Artificial Intelligence-Based Soil Erosion Risk Monitoring Method

Abstract

This invention discloses an artificial intelligence-based method for monitoring soil erosion risk. To address the problems of untimely identification of soil erosion risk, difficulty in quantifying disturbance information, and insufficient early warning accuracy within the responsibility area of ​​production and construction projects, this invention acquires high-resolution remote sensing images and divides the responsibility areas into prevention and control units. It utilizes a disturbance identification model to identify pixel-level engineering disturbances and extract disturbance features such as disturbance area ratio, disturbance morphology, and spatial relationship between disturbances and drainage channels. Combining rainfall, topography, soil, vegetation, and soil and water conservation measures data, it calculates key factors such as rainfall erosivity, vegetation cover management, slope length and gradient, soil erodibility, and measure factors. These factors are then substituted into a modified general soil loss equation to obtain baseline soil erosion. The baseline soil erosion, disturbance features, key factors, and rainfall forecast are input into a deep residual network model to obtain target soil erosion risk indicators. Finally, based on risk classification thresholds, the risk level of each prevention and control responsibility unit is output, achieving the technical effect of unit-scale soil erosion risk monitoring and early warning.

Description

Technical Field

[0001] This invention relates to the field of soil and water conservation monitoring and early warning technology, and in particular to an artificial intelligence-based method for monitoring soil erosion risk. Background Technology

[0002] Soil erosion is a significant factor restricting my country's ecological environment security and sustainable land use. Human disturbance caused by production and construction activities has become a major source of severe localized soil erosion. To address this issue, soil and water conservation management departments have established systems for the preparation and approval of soil and water conservation plans for production and construction projects, in-process and post-process monitoring, acceptance, and the delineation of "prevention and control responsibility areas." These systems monitor the implementation of soil and water conservation measures and the risk of soil erosion within the project's disturbance area. To support these management needs, existing technologies commonly employ the Universal / Modified Universal Soil Loss Equation (USLE / RUSLE), Geographic Information Systems (GIS), and remote sensing monitoring to quantitatively evaluate the intensity of soil erosion at the watershed or administrative unit scale, and based on this, conduct research on soil and water conservation zoning and measure deployment. In recent years, some remote sensing monitoring and drone patrol technologies for soil and water conservation in production and construction projects have also emerged. These technologies are used to identify newly added bare land, spoil heaps, construction access roads, and other project-disturbed areas, and to indirectly reflect the effectiveness of soil and water conservation measures.

[0003] However, existing technologies still have the following shortcomings:

[0004] 1. Insufficient spatial information representation of engineering disturbance and soil and water conservation measures: Existing soil and water conservation monitoring focuses mainly on macro indicators such as disturbance area and vegetation coverage. The spatial morphological characteristics of the engineering disturbance area (such as boundary shape, length-to-width ratio, compactness, and degree of fragmentation) and its spatial relationship with soil and water conservation drainage systems such as drainage ditches and intercepting ditches are not described in detail enough. It is difficult to accurately reflect key soil and water conservation processes such as changes in runoff paths and enhanced slope water collection. It also provides insufficient support for the identification of soil and water conservation risk points and sensitive areas within the "prevention and control responsibility area".

[0005] 2. The disturbance identification results are not closely coupled with rainfall forecasts and soil and water conservation mechanisms: Existing soil and water conservation assessments based on RUSLE mostly use historical rainfall or multi-year averages to estimate rainfall erosivity, and rarely combine future rainfall forecast information to conduct dynamic analysis of short-term heavy rainfall processes. They also lack a mechanism to comprehensively couple disturbance patterns, vegetation restoration status, implementation of soil and water conservation measures, and rainfall temporal characteristics, making it difficult to reflect the changes in soil erosion risk in the "prevention and control responsibility area" under specific future rainfall scenarios in a timely manner.

[0006] 3. Limited integration of mechanistic models and artificial intelligence methods in soil and water conservation: Traditional quantitative evaluation of soil and water conservation relies on empirical mechanistic models such as RUSLE. The determination of parameters is limited by empirical formulas and regional experience, resulting in limited accuracy in complex engineering disturbance scenarios. On the other hand, some soil erosion prediction work that introduces machine learning or deep learning often weakens or ignores the constraints of soil and water conservation mechanisms, which can easily lead to overfitting and poor interpretability of prediction results. This makes it difficult to meet the requirements of reliability, interpretability and traceability of results in soil and water conservation supervision.

[0007] Therefore, it is necessary to propose a soil and water loss risk monitoring method in the context of soil and water conservation operations. This method should be able to identify engineering disturbance characteristics with high precision at the scale of "prevention and control responsibility scope", fully couple disturbance identification results, rainfall forecast information and soil and water conservation mechanism models, and integrate the advantages of mechanism models and artificial intelligence models, so as to improve the scientificity and precision of soil and water conservation supervision. Summary of the Invention

[0008] One objective of this invention is to propose an artificial intelligence-based method for monitoring soil erosion risks. Addressing the problems of existing technologies that rely on manual inspections or single-mechanism models, leading to difficulties in accurately identifying engineering disturbances, incomplete acquisition of influencing factors, and delayed and inaccurate risk warnings, this invention proposes a method that divides prevention and control responsibility units according to prevention and control zones. It uses high-resolution remote sensing images to identify pixel-level disturbances and extract disturbance area proportions, disturbance morphology, and spatial relationship features between disturbances and drainage channels. Combining rainfall forecasts, topography, soil, vegetation, and soil and water conservation measures, it calculates rainfall erosivity, slope length and gradient, soil erodibility, vegetation cover management, and other factors. These are then substituted into a modified general soil loss equation to obtain baseline soil erosion. The baseline amount, along with disturbance characteristics, key factors, and rainfall time series, is input into a deep residual network to obtain target risk indicators. Finally, it performs graded warnings based on a threshold set. This invention offers the technical benefits of refined monitoring at the responsibility unit scale, forward-looking prediction and risk grading output, and improved monitoring and warning accuracy.

[0009] This invention provides an artificial intelligence-based method for monitoring soil erosion risk, comprising:

[0010] S1. Acquire high-resolution remote sensing images covering the prevention and control responsibility area of ​​the production and construction project, perform preprocessing, and divide the project into multiple prevention and control responsibility units according to the prevention and control zones of the project's soil and water conservation plan. Crop the preprocessed images to obtain remote sensing image sub-data corresponding to each prevention and control responsibility unit. S2. Input the remote sensing image sub-data of each prevention and control responsibility unit into a disturbance recognition model pre-trained based on disturbance samples to obtain pixel-level disturbance recognition results. Based on this, determine the disturbance area ratio, disturbance morphology characteristics, and spatial relationship characteristics between disturbance and drainage channels for each prevention and control responsibility unit, collectively referred to as disturbance features. S3. Based on the disturbance features of each prevention and control responsibility unit, and combined with data on rainfall, topography, soil, vegetation, and soil and water conservation measures corresponding to each prevention and control responsibility unit, calculate the rainfall erosivity factor and vegetation density of each prevention and control responsibility unit. Coverage management factors, slope length and gradient factors, soil erodibility factors, and soil and water conservation measures factors are collectively referred to as key factors; S4, Substitute the key factors of each prevention and control responsibility unit into the modified general soil loss equation model to obtain the baseline soil and water loss of each prevention and control responsibility unit; S5, Based on the baseline soil and water loss, disturbance characteristics, key factors, and rainfall forecast data of each prevention and control responsibility unit, construct input feature data and input it into the deep residual network model pre-trained based on historical soil and water loss monitoring samples to obtain the target soil and water loss risk index of each prevention and control responsibility unit; S6, Based on the preset set of soil and water loss risk classification thresholds, classify the target soil and water loss risk index of each prevention and control responsibility unit to obtain the soil and water loss risk level, which is output as the soil and water loss risk monitoring result and early warning result of the prevention and control responsibility area.

[0011] Optionally, S1 includes:

[0012] A high-resolution remote sensing image with a spatial resolution better than 2 meters and containing multiple spectral bands is acquired, covering the entire prevention and control responsibility area of ​​the production and construction project in a planar position. The high-resolution remote sensing image is preprocessed, including at least radiometric correction and geometric correction, and preferably also radiometric calibration and atmospheric correction. The preprocessed high-resolution remote sensing image is then unified to a preset map projection coordinate system and a unified spatial resolution to obtain a corrected remote sensing image. Based on the boundary vector data of the prevention and control responsibility area of ​​the production and construction project and the prevention and control zoning vector data of the approved soil and water conservation plan for the production and construction project, the prevention and control responsibility area is divided into multiple prevention and control responsibility units. The corrected remote sensing image is spatially cropped based on the boundary of each prevention and control responsibility unit, so that each prevention and control responsibility unit corresponds to a continuous image region in the corrected remote sensing image, and each continuous image region is used as a remote sensing image sub-data corresponding to each prevention and control responsibility unit.

