A method and system for predicting soil erosion resistance based on surface parameters

Abstract

The application discloses a soil anti-erosion capacity prediction method and system based on surface parameters, relates to the technical field of soil erosion and water and soil conservation monitoring and prediction, and comprises the following steps: surface parameter collection, parameter preprocessing, erosion response modeling, anti-erosion threshold tensor inversion and anti-erosion capacity prediction. Multi-source surface parameter data is collected; secondly, a unified surface parameter feature set is constructed; an erosion response modeling method combining physical process simulation and lightweight machine learning residual correction is adopted to obtain erosion response simulation data; through the construction of an explicit Bayesian neural network and a variational inference framework, a four-dimensional anti-erosion threshold tensor is inverted by combining observation data; a double-branch spatio-temporal multi-task fusion prediction model of the anti-erosion threshold tensor is introduced to output the prediction value and the prediction grade of the soil anti-erosion capacity; and the application scheme can realize dynamic and spatialized prediction of the soil anti-erosion capacity and provide a scientific basis for water and soil conservation and ecological environment management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of soil erosion and soil and water conservation monitoring and prediction, and particularly relates to a soil erosion resistance prediction method and system based on surface parameters. BACKGROUND

[0002] The soil erosion resistance prediction method based on surface parameters refers to a method for realizing numerical prediction and graded evaluation of soil erosion resistance by collecting multi-source surface parameter data (topography, soil physicochemical properties, vegetation coverage, hydrology and meteorology, and human activities), performing data preprocessing and standardized feature construction, using erosion response modeling and erosion resistance threshold tensor inversion.

[0003] However, in the existing soil erosion resistance prediction method based on surface parameters, there are technical problems of data link fragmentation, lack of event dynamic adaptation of empirical formula, and large prediction result deviation caused by inconsistency between data sources; in the existing erosion response modeling method, there are technical problems of difficulty of single-factor product model in expressing rainstorm pulse and sediment transport continuity, and overfitting of pure machine learning model lacking physical constraints; in the existing erosion resistance threshold tensor inversion method, there are technical problems of threshold staticization, inability to represent soil depth and time dynamic change, and lack of pixel-level uncertainty description; in the existing erosion resistance prediction method, there are technical problems of inconsistency between regression prediction and grade classification, error amplification in threshold adjacent area, and low prediction accuracy of severe erosion grade. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides a soil erosion resistance prediction method and system based on surface parameters, which aims to overcome the technical problems of data link fragmentation, lack of event dynamic adaptation of empirical formula, and large prediction result deviation caused by inconsistency across data sources in the existing soil erosion resistance prediction method based on surface parameters. The present scheme creatively adopts a comprehensive erosion resistance prediction method combining erosion response modeling and erosion threshold tensor inversion, establishes a closed-loop mechanism of "physical simulation-residual correction-tensor inversion-prediction grading", realizes dynamic, spatial and interpretable prediction of soil erosion resistance, and makes the prediction results more accurate and transferable at the event level and basin level. In view of the technical problems of single-factor product model in the existing erosion response modeling method, such as difficulty in expressing storm pulse and sediment transport continuity, and pure machine learning model lacking physical constraints and being prone to overfitting, the present scheme creatively adopts an erosion response simulation method combining dynamic data assimilation and physically guided machine learning for erosion response modeling. By introducing a simplified sediment continuity equation and a dynamic vegetation coverage factor in the physical process simulation, combining a lightweight machine learning model for residual correction, and quality weighting and time smoothing fusion of the dual-source results, the present scheme realizes more precise modeling of the erosion process under extreme rainfall events and intermittent runoff scenarios, effectively improving the prediction ability of peak intensity and time evolution. In view of the technical problems of the existing erosion threshold tensor inversion method, such as threshold staticization, inability to represent soil depth and time dynamic changes, and lack of pixel-level uncertainty description, the present scheme creatively adopts a hybrid inversion method combining spatiotemporal optimization of explicit Bayesian neural network for erosion threshold tensor inversion. By constructing a four-dimensional erosion threshold tensor (spatial coordinates, soil depth, time step), and jointly minimizing the observed negative log-likelihood and spatiotemporal regularization term, the present scheme realizes dynamic inversion and uncertainty quantification of the soil erosion resistance threshold, so as to output the erosion threshold field with confidence interval, and improves the interpretability of the model and the reliability of the engineering application. In view of the technical problems of the existing erosion resistance prediction method, such as inconsistency between regression prediction and grade classification, error amplification in threshold adjacent areas, and low prediction accuracy of severe erosion grade, the present scheme creatively adopts a dual-branch spatiotemporal multi-task fusion prediction model improved by introducing an erosion threshold tensor, for erosion resistance prediction. By constructing a dynamic tensor feature branch and a static feature encoding branch, and simultaneously performing erosion resistance value regression and grade classification prediction in the multi-task output layer, and combining an improved loss function with threshold consistency constraint, the present scheme realizes consistency enhancement of regression results and grade labels, significantly reduces the prediction error rate at the threshold boundary, and improves the recognition ability of severe erosion areas.

