summation operation on the feature change rate within the sliding window according to the corresponding weight coefficients. Arrange all risk scores in chronological order to construct a risk score sequence; S43. Perform first-order difference operation on the risk score sequence to construct the rate of change sequence, and use the Kneedle algorithm to identify trend inflection points; S44. Divide the risk score sequence into multiple score segments based on the trend inflection point, and perform least squares fitting operation to extract the fitting slope and trend direction of each score segment to construct the churn trend sequence. S5. Perform time alignment operation between the soil loss trend sequence and the surface image data, perform visual inspection operation on the time alignment result, use the Otsu method to mark the spatial range of soil erosion and estimate the area, and generate image detection results. S6. Construct a fusion feature vector based on the image detection results and the erosion trend sequence, and determine the soil erosion risk level by combining the preset grade intervals, and generate a monitoring report.

2. The method for monitoring soil erosion based on intelligent sensors according to claim 1, characterized in that, The smart sensor data includes soil moisture, rainfall intensity, and surface runoff velocity collected by multiple smart sensors deployed in the monitoring area, and the surface image data represents a time-series image set of the surface condition covering the monitoring area, acquired by an image acquisition device.

3. The method for monitoring soil erosion based on intelligent sensors according to claim 1, characterized in that, S2 specifically includes: S21. Perform brightness normalization and contrast enhancement operations on each frame of the time-aligned surface image data. S22. Based on the enhanced surface image data, the Canny operator is used to perform multi-scale gradient calculation and non-maximum suppression operations to extract continuous edge responses in the surface image and form edge recognition results, specifically including: Gaussian filtering is performed on the enhanced surface image data, and convolution noise reduction is applied to the image pixels using a two-dimensional Gaussian kernel function with a fixed window size. The horizontal and vertical Sobel operators are applied respectively to calculate the horizontal and vertical gradient values ​​of each pixel in the noise reduction result, forming a gradient response matrix. The gradient magnitude and direction of each pixel are calculated based on the gradient response matrix. Non-maximum suppression is performed on all gradient magnitudes. The response intensity in the local neighborhood of each pixel is determined along the gradient direction corresponding to that pixel. The maximum response intensity in the gradient direction is retained, and all positions except the maximum response intensity are set to zero. Perform a dual-threshold edge classification operation on the suppression results. Set a high threshold and a low threshold. Pixels with a response intensity higher than the high threshold are marked as strong edges, and pixels located between the two thresholds and connected to strong edges are marked as weak edges. Pixels with a response intensity lower than the low threshold are removed. The pixel set consisting of strong edges and weak edges is used as the edge recognition result; S23. Perform dilated convolution operation on the enhanced surface image data using the DeepLabV3+ model to extract semantic response features at different spatial scales, and generate a regional feature map consistent with the spatial structure of the surface image data in the decoding path of the model. S24. Perform spatial alignment and feature stitching operations on the region feature map and edge recognition results to form fused image features; S25. Perform joint mapping processing on the fused image features and the corresponding smart sensor data according to the time step to construct the monitoring feature sequence.

4. The method for monitoring soil erosion based on intelligent sensors according to claim 3, characterized in that, S23 specifically includes: S231. Input the enhanced surface image data into the encoding path of the DeepLabV3+ model, and extract the initial features through multi-layer convolution operations; S232. Perform dilated convolution on the enhanced surface image data in the encoding path to construct multi-scale receptive fields with different dilation rates and extract semantic features with spatial context dependencies in the surface image. S233. Concatenate the initial features and semantic features, and input the concatenation result into the decoding path of the DeepLabV3+ model. Perform upsampling and convolution operations in sequence to generate region feature maps.

5. The method for monitoring soil erosion based on intelligent sensors according to claim 1, characterized in that, S3 specifically includes: S31. Construct a fixed-length sliding window for the monitoring feature sequence according to the time step, and extract the feature subsequence of the corresponding time step in each sliding window; S32. Use the Hampel filter to detect outliers at each time step in the feature subsequence, calculate the median and perform deviation analysis to identify outlier feature values. S33. Perform difference calculation between the abnormal feature values ​​and the feature values ​​of adjacent time steps to obtain the feature change rate of the corresponding time step and generate the feature change sequence; S34. Perform standardization on the feature change sequence, and use the Z-Score method to calculate the mean and standard deviation of the feature change rate at each time step to obtain a standard score, specifically including: Perform a mean calculation operation on all feature change rates in the feature change sequence, and calculate the square of the difference between the feature change rate and the mean at each time step; Sum all the squared results, average them, and then perform a square root operation to obtain the standard deviation. The standard score is obtained by subtracting the mean of the feature change rate from the feature change rate corresponding to each time step and dividing by the standard deviation. S35. Based on each standard score and the set mutation threshold, filter the time intervals in which the feature change rate continuously exceeds the mutation threshold and mark them as mutation fragments.