[0013] Optionally, S2 includes:

[0014] Within each prevention and control responsibility unit, the processed remote sensing image sub-data is input into a disturbance recognition model pre-trained based on disturbance samples. This disturbance recognition model is a semantic segmentation neural network employing an encoder and decoder structure. It extracts spatial features based on a convolutional neural network and an attention mechanism, outputting pixel-level classification results consistent with the spatial resolution of the remote sensing image sub-data. Each pixel in the pixel-level classification result is labeled as either an engineering disturbance area pixel or a non-engineering disturbance area pixel, forming the pixel-level disturbance recognition result for that prevention and control responsibility unit. Within each prevention and control responsibility unit, the number of pixels labeled as engineering disturbance areas is counted based on the pixel-level disturbance recognition results. The total area of ​​the engineering disturbance area is calculated by combining this with the area of ​​individual pixels in the remote sensing image sub-data. The ratio of the total area of ​​the engineering disturbance area to the total area of ​​the prevention and control responsibility unit is calculated to obtain the disturbance area ratio for that unit. Within each prevention and control responsibility unit, the outer boundary of the engineering disturbance area is extracted based on the pixel-level disturbance recognition results, and the engineering disturbance area is calculated based on the geometry of the outer boundary. The disturbance morphology of the prevention and control responsibility unit is obtained by considering at least one of the following: perimeter, area, major axis direction, aspect ratio, compactness, and fragmentation degree. Multispectral aerial photographs of the UAV covering the prevention and control responsibility unit are acquired, including at least red and near-infrared bands. Terrain data is generated based on the multispectral aerial photographs, and these photographs are input into a pre-trained drainage channel recognition model to automatically identify drainage channels, thus obtaining spatial location data of the drainage channels. Within each prevention and control responsibility unit, the spatial location data of the drainage channels is spatially superimposed with the pixel-level disturbance recognition results of that unit. At least one of the following is calculated: minimum distance from the boundary of the engineering disturbance area to the nearest drainage channel, intersection length of the engineering disturbance area and the drainage channel, and distribution length of the engineering disturbance area along the drainage channel direction, to obtain the spatial relationship characteristics between the disturbance and the drainage channel of the prevention and control responsibility unit. Finally, the disturbance area ratio, disturbance morphology characteristics, and spatial relationship characteristics between the disturbance and the drainage channel of each prevention and control responsibility unit are output.

[0015] Optionally, S3 includes:

[0016] Within each prevention and control responsibility unit, using the proportion of disturbed area, morphological characteristics of the disturbance, and spatial relationship characteristics between the disturbance and drainage channels as inputs, historical rainfall monitoring data, rainfall forecast data, topographic data, soil type data, information on the implementation of soil and water conservation measures, and remote sensing vegetation index data corresponding to that responsibility unit are acquired. The topographic data and remote sensing vegetation index data are generated from UAV multispectral aerial photography, and the information on the implementation of soil and water conservation measures is obtained from UAV multispectral aerial photography through deep learning recognition. Based on the historical rainfall monitoring data and the rainfall forecast data, combined with the proportion of disturbed area and morphological characteristics... Based on the spatial relationship characteristics between disturbances and drainage channels, the rainfall kinetic energy, maximum rainfall intensity, and rainfall concentration of the prevention and control responsibility unit within the target forecast period are calculated. The rainfall kinetic energy, maximum rainfall intensity, and rainfall concentration are then converted into a rainfall erosivity factor for the prevention and control responsibility unit according to a preset rainfall erosivity factor calculation formula. Based on the pixel-level disturbance identification results, the engineering disturbance area within the prevention and control responsibility unit is determined. The vegetation cover of the disturbance area is calculated within the engineering disturbance area using remote sensing vegetation index data. Finally, the vegetation cover of the disturbance area is converted into a vegetation cover factor for the prevention and control responsibility unit according to a preset vegetation cover management factor grading rule. The management factors are as follows: A digital elevation model of the prevention and control responsibility unit is constructed based on the topographic data. The slope and runoff length of each pixel within the prevention and control responsibility unit are calculated. The slope and runoff length are then calculated according to a preset slope length and slope factor calculation formula to obtain the slope length and slope factor of the prevention and control responsibility unit. Based on the soil type data, a soil attribute database is queried to obtain parameters such as soil texture, organic matter content, and structural gradation corresponding to different soil types. These parameters are then calculated according to a preset soil erodibility factor calculation formula to obtain the soil erodibility factor of the prevention and control responsibility unit. Based on the soil and water conservation... The information on the implementation of measures identifies the types, spatial locations, and coverage areas of soil and water conservation measures such as slope drainage ditches, intercepting ditches, retaining walls, temporary covering measures, and vegetation slope protection within the prevention and control responsibility unit. It also calculates the proportion of the engineering disturbance area covered by the soil and water conservation measures based on the proportion of the disturbance area covered by the measures, and converts the proportion of the engineering disturbance area covered by the measures and the type of soil and water conservation measures into soil and water conservation measure factors for the prevention and control responsibility unit according to the preset soil and water conservation measure factor calculation rules. This results in the output of rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erodibility factor, and soil and water conservation measure factor for each prevention and control responsibility unit.

[0017] Optionally, S4 includes:

[0018] Within each prevention and control responsibility unit, the baseline soil and water loss is calculated using the following inputs: rainfall erosivity factor, slope length and gradient factor, soil erodibility factor, vegetation cover management factor, soil and water conservation measures factor, and the area of ​​the engineering disturbance zone within the unit. The modified general soil loss equation model is used to calculate the baseline soil and water loss of the unit. The engineering disturbance zone area is the total area of ​​the engineering disturbance zone obtained from the pixel-level disturbance identification results. The modified general soil loss equation model is based on the general soil loss equation A = R × K × LS × C × P, with rainfall erosivity factor as R, soil erodibility factor as K, slope length and gradient factor as LS, vegetation cover management factor as C, and soil and water conservation measures factor as P. First, the baseline soil and water loss per unit area within the prevention and control responsibility unit during the target forecast period is calculated based on the general soil loss equation. Then, the baseline soil and water loss per unit area is multiplied by the area of ​​the engineering disturbance zone within the unit to obtain the baseline soil and water loss of the unit. The baseline soil and water loss of each prevention and control responsibility unit is then output.

[0019] Optionally, S5 includes:

[0020] Within each prevention and control responsibility unit, based on baseline soil erosion, disturbed area proportion, disturbance morphology, spatial relationship between disturbance and drainage channels, as well as rainfall erosivity factors, vegetation cover management factors, slope length and gradient factors, soil erosibility factors, and soil and water conservation measures factors, rainfall forecast data corresponding to that responsibility unit is obtained. The rainfall forecast data undergoes time series processing and numerical normalization to calculate the rainfall temporal characteristics of that responsibility unit within the forecast period. Within each responsibility unit, baseline soil erosion, disturbed area proportion, disturbance morphology, spatial relationship between disturbance and drainage channels, rainfall erosivity factors, vegetation cover management factors, slope length and gradient factors, soil erosibility factors, soil and water conservation measures factors, and rainfall temporal characteristics are concatenated according to a preset feature order to construct the input feature data for that responsibility unit. The input feature data of each responsibility unit is then input into a deep residual network model pre-trained based on historical soil erosion monitoring samples. The deep residual network model includes a mechanism branch and a data branch, wherein... The mechanism branch takes baseline soil erosion, rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erosibility factor, and soil and water conservation measure factor as mechanism constraint inputs. It performs a nonlinear transformation on the baseline soil erosion through a fully connected neural network with multi-layer residual connections to obtain the soil erosion estimation result based on mechanism constraints. The data branch takes baseline soil erosion, disturbance area ratio, disturbance morphology characteristics, spatial relationship characteristics between disturbance and drainage channels, rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erosibility factor, soil and water conservation measure factor, and rainfall time series characteristics as inputs. It learns the deviation between the baseline soil erosion and the actual soil erosion through a neural network with multi-layer residual connections to obtain the soil erosion deviation estimation result. In the fusion layer of the deep residual network model, the soil erosion estimation result based on mechanism constraints output by the mechanism branch and the soil erosion deviation estimation result output by the data branch are residually fused to obtain the target soil erosion risk index of the prevention and control responsibility unit, and output the target soil erosion risk index of each prevention and control responsibility unit.