[0005] The technical scheme adopted by the present application is as follows: The present application provides a soil erosion resistance prediction method based on surface parameters, which comprises the following steps:

[0006] Step S1: collecting surface parameters;

[0007] Step S2: parameter preprocessing;

[0008] Step S3: erosion response modeling;

[0009] Step S4: anti-erosion threshold tensor inversion;

[0010] Step S5: anti-erosion capacity prediction.

[0011] Further, in step S1, the surface parameter collection is used to obtain original surface parameter data related to soil anti-erosion capacity, specifically by multi-source data collection to obtain original surface parameter data;

[0012] The original surface parameter data specifically includes modeling original data and actual observation erosion data;

[0013] The modeling original data specifically includes topographic and geomorphic parameters, soil physicochemical parameters, vegetation coverage parameters, hydrological and meteorological parameters, and human activity parameters.

[0014] Further, in step S2, the parameter preprocessing is used to process the collected original surface parameter data into a standardized feature data set, specifically according to the original surface parameter data, by sequentially performing geometric correction, atmospheric correction, data fusion, derivative feature extraction, and feature standardization to obtain a standardized surface parameter feature set, including the following steps:

[0015] Step S21: geometric correction, specifically spatial position correction of remote sensing data and unmanned aerial survey data to eliminate geometric distortion and topographic relief effects to obtain geometric correction data;

[0016] Step S22: atmospheric correction, used to eliminate the effects of atmospheric scattering and absorption on spectral reflectance, specifically by reflectance correction to obtain real reflectance optimization data of ground objects;

[0017] Step S23: data fusion, used to fuse remote sensing data, unmanned aerial survey data, ground sensor monitoring data, and laboratory detection data to obtain a fusion data set;

[0018] Step S24: derivative feature extraction, used to generate composite features from the fusion data set, the composite features including slope, aspect, curvature, and convergence area based on digital elevation model, normalized vegetation index, soil adjusted vegetation index, and humidity index based on spectral data, and rainfall erosion force factor, soil erodibility factor, and structure stability index based on meteorological and soil parameters, to obtain derivative feature data;

[0019] Step S25: feature standardization, for normalizing or standardizing features of different sources and dimensions, so that the features are in a unified scale, specifically, performing a standardization operation on the derived feature data to obtain a standardized land surface parameter feature set;

[0020] The standardized land surface parameter feature set specifically includes topographic and geomorphic features, soil physicochemical features, vegetation cover features, water level and weather features, and human activity features.

[0021] Further, in step S3, the erosion response modeling is used to establish the physical relationship between the land surface parameters and the soil erosion process and to simulate and quantify the possible erosion dynamic response of the land surface under different environmental conditions, specifically, according to the standardized land surface parameter feature set, an erosion response simulation method combining dynamic data assimilation and physical guided machine learning is used to perform erosion response modeling to obtain erosion response simulation data, including the following steps:

[0022] Step S31: erosion response physical simulation, specifically, by integrating rainfall resistance factor, soil erodibility factor, slope length factor, slope factor, water and soil conservation factor, and dynamic vegetation cover factor, and introducing a simplified sediment continuity equation, the deposition and sediment movement process of soil erosion is performed, the erosion response physical simulation is performed, and prediction soil erosion amount data is obtained;

[0023] Step S32: constructing a physical correction prediction model, specifically, by constructing a lightweight machine learning model, the error of the prediction soil erosion amount data is learned, and the prediction data correction of the prediction soil erosion amount data is performed, and physical correction prediction erosion data is obtained;

[0024] Step S33: data smoothing fusion, specifically, by performing quality evaluation, weighted fusion and time dimension smoothing on the physical correction prediction erosion data and the prediction soil erosion amount data, fusion prediction data is obtained;

[0025] Step S34: erosion response modeling, specifically, according to the fusion prediction data, erosion response comprehensive modeling is performed, and erosion response simulation data is obtained;

[0026] The erosion response simulation data specifically includes erosion response simulation data and response posterior parameters.