6. The method for monitoring soil erosion based on intelligent sensors according to claim 5, characterized in that, The process of identifying abnormal feature values ​​using the Hampel filter specifically includes: Sort all feature values ​​in the feature subsequence and select the feature value in the middle position as the feature median of the current sliding window. The feature values ​​at each time step in the feature subsequence are compared with the feature median, and the absolute value of the difference is taken to form a set of deviation values. Sort the set of deviation values ​​and select the deviation value in the middle position as the median absolute deviation value. Multiply the median absolute deviation value by the preset anomaly coefficient factor to obtain the anomaly reference value. Use the median of the features minus the anomaly reference value as the lower limit of the interval, and the median of the features plus the anomaly reference value as the upper limit of the interval to construct the anomaly judgment interval. Map the deviation values ​​corresponding to each time step to the anomaly judgment interval, and mark the feature values ​​corresponding to the deviation values ​​that are not in the anomaly judgment interval as abnormal feature values, otherwise mark them as non-abnormal feature values.

7. The method for monitoring soil erosion based on intelligent sensors according to claim 1, characterized in that, Specifically, S43 includes: S431. Perform a first-order difference operation on the risk score sequence at each time step, subtract the risk score values ​​of adjacent time steps to form a rate of change sequence; S432. Using the Kneedle algorithm, a risk scoring curve is constructed based on the rate of change sequence. The curvature value of each time step is calculated in the risk scoring curve with a fixed sliding interval to construct a curvature sequence. S433. Perform a maximum value search operation on the curvature sequence and extract the maximum points in the curvature sequence as candidate inflection points; S434. Perform a change direction judgment operation on all candidate inflection points, and retain candidate inflection points whose risk score value changes in direction in adjacent time steps greater than a preset direction threshold as trend inflection points.

8. The method for monitoring soil erosion based on intelligent sensors according to claim 1, characterized in that, S5 specifically includes: S51. Perform time alignment operation on the loss trend sequence and the surface image data to establish a one-to-one mapping relationship between each time step in the loss trend sequence and the corresponding surface image data. S52. Perform visual detection operations frame by frame on the time-aligned surface image data, and determine the detection interest area in the surface image based on the corresponding loss trend information in the loss trend sequence. S53. Within the detection area of ​​interest, perform grayscale processing on the surface image data, and use the Otsu method to perform threshold segmentation on the grayscale result to generate a binary segmentation result for the soil erosion area, specifically including: Extract the corresponding surface image portion from the region of interest and convert it into a grayscale image. Count the number of times each brightness value appears in the grayscale image to form grayscale distribution data. Based on the grayscale distribution data, the Otsu method is used to calculate the brightness threshold, specifically including: Based on the grayscale distribution data, each brightness value is selected as a candidate boundary value in turn; Based on the candidate threshold value, the pixels in the grayscale image are divided into two categories: one is the set of pixels whose brightness value is not greater than the candidate threshold value, and the other is the set of pixels whose brightness value is greater than the candidate threshold value. Calculate the pixel proportion and corresponding average brightness value of the two pixel sets in the grayscale image, and calculate the inter-class variance between the two pixel sets based on the pixel proportion and average brightness value; The candidate boundary value with the largest inter-class variance is selected as the brightness boundary value; All pixels in the grayscale image are classified according to the brightness threshold. Pixels with brightness values ​​higher than the threshold are set as foreground and pixels with brightness values ​​lower than the threshold are set as background, generating a black and white binary image as the binary segmentation result. S54. Perform connected component analysis on the binary segmentation results to extract continuous pixel regions and determine the spatial range of soil erosion. S55. Perform statistical calculations on the number of pixels within the spatial range to obtain the corresponding soil erosion area, and use the spatial range and the soil erosion area together to form the image detection result.

9. A soil erosion monitoring system based on intelligent sensors, comprising the soil erosion monitoring method based on intelligent sensors as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire smart sensor data and surface image data within a fixed time period at a preset frequency, and to perform time alignment operations. The feature recognition module is used to enhance the surface image data, perform edge recognition operation using the Canny operator, extract regional feature maps of the surface image data through the DeepLabV3+ model, and construct monitoring feature sequences by combining the recognition results with smart sensor data. The mutation screening module is used to perform sliding window analysis on the monitored feature sequences, identify abnormal feature values ​​through Hampel filter to calculate feature change rate, and screen mutation fragments using Z-Score normalization method; The trend analysis module is used to construct a risk score sequence based on the characteristic change rate and mutation fragments, combined with historical monitoring thresholds. It uses the Kneedle algorithm to extract the trend inflection point of the risk score sequence and performs multi-segment regression fitting to generate a loss trend sequence. The visual detection module is used to perform time alignment operations between the soil loss trend sequence and surface image data, perform visual detection operations on the time alignment results, use the Otsu method to identify the spatial range of soil erosion and estimate the area, and generate image detection results. The risk assessment module is used to construct a fused feature vector based on image detection results and erosion trend sequence, and to determine the risk level of soil erosion by combining it with preset grade intervals, and generate a monitoring report.

Images