[0021] Optionally, S6 includes:

[0022] Within each prevention and control responsibility unit, using the target soil erosion risk indicator as input, a set of soil erosion risk classification thresholds, pre-determined based on historical soil erosion monitoring samples, is obtained. This set includes multiple threshold intervals arranged in ascending order of risk. Within each responsibility unit, the target soil erosion risk indicator is compared with the set of thresholds to determine the threshold interval into which the target indicator falls. Target soil erosion risk indicators falling into different threshold intervals are then assigned to at least three soil erosion risk levels: low risk, relatively low risk, medium risk, relatively high risk, and high risk, respectively, thus obtaining the soil erosion risk level for each responsibility unit. At the scope of responsibility, the soil erosion risk levels of each responsibility unit are summarized. Based on these levels, corresponding monitoring and early warning indicators are assigned to each responsibility unit. The soil erosion risk levels of each responsibility unit, along with their corresponding monitoring and early warning indicators, are output as the soil erosion risk monitoring and early warning results for the scope of responsibility.

[0023] Optionally, the terrain data generated based on UAV multispectral aerial photography includes:

[0024] Aerial triangulation and orthorectification are performed on multispectral aerial photographs taken by UAVs to generate orthophotos, and digital surface model (DSM) and / or digital elevation model (DEM) are generated based on photogrammetric 3D reconstruction, wherein the terrain data includes the DEM;

[0025] The drainage channel identification model is a semantic segmentation neural network and / or an object detection neural network, which outputs the pixel-level identification results of the drainage channel and vectorizes the pixel-level identification results to obtain the spatial location data of the drainage channel.

[0026] The beneficial effects of this invention are:

[0027] 1. Achieve refined and automated risk monitoring and early warning at the responsibility unit scale: By cropping high-resolution remote sensing images to the prevention and control responsibility unit, and performing pixel-level engineering disturbance identification and disturbance area ratio statistics, the monitoring results can be directly correlated with the prevention and control zones and responsibility units of the soil and water conservation plan, reducing reliance on manual inspections and improving the timeliness and feasibility of monitoring.

[0028] 2. Improve the accuracy and robustness of soil erosion risk assessment: At the mechanism level, the baseline soil erosion is calculated based on the modified general soil loss equation. At the data level, information such as disturbance morphology, spatial relationship between disturbance and drainage channels, and rainfall time series are introduced. The risk index is output by learning the deviation between the baseline and the actual situation through a deep residual network, thereby taking into account both mechanism constraints and data-driven correction, and reducing the systematic error caused by a single model.

[0029] 3. Enhance the foresight and interpretability of early warning: Combine rainfall forecast data to calculate rainfall erosivity and rainfall time sequence characteristics, and output graded early warning results based on the risk classification threshold set. This enables risk assessment to not only reflect the current disturbance and the implementation status of measures, but also to provide early warning of soil and water loss risk during the target forecast period. Furthermore, key factors (R, K, LS, C, P) and disturbance characteristics provide interpretable evidence for the source of risk. Attached Figure Description

[0030] 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:

[0031] Figure 1 This is a flowchart of an artificial intelligence-based method for monitoring soil erosion risk proposed in this invention. Detailed Implementation

[0032] 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.

[0033] refer to Figure 1 An artificial intelligence-based method for monitoring soil erosion risk includes:

[0034] S1. Acquire high-resolution remote sensing images covering the prevention and control responsibility area of ​​the production and construction project, perform preprocessing, and divide the project into multiple prevention and control responsibility units according to the prevention and control zones of the project's soil and water conservation plan. Crop the preprocessed images to obtain remote sensing image sub-data corresponding to each prevention and control responsibility unit. S2. Input the remote sensing image sub-data of each prevention and control responsibility unit into a disturbance recognition model pre-trained based on disturbance samples to obtain pixel-level disturbance recognition results. Based on this, determine the disturbance area ratio, disturbance morphology characteristics, and spatial relationship characteristics between disturbance and drainage channels for each prevention and control responsibility unit, collectively referred to as disturbance features. S3. Based on the disturbance features of each prevention and control responsibility unit, and combined with data on rainfall, topography, soil, vegetation, and soil and water conservation measures corresponding to each prevention and control responsibility unit, calculate the rainfall erosivity factor and vegetation density of each prevention and control responsibility unit. Coverage management factors, slope length and gradient factors, soil erodibility factors, and soil and water conservation measures factors are collectively referred to as key factors; S4, Substitute the key factors of each prevention and control responsibility unit into the modified general soil loss equation model to obtain the baseline soil and water loss of each prevention and control responsibility unit; S5, Based on the baseline soil and water loss, disturbance characteristics, key factors, and rainfall forecast data of each prevention and control responsibility unit, construct input feature data and input it into the deep residual network model pre-trained based on historical soil and water loss monitoring samples to obtain the target soil and water loss risk index of each prevention and control responsibility unit; S6, Based on the preset set of soil and water loss risk classification thresholds, classify the target soil and water loss risk index of each prevention and control responsibility unit to obtain the soil and water loss risk level, which is output as the soil and water loss risk monitoring result and early warning result of the prevention and control responsibility area.

[0035] In this specific embodiment, S1 includes:

[0036] The remote sensing data uses high-resolution remote sensing images from the same imaging phase. The high-resolution remote sensing images have a spatial resolution of 1 m and include four spectral bands: blue, green, red, and near-infrared. They are provided in the form of orthophoto products. The high-resolution remote sensing images completely cover the prevention and control responsibility area of ​​the production and construction project in a planar position. The prevention and control responsibility area is represented by the boundary vector data of the prevention and control responsibility area as a closed polygon surface element with unified coordinate reference information. The image coverage is determined by spatially superimposing and verifying the image's outer rectangle with the boundary vector data of the prevention and control responsibility area. The image files are stored in GeoTIFF format, and the metadata records radiometric calibration parameters such as imaging time, solar altitude angle, sensor gain and offset, as well as RPC geometric parameters.

[0037] The high-resolution remote sensing image is preprocessed to obtain a corrected remote sensing image. The preprocessing includes radiometric calibration, atmospheric correction, radiometric consistency correction, and geometric correction. Radiometric calibration converts the digital quantization values ​​of image pixels into sensor-received radiance values ​​for each band. The conversion relationship is as follows:

[0038] ;

[0039] in Indicates spectral band The corresponding sensor received radiance, This indicates the spectral band identifier, and in this embodiment, it is one of the four bands: blue light, green light, red light, and near-infrared light. Indicates band The gain coefficient is given by the image metadata. Indicates the corresponding pixel in the band The digital quantization value below, Indicates band The bias coefficient is given by the image metadata;

[0040] Atmospheric correction employs a correction process based on radiative transfer, which combines radiance with metadata such as imaging time and solar elevation angle to calculate surface reflectance in order to eliminate the effects of atmospheric scattering and absorption.

[0041] Radiometric consistency correction addresses situations where there is strip noise and inter-scene brightness differences. It performs strip noise suppression and histogram matching on each band to achieve radiometric consistency within the same imaging product.

[0042] Geometric correction uses RPC geometric parameters and digital elevation model (DEM) for orthorectification to eliminate geometric distortion caused by terrain undulations, and controls the correction result to a planar positioning error of no more than one pixel.

[0043] After completing radiometric and geometric corrections, the corrected remote sensing images are unified to a preset map projection coordinate system and a unified spatial resolution. The map projection coordinate system adopts the UTM projection coordinate system based on the CGCS2000 standard. The UTM projection zone number is determined based on the longitude of the geometric center point of the boundary vector data of the prevention and control responsibility area. The image pixel size is unified to [missing information]. When the original pixel size is inconsistent with the uniform pixel size, bilinear resampling is used to achieve resolution uniformity. After resampling, the band order, row and column number of the four spectral bands are kept consistent, and the georeferenced information is completely written into the GeoTIFF header file to form the corrected remote sensing image.