[0027] Further, in step S4, the erosion resistance threshold tensor inversion is used to inversely deduce the comprehensive erosion resistance threshold of the soil according to the erosion response modeling result and the actually observed erosion result, and construct an erosion resistance threshold tensor, specifically, the erosion resistance threshold tensor inversion is performed by using a hybrid inversion method combining space-time optimization and an explicit Bayesian neural network, according to the erosion response simulation data and the actually observed erosion data in the original ground surface parameter data, to obtain the erosion resistance threshold tensor, including the following steps:

[0028] Step S41: erosion resistance threshold space-time tensor definition, specifically, the erosion resistance threshold tensor is defined by constructing a four-dimensional data structure, the four-dimensional data structure includes soil horizontal coordinates, soil vertical coordinates, soil layer depth and time steps, and is used to represent the comprehensive erosion resistance, to obtain an initial erosion resistance tensor;

[0029] Step S42: constructing a simulated observation likelihood function, specifically, the simulated observation likelihood function is constructed to calculate the probability of observing the actual erosion data under the condition of a given initial erosion resistance tensor, to obtain an observation likelihood probability;

[0030] Step S43: constructing a variational inference neural network, specifically, the observation likelihood probability is approximately inferred by constructing an explicit Bayesian neural network and a variational inference framework, to obtain a variational inference neural network;

[0031] Step S44: erosion resistance threshold tensor inversion, specifically, the variational inference neural network is used as an inverter to jointly minimize the observation negative log-likelihood and the space-time regularization term, to perform the inversion solution of the erosion resistance threshold tensor, to obtain the erosion resistance threshold tensor.

[0032] Further, in step S5, the erosion resistance capacity prediction is used to comprehensively predict and classify the soil erosion resistance capacity according to the erosion resistance threshold tensor, specifically, the erosion resistance capacity prediction is performed by using a dual-branch space-time multi-task fusion prediction model improved by introducing the erosion resistance threshold tensor, according to the erosion resistance threshold tensor, to obtain soil erosion resistance capacity comprehensive prediction data, including the following steps:

[0033] Step S51: constructing a dynamic tensor feature extraction branch, specifically, a three-dimensional convolution feature extraction layer and a time dynamic feature extraction layer are sequentially constructed, and a dynamic tensor feature extraction branch is constructed by space dimension compression;

[0034] Step S52: constructing a static feature encoding branch, specifically, two fully connected layers are constructed to process a static input vector, and a static feature encoding branch is constructed;

[0035] Step S53: constructing a feature fusion layer, specifically, constructing the feature fusion layer by feature splicing the dynamic feature vector output by the dynamic tensor feature extraction branch and the static feature vector output by the static feature encoding branch, and outputting a fusion feature;

[0036] Step S54: constructing a multi-task output layer, specifically, constructing the multi-task output layer by sequentially constructing a prediction anti-erosion capacity value regression task output layer and a prediction anti-erosion capacity level classification task output layer according to the fusion feature;

[0037] Step S55: constructing an anti-erosion threshold improvement loss, specifically, constructing the anti-erosion threshold improvement loss by introducing a weighted combination loss of improved regression and classification loss, to obtain an anti-erosion threshold improvement loss function;

[0038] Step S56: anti-erosion capacity prediction, specifically, training an anti-erosion capacity prediction model according to the dynamic tensor feature extraction branch, the static feature encoding branch, the feature fusion layer, the multi-task output layer and the anti-erosion threshold improvement loss function, obtaining the anti-erosion capacity prediction model, and predicting the anti-erosion capacity according to the anti-erosion threshold tensor by using the anti-erosion capacity prediction, to obtain soil anti-erosion capacity comprehensive prediction data;

[0039] The soil anti-erosion capacity comprehensive prediction data specifically includes an anti-erosion capacity prediction value and an anti-erosion capacity prediction level.

[0040] The application provides a soil anti-erosion capacity prediction system based on surface parameters, which comprises a surface parameter acquisition module, a parameter preprocessing module, an erosion response modeling module, an anti-erosion threshold tensor inversion module and an anti-erosion capacity prediction module.

[0041] The surface parameter acquisition module is used for surface parameter acquisition, and through surface parameter acquisition, original surface parameter data is obtained, and the original surface parameter data is sent to the parameter preprocessing module and the anti-erosion threshold tensor inversion module.