[0044] Prevention and control responsibility units are divided based on the boundary vector data of the prevention and control responsibility scope of the production and construction project and the prevention and control zoning vector data of the approved water and soil conservation plan of the production and construction project. The prevention and control zoning vector data is a set of surface features with a prevention and control zoning identifier field. The prevention and control responsibility units are obtained by performing spatial clipping of the prevention and control zoning vector data and the boundary vector data of the prevention and control responsibility scope. The clipped surface features are decomposed according to connectivity to ensure that each prevention and control responsibility unit corresponds to a continuous surface area. Then, a unique prevention and control responsibility unit identifier is written to each continuous surface area and the prevention and control responsibility unit boundary is generated.

[0045] Based on the boundaries of each prevention and control responsibility unit, the corrected remote sensing image is masked and cropped. During cropping, the spatial range of the output sub-image is determined by the minimum outer rectangle of the prevention and control responsibility unit boundary, and the prevention and control responsibility unit boundary is used as the effective area mask. Pixels outside the boundary are set to NoData, while the four-band reflectance values ​​and georeferenced information of pixels inside the boundary are retained. This results in remote sensing image sub-data that corresponds one-to-one with each prevention and control responsibility unit and is associated with the prevention and control responsibility unit identifier.

[0046] In this specific embodiment, S2 includes:

[0047] For each prevention and control responsibility unit's remote sensing image sub-data, the remote sensing image sub-data is first constructed by inputting the remote sensing image sub-data. The remote sensing image sub-data maintains the band order of the four spectral bands as blue light, green light, red light, and near-infrared light, and NoData pixels are filled with all-zero vectors to avoid abnormal responses to convolution features.

[0048] Linear normalization is performed on each spectral band separately, scaling the pixel values ​​of each band to [value missing]. The scaling factor should be consistent across the same prevention and control responsibility unit;

[0049] The normalized four-band images are then processed according to... The pixel size is used for sliding window slicing, with a sliding window step size of 256 pixels, and edge insufficient processing is applied. The regions are filled with mirror images to ensure that the input block size is consistent, thus obtaining an image block sequence for inference of the perturbation recognition model;

[0050] The perturbation recognition model employs a semantic segmentation neural network with an encoder-decoder structure. The encoder uses a ResNet-34 backbone network and inserts a CBAM attention module at the output of each residual stage to simultaneously model channel attention and spatial attention. The decoder uses a four-level upsampling structure and achieves detail recovery through skip connections of features at the same scale as the encoder. Upsampling uses bilinear interpolation and is concatenated after each upsampling stage. Convolutional layers, batch normalization layers, and ReLU activation layers; network input is four channels. The image block is output as a two-channel pixel-level category probability map, and the pixel probability of the engineering disturbance area and the pixel probability of the non-engineering disturbance area are obtained by Softmax.

[0051] The disturbance identification model is trained based on disturbance samples. The disturbance samples consist of four-band remote sensing image blocks of the same resolution and pixel-level labeled masks. The labeled categories include pixels in engineering disturbance areas and pixels in non-engineering disturbance areas. The pixels in engineering disturbance areas are defined as the set of pixels in construction bare land, spoil dumping areas, construction access roads and temporary land occupation areas.

[0052] During training, the batch size was set to 8, the number of training epochs was set to 100, the optimizer was Adam, and the learning rate was set to... The loss function is a weighted sum of cross-entropy loss and Dice loss, with the weight coefficients set to 0.5 and 0.5 respectively.

[0053] During the inference process, the output pixel probability map of the engineering disturbance area is binarized with a threshold of 0.5 to obtain the pixel-level disturbance recognition result. Opening and closing operations are then performed on the binarized result to eliminate isolated noise and holes. The structuring element uses... A square kernel is used with the number of iterations set to 1. Then, engineering perturbation connected components with a connected component area of ​​less than 50 pixels are deleted to suppress false detections.

[0054] Within each prevention and control responsibility unit, based on the pixel-level disturbance identification results, the number of pixels in the engineering disturbance area is counted and the total area of ​​the engineering disturbance area is calculated. Then, the disturbance area ratio is calculated, and the disturbance area ratio is determined according to the formula... Calculation, where This indicates the proportion of the disturbed area within the prevention and control responsibility unit. This represents the total area of ​​the engineering disturbance zone within the prevention and control responsibility unit, and is obtained by multiplying the number of pixels in the engineering disturbance zone by the area of ​​a single pixel in the remote sensing image sub-data. This represents the total area of ​​the prevention and control responsibility unit and is determined by the area attribute of the boundary vector of the prevention and control responsibility unit;

[0055] In the disturbance morphology feature extraction, the binary pixel mask of the engineering disturbance area is vectorized to generate a polygon of the engineering disturbance area, and the outer boundary is used as the disturbance boundary. The disturbance perimeter is taken as the length value of the outer boundary, and the disturbance area is taken as the area value of the polygon. The direction of the disturbance major axis is obtained by performing principal component analysis on the set of pixel coordinates of the disturbance area to obtain the first principal component direction angle, and the north direction is taken as zero degrees and the clockwise direction is taken as the positive direction. The aspect ratio of the disturbance is obtained by calculating the minimum bounding rectangle of the disturbance area polygon and taking the ratio of its long side length to the short side length. The degree of disturbance fragmentation is obtained by analyzing the connected components to count the number of connected components of the engineering disturbance in the prevention and control responsibility unit and normalizing it with the total area of ​​the prevention and control responsibility unit to obtain the number of connected components per unit area, thereby obtaining the disturbance morphology feature set of the prevention and control responsibility unit.

[0056] Within each prevention and control responsibility unit, in order to obtain spatial location data of the drainage channel, a UAV multispectral aerial photograph covering the prevention and control responsibility unit is acquired, including four bands: blue light, green light, red light, and near-infrared light. The ground resolution of the UAV multispectral aerial photograph is set to 0.1 m, and the exposure position and attitude are recorded by airborne RTK and 6 ground image control points are set up to achieve absolute orientation. Based on aerial triangulation and orthorectification, an UAV orthophoto is generated and is consistent with the map projection coordinate system in step 51.

[0057] Orthophotos of drones Pixel segments are input into the drainage channel recognition model to output pixel-level recognition results of the drainage channels. The drainage channel recognition model uses a DeepLabV3+ semantic segmentation network with a ResNet-50 backbone network and a dilated convolution output stride of 16. The training batch size is set to 6, the number of training epochs is set to 80, the optimizer is SGD with a momentum of 0.9, and the learning rate is set to... The loss function used is cross-entropy loss;

[0058] The pixel-level recognition results of the drainage channel are refined and extracted into a skeleton, and vectorized by fitting a piecewise polyline to obtain the spatial location data of the drainage channel and represent it as a polyline feature.

[0059] In the extraction of spatial relationship features between disturbance and drainage channel, the polygonal features of the engineering disturbance area and the polyline features of the drainage channel are spatially superimposed and analyzed. The minimum distance from the boundary of the engineering disturbance area to the nearest drainage channel is obtained by calculating the shortest Euclidean distance between the polyline of the disturbance boundary and the polyline of the drainage channel. The intersection length of the engineering disturbance area and the drainage channel is obtained by calculating the sum of the lengths of the line segments of the drainage channel polyline that fall inside the polygon of the engineering disturbance area. The distribution length of the engineering disturbance area along the direction of the drainage channel is obtained by projecting the intersection line segments inside the polygon of the engineering disturbance area according to the mileage direction of the drainage channel polyline and taking the length of the projection range. The disturbance area ratio, disturbance morphology features and spatial relationship features between disturbance and drainage channel of each prevention and control responsibility unit are output.

[0060] In this specific embodiment, S3 includes:

[0061] For each prevention and control responsibility unit, the disturbance area ratio, disturbance morphology characteristics, and spatial relationship characteristics between the disturbance and the drainage channel are used as disturbance-side inputs. Historical rainfall monitoring data, rainfall forecast data, topographic data, soil type data, water and soil conservation measures implementation information, and remote sensing vegetation index data consistent with the spatial range of the prevention and control responsibility unit are acquired. Among them, the historical rainfall monitoring data comes from the 10-minute rainfall sequence of automatic rain gauges within the prevention and control responsibility area and covers the continuous 72 hours before the current time. The rainfall forecast data comes from the hourly rainfall of the numerical weather prediction grid product and covers the continuous 24 hours after the current time. The two are unified into the 10-minute target forecast period rainfall sequence through time alignment and linear interpolation, and outliers are removed by missing measurement markers.

[0062] The terrain data is generated from UAV multispectral aerial photographs covering the prevention and control responsibility unit through aerial triangulation, orthorectification, and photogrammetric 3D reconstruction to create a digital elevation model (DEM). The model is then resampled to the same spatial resolution as the remote sensing image sub-data of the prevention and control responsibility unit. Based on the DEM, the slope and runoff length of each pixel are calculated using the D8 single-flow direction algorithm. The area-weighted average of the slope and runoff length within the prevention and control responsibility unit is then used as the input for calculating the slope length and slope factor.

[0063] Soil type data is represented by soil type surface features and rasterized to the same spatial resolution as the prevention and control responsibility unit. Based on the soil type code, the corresponding soil texture, organic matter content and structural gradation parameters are obtained by querying the soil attribute database. The soil erodibility factor of the prevention and control responsibility unit is directly determined according to the soil erodibility factor value that corresponds one-to-one with the soil type in the soil attribute database.

[0064] Remote sensing vegetation index data was calculated from UAV multispectral aerial imagery to obtain NDVI gratings, which were then registered with the prevention and control responsibility units. Based on the pixel-level disturbance identification results, the engineering disturbance area was determined, and the proportion of pixels with NDVI greater than or equal to 0.30 within the engineering disturbance area was counted as the vegetation coverage of the disturbance area. According to the vegetation coverage classification rules of the disturbance area, the vegetation coverage was converted into vegetation coverage management factor values ​​from low to high. The four levels are used to determine the vegetation cover management factors for the prevention and control responsibility unit;

[0065] Information on the implementation of soil and water conservation measures is obtained from UAV multispectral aerial photography through a soil and water conservation measure identification model. This model employs a Mask R-CNN instance segmentation network with a ResNet-50 backbone network, combined with an FPN feature pyramid. The category set is fixed at five categories: slope drainage ditches, intercepting ditches, retaining walls, temporary tarpaulin measures, and vegetation slope protection. The input is... The model generates orthophoto slices of drone images at pixel level and outputs instance masks and category labels for various measures. The batch size for model training is set to 4, the number of training epochs is set to 60, the optimizer uses SGD with momentum set to 0.9, and the learning rate is set to... During the inference phase, instances with a confidence level less than 0.5 are discarded, and the remaining instances are vectorized into masked polygons to obtain measure polygons. Subsequently, these polygons are spatially superimposed with the polygons representing the engineering disturbance area to calculate the proportion of the engineering disturbance area covered by the measures. The water and soil conservation measure factor is determined according to the rules based on the measure type. Specifically, when the proportion of the engineering disturbance area covered by the measures is less than 0.30, the water and soil conservation measure factor is 0.8; when the proportion is between 0.30 and 0.70, the factor is 0.6; and when it is greater than or equal to 0.70, the factor is 0.4. Furthermore, when a retaining wall instance is identified, the water and soil conservation measure factor is multiplied by 0.9 to reflect the contribution of the retaining measures.

[0066] The calculation of the rainfall erosivity factor is based on the rainfall sequence of the target forecast period. Rainfall intensity is calculated every 10 minutes, and the rainfall kinetic energy for the target forecast period is accumulated according to the USLE kinetic energy function. Simultaneously, the maximum rainfall intensity is calculated within a 30-minute sliding window and used as the maximum rainfall intensity index. Rainfall concentration is determined by statistically ranking the rainfall intensities within the target forecast period. The proportion of time-step rainfall to total rainfall is obtained, and the rainfall kinetic energy is amplified and corrected by combining the fragmentation degree in the disturbance morphology characteristics and the minimum distance from the boundary of the engineering disturbance area to the nearest drainage channel in the spatial relationship characteristics between the disturbance and the drainage channel. Specifically, when the minimum distance is less than or equal to 10 m, the rainfall kinetic energy is multiplied by 1.2, and when the fragmentation degree is greater than or equal to 3 connected domains per hectare, the rainfall kinetic energy is multiplied by 1.1. Then, according to the preset rainfall erosivity factor calculation formula, the rainfall kinetic energy, maximum rainfall intensity, and rainfall concentration degree are converted into a rainfall erosivity factor that satisfies the following:

[0067] ;

[0068] in This represents the rainfall erosivity factor for the prevention and control responsibility unit during the target forecast period, and the unit is... This represents the rainfall kinetic energy for the target forecast period after confluence amplification correction, and the unit is... This represents the maximum rainfall intensity within a 30-minute sliding window during the target forecast period, and the unit is... The dimensionless concentration coefficient, determined by the concentration level, is expressed by mapping the concentration level from low to high as... The value is obtained from five levels;

[0069] The slope length and gradient factor are calculated using the RUSLE slope length and gradient factor algorithm based on the area-weighted average of the slope and runoff length, with the slope length index set to 0.4 and the gradient term index set to 1.3, to obtain the slope length and gradient factor of the prevention and control responsibility unit. The output of the rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erodibility factor, and soil and water conservation measures factor of the prevention and control responsibility unit are also obtained.

[0070] In this specific embodiment, S4 includes:

[0071] For each prevention and control responsibility unit, the rainfall erosivity factor, soil erodibility factor, slope length and slope factor, vegetation cover management factor and soil and water conservation measure factor are read. The total area of ​​the engineering disturbance area obtained by statistical analysis based on the pixel-level disturbance identification results is read as the engineering disturbance area. The engineering disturbance area is measured in square meters and converted to hectares by dividing by 10,000 to ensure consistency with the measurement system of rainfall erosivity factor and soil erodibility factor.

[0072] Before numerical calculations, the validity of each factor was verified. Specifically, the rainfall erosivity factor was verified to be a non-negative real number, the soil erodibility factor was verified to be a non-negative real number, the slope length and gradient factor was verified to be a real number greater than zero, and the vegetation cover management factor and the soil and water conservation measures factor were verified to be intervals. The real numbers within the range are truncated and corrected according to their physical meaning to avoid meaningless results in subsequent multiplication operations;

[0073] Subsequently, the baseline soil and water loss of the prevention and control responsibility unit is calculated within the target forecast period according to the modified general soil loss equation model. The modified general soil loss equation model is based on the general soil loss equation and multiplies the baseline soil and water loss per unit area by the area of ​​the engineering disturbance zone to obtain the baseline soil and water loss at the area scale. The calculation formula is as follows:

[0074] ;

[0075] in This represents the baseline soil erosion volume of the prevention and control responsibility unit during the target forecast period, with units of... This indicates the rainfall erosivity factor for the responsible prevention and control unit during the target forecast period. This indicates the soil erodibility factor of the prevention and control responsibility unit. This represents the slope length and slope factor of the prevention and control responsibility unit. This indicates the vegetation cover management factor for this prevention and control responsibility unit. This indicates the soil and water conservation measures factor for the responsible unit for prevention and control. This represents the area of ​​the engineering disturbance zone within the prevention and control responsibility unit, expressed in ha. The multiplication calculation is implemented using double-precision floating-point numbers and retains at least 6 decimal places to avoid rounding errors in small-area disturbance scenarios. Simultaneously, the area of ​​each prevention and control responsibility unit is... Establish a one-to-one correspondence between the identification of the prevention and control responsibility unit and record it in the baseline soil and water loss result table.

[0076] In this specific embodiment, S5 includes:

[0077] For each prevention and control responsibility unit, baseline soil erosion data were collected. and the proportion of the disturbance area The characteristics of disturbance morphology and spatial relationship between disturbance and drainage channels, and the reading of rainfall erosivity factor. Vegetation management factors Slope length and slope factor Soil erodibility factors Factors related to soil and water conservation measures ;

[0078] The disturbance morphology features are constructed into a numerical vector of length 5, which are, in order, the perimeter of the engineering disturbance area, the area of ​​the engineering disturbance area, the major axis direction angle, the aspect ratio and the degree of fragmentation. The spatial relationship features between the disturbance and the drainage channel are constructed into a numerical vector of length 3, which are, in order, the minimum distance from the boundary of the engineering disturbance area to the nearest drainage channel, the intersection length of the engineering disturbance area and the drainage channel, and the distribution length of the engineering disturbance area along the direction of the drainage channel.

[0079] Acquire rainfall forecast data corresponding to the prevention and control responsibility unit, fix the forecast period to the next 24 hours, fix the time step to 10 minutes, and obtain a length of [missing information]. The rainfall sequence was processed, and the sequence was organized into a rainfall time-series feature vector in chronological order. Missing time steps were padded with zeros to ensure consistent vector length. The rainfall data for each time step in the rainfall time-series feature vector was then pruned to a specified value. Millimeters to suppress outliers;

[0080] All numerical features used as model input are concatenated in a preset order to form input feature data, wherein the preset order is, in order, baseline soil erosion. Disturbance area ratio Disturbance morphology feature vector, spatial relationship feature vector between disturbance and drainage channel, rainfall erosivity factor Vegetation management factors Slope length and slope factor Soil erodibility factors Soil and water conservation measures factors With rainfall time-series feature vectors;

[0081] Numerical normalization is performed on the input feature data. The normalization parameters are the feature mean and feature standard deviation obtained from the training set during the training phase of historical soil and water loss monitoring samples and are stored in a fixed manner. During normalization, the normalization result of features with a standard deviation of zero is set to zero to avoid numerical overflow.

[0082] The normalized input feature data is fed into a pre-trained deep residual network model to obtain the target soil erosion risk index for the prevention and control responsibility unit. The deep residual network model includes a mechanism branch, a data branch, and a fusion layer, wherein the input of the mechanism branch only contains... and , The mechanism branch consists of an input layer, six fully connected residual blocks, and an output layer. Each fully connected residual block has a hidden dimension of 128 and uses linear transformation, batch normalization, and ReLU activation. The block input and block output are added to form a residual connection. The output of the mechanism branch is the soil and water loss estimation result based on mechanism constraints.

[0083] The input to the data branch contains Disturbance morphological feature vector, spatial relationship feature vector between disturbance and drainage channel, The data branch consists of an input layer, 8 fully connected residual blocks, and an output layer, with each fully connected residual block having a hidden dimension of 256. It adopts the same residual connection method as the mechanism branch. The output of the data branch is the soil erosion bias estimation result.

[0084] The fusion layer performs residual fusion of the mechanism branch output and the data branch output to obtain the target soil and water loss risk index, which satisfies the following:

[0085] ;

[0086] in This indicates the target soil and water loss risk index for the prevention and control responsibility unit, and outputs the amount of soil and water loss during the target forecast period as the unit of measurement. This represents the mechanism-constrained soil erosion estimation results output by the mechanism branch. This represents the soil erosion deviation estimation result output by the data branch;

[0087] The training samples of the deep residual network model consist of historical soil erosion monitoring samples, and each sample corresponds to the input feature data and measured soil erosion amount label of a prevention and control responsibility unit within a target forecast period. The measured soil erosion amount label is obtained by integrating the sediment content and flow rate of the downstream sediment monitoring section of the prevention and control responsibility unit and converted into tons.

[0088] Training uses the mean squared error loss function and the Adam optimizer, with a learning rate set to [value missing]. The batch size was set to 64, the number of training rounds was set to 200, and the loss was calculated on the validation set after each training round to save the model parameters with the minimum loss on the validation set as the parameters of the pre-trained deep residual network model.

[0089] During the inference phase, a forward computation is performed independently for each prevention and control responsibility unit, and its output is generated. and will Establish a one-to-one correspondence between the risk indicator results table and the identification of the prevention and control responsibility unit.

[0090] In this specific embodiment, S6 includes:

[0091] For each prevention and control responsibility unit within the scope of responsibility, read the target soil and water loss risk indicators. And read the pre-solidified and stored set of water and soil erosion risk classification thresholds. The set of thresholds for soil erosion risk classification The threshold risk indicators are determined by statistical analysis of historical soil erosion monitoring samples and calculated in one go during system deployment, then written into the threshold configuration file. The calculation process involves sorting all target soil erosion risk indicators obtained during the training phase using the same caliber from smallest to largest, with the sample size set to [value missing]. Take the sorted sequence number as follows: The sample values ​​are used as four thresholds and denoted as Thus, the threshold ranges are obtained in ascending order of risk.

[0092] During the operational phase, risk classification comparisons are performed for each prevention and control responsibility unit, and the soil erosion risk level is output. The risk classification rules satisfy the following:

[0093] ;

[0094] Where L represents the level number of the soil and water loss risk level and its value ranges from 1 to... This indicates the target soil and water loss risk index for the prevention and control responsibility unit, with units of... This represents the threshold separating low risk from relatively low risk, and the unit is 1. This represents the threshold separating low and medium risk, and the unit is... This represents the threshold separating medium and high risk, and the unit is 1. This represents the threshold separating higher and higher risks, and its unit is t.

[0095] Number the level The mapping is to a fixed name for the soil erosion risk level and maintains a one-to-one correspondence, whereby... Corresponding to low risk, Corresponding to lower risk, Corresponding to medium risk Corresponding to higher risks, Corresponding to high risk;

[0096] The soil and water loss risk levels of all prevention and control responsibility units are summarized at the scale of prevention and control responsibility scope, and each prevention and control responsibility unit is assigned a monitoring identifier and an early warning identifier. The monitoring identifier is a symbol code used for spatial display and is based on... The fixed values ​​are M1, M2, M3, M4, and M5, and the corresponding fill colors are green, blue, yellow, orange, and red, respectively, for rendering the prevention and control responsibility unit polygon elements in the GIS layer. The warning identifier is an event code used to trigger the alarm and is set according to... Fixed value ,in and The warning symbol W0 indicates no alarm. The warning indicator is W1, which indicates an alarm to be monitored. The warning symbol W2 indicates a warning. The warning identifier W3 indicates an emergency alarm, and the identifier of the prevention and control responsibility unit will be used when an emergency alarm is generated. , The start and end times of the target forecast period are written to the alarm log to ensure traceability;

[0097] The identification of each prevention and control responsibility unit and the target soil and water loss risk indicators shall be set. The names of soil and water loss risk levels, monitoring symbols, and early warning symbols are written into the risk classification result table, and a vector layer of prevention and control responsibility units with attribute fields is generated simultaneously as the output of soil and water loss risk monitoring results and early warning results within the scope of prevention and control responsibility.

[0098] In this specific embodiment, the process of generating terrain data and identifying drainage channels based on UAV multispectral aerial photography is as follows: A UAV equipped with a four-channel multispectral camera performs aerial surveying over the prevention and control responsibility unit to acquire UAV multispectral aerial photography images. The four-channel multispectral camera has fixed bands of blue light, green light, red light, and near-infrared, and each image records the RTK positioning information and attitude information at the exposure time. The flight altitude is set to achieve a ground resolution of 0.10 m, and the heading overlap is set to... And the lateral overlap is set to Six ground image control points were set up within the scope of the prevention and control responsibility unit, and the three-dimensional coordinates of the image control points were measured using the CGCS2000 coordinate system.

[0099] Aerial triangulation was performed on UAV multispectral aerial images. Bundle adjustment was performed using image-to-image matching and control point constraints, and the root mean square error of image point reprojection was controlled within 0.5 pixels, thereby obtaining the exterior orientation elements and sparse point cloud of each image.

[0100] Based on the aerial triangulation results, dense matching is performed to generate dense point clouds, and the dense point clouds are rasterized to generate a digital surface model (DSM). The raster resolution of the DSM is set to 0.10 m, and the raster value of the DSM is the highest elevation of the point cloud within the corresponding raster.

[0101] Ground point classification is performed on the DSM to generate a digital elevation model (DEM). The ground point classification uses a progressive TIN densification algorithm and a slope threshold. An elevation difference threshold of 0.5 m is used as the iterative update condition to remove non-ground points such as buildings and vegetation. Then, TIN interpolation is performed on the ground points and rasterized to obtain the DEM. The raster resolution of the DEM is 0.10 m and it uses the same map projection coordinate system as the prevention and control responsibility unit. The DEM is output as terrain data and used in subsequent steps for slope and runoff length calculation.

[0102] Based on the aerial triangulation results, orthorectification is performed on the UAV multispectral aerial photography to generate UAV orthophotos. The orthorectification uses the DEM to eliminate the displacement caused by terrain undulation and outputs the orthophotos in GeoTIFF format while maintaining the four-band order and the integrity of spatial reference information.

[0103] Orthophotos of drones Pixel segments are input into a drainage channel recognition model to output pixel-level recognition results of the drainage channels. The drainage channel recognition model is a semantic segmentation neural network with a DeepLabV3+ network structure. Its backbone network is ResNet-50 with an output stride of 16. The holes in the ASPP module are set to 6, 12, and 18, and a parallel global average pooling branch is used. The decoder employs a single-pass... Bilinear upsampling and two layers Convolution is used to refine the boundaries. The network input is a four-channel image block and the output is a single-channel drainage channel pixel probability map.

[0104] The drainage channel recognition model was trained using manually labeled drainage channel pixel masks. The training batch size was set to 6, the number of training epochs was set to 80, the optimizer used was SGD with a momentum of 0.9, and the initial learning rate was set to... The loss function uses cross-entropy loss and a fixed random seed is used during training to ensure that the parameter convergence results are reproducible.

[0105] During the inference phase, the probability map of drainage channel pixels is set according to a threshold. Binarization yields pixel-level recognition results for the drainage channel and satisfies... If and only if otherwise ,in Indicates that it is located at the th Line number Whether the cells in the column are binary labels for drainage channel cells, Indicates that it is located at the th Line number The probability value of a column's cells belonging to a drainage channel. This represents the probability threshold used for binarization;

[0106] The pixel-level recognition results are filtered for connected components, and connected components with an area of ​​less than 200 pixels are deleted to suppress isolated noise. Then, the retained area is subjected to refined skeleton extraction to obtain a single-pixel width centerline grid. The centerline grid is converted into polyline features in a grid-to-vector manner. The polyline simplification adopts the Douglas-Peucker algorithm and the simplification tolerance is set to 0.20 m to balance the smoothness of the line shape and the geometric accuracy. Thus, the spatial location data of the drainage channel is obtained and output in the form of a polyline feature set for spatial overlay analysis in step S2.

[0107] 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.

[0108] This invention forms a closed-loop algorithm combination of "pixel-level engineering disturbance identification - key factor calculation - baseline estimation of modified general soil loss equation - deep residual network risk prediction - threshold-based early warning": First, using the prevention and control responsibility unit as the monitoring and decision-making granularity, high-resolution remote sensing semantic segmentation is used to achieve fine extraction of engineering disturbances, and further quantifies the disturbance area ratio, disturbance morphology, and spatial relationship between disturbance and drainage channels, so that the impact of disturbance on runoff generation, runoff confluence, and erosion risk is expressed in a structured way; then, key factors such as R, K, LS, C, and P are calculated by combining information on rainfall, topography, soil, vegetation, and soil and water conservation measures, and baseline soil and water loss is obtained to achieve a mechanism characterization of risk; then, the baseline amount, disturbance characteristics, key factors, and rainfall forecast time series characteristics are input into the deep model to output target risk indicators, and risk classification is completed through a threshold set, thereby improving automation and timeliness, and achieving quantifiable early warning and higher accuracy risk identification for future periods.

[0109] In terms of algorithm structure, this invention makes targeted improvements to address the technical problems that "simple mechanistic models are difficult to reflect the complexity of engineering disturbances, and simple data models are unstable in generalization and lack constraints": a deep residual network containing mechanistic and data branches is constructed, where the mechanistic branch uses the RUSLE baseline and key factors as constraint inputs to ensure that the output is consistent with the erosion mechanism and improves cross-scenario robustness; the data branch introduces risk-sensitive features such as disturbance morphology, drainage channel relationships, and rainfall timing, and learns the "baseline-measured" deviation to compensate for the systematic errors of the mechanistic model under engineering disturbance scenarios; finally, residual fusion is performed in the fusion layer to obtain risk indicators, so that the model has both interpretable mechanistic consistency and adaptability to risk changes driven by complex disturbances and forecasted rainfall, thereby better achieving the effect of refined and forward-looking soil erosion risk monitoring and early warning technology.

Claims

1. A method for monitoring soil erosion risk based on artificial intelligence, characterized in that, include: S1. Acquire high-resolution remote sensing images covering the prevention and control responsibility area of ​​the production and construction project, perform preprocessing, and divide the project into multiple prevention and control responsibility units according to the prevention and control zoning of the soil and water conservation plan. Crop the preprocessed images to obtain remote sensing image sub-data corresponding to each prevention and control responsibility unit. S2. Input the remote sensing image sub-data of each prevention and control responsibility unit into a disturbance recognition model pre-trained based on disturbance samples to obtain pixel-level disturbance recognition results. Based on this, determine the disturbance area ratio, disturbance morphology characteristics, and spatial relationship characteristics between disturbance and drainage channels of each prevention and control responsibility unit, collectively referred to as disturbance features. S3. Based on the disturbance features of each prevention and control responsibility unit, and combined with the rainfall, topography, soil, vegetation, and soil and water conservation measures data corresponding to each prevention and control responsibility unit, calculate the rainfall erosivity factor, vegetation cover management factor, slope length and slope factor, soil erodibility factor, and soil and water conservation measure factor of each prevention and control responsibility unit, collectively referred to as key factors. S4. Substitute the key factors of each prevention and control responsibility unit into the modified general soil loss equation model to obtain the baseline soil and water loss of each prevention and control responsibility unit; S5. Based on the baseline soil and water loss, disturbance characteristics, key factors, and rainfall forecast data of each prevention and control responsibility unit, construct input feature data and input it into the deep residual network model trained in advance based on historical soil and water loss monitoring samples to obtain the target soil and water loss risk index of each prevention and control responsibility unit; S6. Based on the preset set of soil and water loss risk classification thresholds, classify the target soil and water loss risk index of each prevention and control responsibility unit to obtain the soil and water loss risk level, which is output as the soil and water loss risk monitoring result and early warning result of the prevention and control responsibility area.

2. The method for monitoring soil erosion risk based on artificial intelligence according to claim 1, characterized in that, S1 specifically refers to: A high-resolution remote sensing image with a spatial resolution better than 2 meters and containing multiple spectral bands is acquired, covering the entire prevention and control responsibility area of ​​the production and construction project in a planar position. The high-resolution remote sensing image is preprocessed, including at least radiometric correction and geometric correction, and preferably also radiometric calibration and atmospheric correction. The preprocessed high-resolution remote sensing image is then unified to a preset map projection coordinate system and a unified spatial resolution to obtain a corrected remote sensing image. Based on the boundary vector data of the prevention and control responsibility area of ​​the production and construction project and the prevention and control zoning vector data of the approved soil and water conservation plan for the production and construction project, the prevention and control responsibility area is divided into multiple prevention and control responsibility units. The corrected remote sensing image is spatially cropped based on the boundary of each prevention and control responsibility unit, so that each prevention and control responsibility unit corresponds to a continuous image region in the corrected remote sensing image, and each continuous image region is used as a remote sensing image sub-data corresponding to each prevention and control responsibility unit.

3. The method for monitoring soil erosion risk based on artificial intelligence according to claim 1, characterized in that, S2 specifically refers to: Within each prevention and control responsibility unit, the processed remote sensing image sub-data is input into a disturbance recognition model pre-trained based on disturbance samples. The disturbance recognition model is a semantic segmentation neural network with an encoder and decoder structure. It extracts spatial features based on a convolutional neural network and an attention mechanism and outputs a pixel-level classification result consistent with the spatial resolution of the remote sensing image sub-data. In the pixel-level classification result, each pixel is labeled as an engineering disturbance area pixel or a non-engineering disturbance area pixel, forming the pixel-level disturbance recognition result of the prevention and control responsibility unit. Within each prevention and control responsibility unit, the number of pixels marked as engineering disturbance areas is counted based on the pixel-level disturbance identification results of that prevention and control responsibility unit. The total area of ​​the engineering disturbance area is calculated by combining the area of ​​a single pixel in the remote sensing image sub-data. The ratio of the total area of ​​the engineering disturbance area to the total area of ​​the prevention and control responsibility unit is calculated to obtain the disturbance area ratio of the prevention and control responsibility unit. Within each prevention and control responsibility unit, the outer boundary of the engineering disturbance area is extracted based on the pixel-level disturbance identification results. At least one of the following is calculated based on the geometry of the outer boundary: perimeter, area, major axis direction, aspect ratio, compactness, and fragmentation degree of the engineering disturbance area, to obtain the disturbance morphology characteristics of that prevention and control responsibility unit. Multispectral aerial photographs of the UAV covering the prevention and control responsibility unit are acquired, and these photographs contain at least red and near-infrared bands. Terrain data is generated based on the multispectral aerial photographs, and the multispectral aerial photographs are input into a pre-trained drainage channel identification model to automatically identify drainage channels, thereby obtaining the spatial location data of the drainage channels. Within each prevention and control responsibility unit, the spatial location data of the drainage channel is spatially superimposed with the pixel-level disturbance identification results of that prevention and control responsibility unit. At least one of the following is calculated: the minimum distance from the boundary of the engineering disturbance area to the nearest drainage channel, the intersection length of the engineering disturbance area and the drainage channel, and the distribution length of the engineering disturbance area along the direction of the drainage channel. This yields the spatial relationship characteristics between the disturbance and the drainage channel of that prevention and control responsibility unit. This allows for the output of the disturbance area ratio, disturbance morphology characteristics, and spatial relationship characteristics between the disturbance and the drainage channel for each prevention and control responsibility unit.

4. The method for monitoring soil erosion risk based on artificial intelligence according to claim 1, characterized in that, S3 specifically refers to: Within each prevention and control responsibility unit, using the proportion of disturbed area, morphological characteristics of the disturbance, and spatial relationship characteristics between the disturbance and drainage channels as inputs, historical rainfall monitoring data, rainfall forecast data, topographic data, soil type data, information on the implementation of soil and water conservation measures, and remote sensing vegetation index data corresponding to that responsibility unit are acquired. The topographic data and remote sensing vegetation index data are generated from UAV multispectral aerial photography, and the information on the implementation of soil and water conservation measures is obtained from UAV multispectral aerial photography through deep learning recognition. Based on the historical rainfall monitoring data and the rainfall forecast data, combined with the proportion of disturbed area and morphological characteristics... Based on the spatial relationship characteristics between disturbances and drainage channels, the rainfall kinetic energy, maximum rainfall intensity, and rainfall concentration of the prevention and control responsibility unit within the target forecast period are calculated. The rainfall kinetic energy, maximum rainfall intensity, and rainfall concentration are then converted into a rainfall erosivity factor for the prevention and control responsibility unit according to a preset rainfall erosivity factor calculation formula. Based on the pixel-level disturbance identification results, the engineering disturbance area within the prevention and control responsibility unit is determined. The vegetation cover of the disturbance area is calculated within the engineering disturbance area using remote sensing vegetation index data. Finally, the vegetation cover of the disturbance area is converted into a vegetation cover factor for the prevention and control responsibility unit according to a preset vegetation cover management factor grading rule. The management factors are as follows: A digital elevation model of the prevention and control responsibility unit is constructed based on the topographic data. The slope and runoff length of each pixel within the prevention and control responsibility unit are calculated. The slope and runoff length are then calculated according to a preset slope length and slope factor calculation formula to obtain the slope length and slope factor of the prevention and control responsibility unit. Based on the soil type data, a soil attribute database is queried to obtain parameters such as soil texture, organic matter content, and structural gradation corresponding to different soil types. These parameters are then calculated according to a preset soil erodibility factor calculation formula to obtain the soil erodibility factor of the prevention and control responsibility unit. Based on the soil and water conservation... The information on the implementation of measures identifies the types, spatial locations, and coverage areas of soil and water conservation measures such as slope drainage ditches, intercepting ditches, retaining walls, temporary covering measures, and vegetation slope protection within the prevention and control responsibility unit. It also calculates the proportion of the engineering disturbance area covered by the soil and water conservation measures based on the proportion of the disturbance area covered by the measures, and converts the proportion of the engineering disturbance area covered by the measures and the type of soil and water conservation measures into soil and water conservation measure factors for the prevention and control responsibility unit according to the preset soil and water conservation measure factor calculation rules. This results in the output of rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erodibility factor, and soil and water conservation measure factor for each prevention and control responsibility unit.

5. The method for monitoring soil erosion risk based on artificial intelligence according to claim 1, characterized in that, S4 specifically refers to: Within each prevention and control responsibility unit, the baseline soil and water loss is calculated using the following inputs: rainfall erosivity factor, slope length and gradient factor, soil erodibility factor, vegetation cover management factor, soil and water conservation measures factor, and the area of ​​the engineering disturbance zone within the unit. The modified general soil loss equation model is used to calculate the baseline soil and water loss of the unit. The engineering disturbance zone area is the total area of ​​the engineering disturbance zone obtained from the pixel-level disturbance identification results. The modified general soil loss equation model is based on the general soil loss equation A = R × K × LS × C × P, with rainfall erosivity factor as R, soil erodibility factor as K, slope length and gradient factor as LS, vegetation cover management factor as C, and soil and water conservation measures factor as P. First, the baseline soil and water loss per unit area within the prevention and control responsibility unit during the target forecast period is calculated based on the general soil loss equation. Then, the baseline soil and water loss per unit area is multiplied by the area of ​​the engineering disturbance zone within the unit to obtain the baseline soil and water loss of the unit. The baseline soil and water loss of each prevention and control responsibility unit is then output.

6. The method for monitoring soil erosion risk based on artificial intelligence according to claim 1, characterized in that, S5 specifically refers to: Within each prevention and control responsibility unit, based on baseline soil erosion, proportion of disturbed area, morphological characteristics of disturbance, spatial relationship characteristics between disturbance and drainage channels, as well as rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erodibility factor, and soil and water conservation measures factor, rainfall forecast data corresponding to the prevention and control responsibility unit is obtained. The rainfall forecast data is then processed by time series analysis and numerical normalization to calculate the rainfall time series characteristics of the prevention and control responsibility unit within the forecast period. Within each prevention and control responsibility unit, baseline soil erosion, disturbed area proportion, disturbance morphology, spatial relationship between disturbance and drainage channels, rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erosibility factor, soil and water conservation measures factor, and rainfall temporal characteristics are concatenated according to a preset feature order to construct the input feature data for that prevention and control responsibility unit. The input feature data of each prevention and control responsibility unit are then input into a deep residual network model pre-trained based on historical soil erosion monitoring samples. This deep residual network model includes a mechanism branch and a data branch. The mechanism branch uses baseline soil erosion, rainfall erosivity factor, vegetation cover management factor, slope length and gradient factor, soil erosibility factor, and soil and water conservation measures factor as mechanism constraints as inputs. The model is then processed through a fully connected neural network with multiple residual connections. The soil erosion volume is nonlinearly transformed to obtain a soil erosion estimation result based on mechanism constraints. The data branch takes the baseline soil erosion volume, the proportion of disturbed area, the morphological characteristics of the disturbance, the spatial relationship characteristics between the disturbance and the drainage channel, the rainfall erosivity factor, the vegetation cover management factor, the slope length and slope factor, the soil erodibility factor, the soil and water conservation measures factor, and the rainfall time series characteristics as inputs. It learns the deviation between the baseline soil erosion volume and the actual soil erosion volume through a multi-layer residual connected neural network to obtain the soil erosion deviation estimation result. In the fusion layer of the deep residual network model, the soil erosion estimation result based on mechanism constraints output by the mechanism branch and the soil erosion deviation estimation result output by the data branch are residually fused to obtain the target soil erosion risk index of the prevention and control responsibility unit. The target soil erosion risk index of each prevention and control responsibility unit is output.

7. The method for monitoring soil erosion risk based on artificial intelligence according to claim 1, characterized in that, S6 specifically refers to: Within each prevention and control responsibility unit, the target soil and water loss risk index is used as input to obtain a set of soil and water loss risk classification thresholds that are pre-determined based on historical soil and water loss monitoring samples. The set of soil and water loss risk classification thresholds includes multiple threshold intervals arranged in order of risk from low to high. Within each prevention and control responsibility unit, the target soil and water loss risk index of the prevention and control responsibility unit is compared with the set of soil and water loss risk classification thresholds to determine the threshold range into which the target soil and water loss risk index of the prevention and control responsibility unit falls. The target soil and water loss risk index falling into different threshold ranges is respectively assigned to at least three soil and water loss risk levels: low risk, relatively low risk, medium risk, relatively high risk, and high risk, to obtain the soil and water loss risk level of each prevention and control responsibility unit. At the level of prevention and control responsibility scope, the soil and water loss risk levels of each prevention and control responsibility unit are summarized. Based on the soil and water loss risk level of each prevention and control responsibility unit, corresponding monitoring and early warning indicators are assigned to each prevention and control responsibility unit. The soil and water loss risk level of each prevention and control responsibility unit, as well as the corresponding monitoring and early warning indicators, are output as the soil and water loss risk monitoring results and early warning results of the prevention and control responsibility scope.

8. The method for monitoring soil erosion risk based on artificial intelligence according to claim 3, characterized in that, The terrain data generated based on UAV multispectral aerial photography includes: Aerial triangulation and orthorectification are performed on multispectral aerial photographs taken by UAVs to generate orthophotos, and digital surface model (DSM) and / or digital elevation model (DEM) are generated based on photogrammetric 3D reconstruction, wherein the terrain data includes the DEM; The drainage channel identification model is a semantic segmentation neural network and / or an object detection neural network, which outputs the pixel-level identification results of the drainage channel and vectorizes the pixel-level identification results to obtain the spatial location data of the drainage channel.

Images