[0042] The parameter preprocessing module is used for parameter preprocessing, and through parameter preprocessing, a standardized surface parameter feature set is obtained, and the standardized surface parameter feature set is sent to the erosion response modeling module.

[0043] The erosion response modeling module is used for erosion response modeling, and through erosion response modeling, erosion response simulation data is obtained, and the erosion response simulation data is sent to the anti-erosion threshold tensor inversion module.

[0044] The anti-erosion threshold tensor inversion module is configured to perform anti-erosion threshold tensor inversion, obtain an anti-erosion threshold tensor through the anti-erosion threshold tensor inversion, and send the anti-erosion threshold tensor to the anti-erosion capacity prediction module;

[0045] The anti-erosion capacity prediction module is configured to perform anti-erosion capacity prediction, and obtain soil anti-erosion capacity comprehensive prediction data through the anti-erosion capacity prediction.

[0046] The above scheme has the following beneficial effects:

[0047] (1) In the existing soil anti-erosion capacity prediction method based on surface parameters, there are technical problems such as data link fragmentation, lack of event dynamic adaptation of empirical formula, and large prediction result deviation caused by inconsistency across data sources. The present scheme creatively adopts a comprehensive anti-erosion capacity prediction method combining erosion response modeling and anti-erosion threshold tensor inversion, establishes a closed-loop mechanism of "physical simulation-residual correction-tensor inversion-prediction grading", realizes dynamic, spatial, and interpretable prediction of soil anti-erosion capacity, and makes the prediction result have higher accuracy and migratability at the event level and the basin level;

[0048] (2) In the existing erosion response modeling method, there are technical problems such as difficulty of single-factor product model in expressing storm pulse and sediment transport continuity, and overfitting of pure machine learning model lacking physical constraints. The present scheme creatively adopts an erosion response simulation method combining dynamic data assimilation and physically guided machine learning, performs erosion response modeling, introduces a simplified sediment continuity equation and a dynamic vegetation coverage factor in physical process simulation, combines a lightweight machine learning model for residual correction, and performs quality weighting and time smoothing fusion on the dual-source results, thereby realizing more precise modeling of the erosion process under extreme rainfall events and intermittent runoff generation scenarios, and effectively improving the prediction ability of peak intensity and time evolution;

[0049] (3) In the existing anti-erosion threshold tensor inversion method, there are technical problems such as threshold staticization, inability to represent soil depth and time dynamic changes, and lack of pixel-level uncertainty description. The present scheme creatively adopts a hybrid inversion method combining spatiotemporal optimization of explicit Bayesian neural network, performs anti-erosion threshold tensor inversion, constructs a four-dimensional anti-erosion threshold tensor (spatial coordinates, soil depth, time step), and jointly minimizes the observation negative log-likelihood and the spatiotemporal regularization term, thereby realizing dynamic inversion and uncertainty quantification of the soil anti-erosion capacity threshold, and being capable of outputting an anti-erosion threshold field with a confidence interval, and improving the interpretability of the model and the reliability of the engineering application;

[0050] (4) In the existing anti-erosion capacity prediction method, there are technical problems of inconsistency between regression prediction and grade classification, error amplification in threshold adjacent area, and low prediction accuracy of severe erosion grade. The scheme creatively uses the improved double-branch space-time multi-task fusion prediction model with the introduction of anti-erosion threshold tensor to predict the anti-erosion capacity. By constructing a dynamic tensor feature branch and a static feature encoding branch, and simultaneously performing anti-erosion capacity value regression and grade classification prediction in the multi-task output layer, combined with the improved loss function with the introduction of threshold consistency constraint, the consistency of regression result and grade label is enhanced, the prediction error rate at the threshold boundary is significantly reduced, and the recognition ability of the severe erosion area is improved. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of a soil anti-erosion capacity prediction method based on ground surface parameters is provided.

[0052] Figure 2 A schematic diagram of a soil anti-erosion capacity prediction system based on ground surface parameters is provided.

[0053] Figure 3 A flowchart of step S2 parameter preprocessing is provided.

[0054] Figure 4 A flowchart of step S3 erosion response modeling is provided.

[0055] Figure 5 A flowchart of step S4 anti-erosion threshold tensor inversion is provided.

[0056] Figure 6 A flowchart of step S5 anti-erosion capacity prediction is provided.

[0057] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0059] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0060] Embodiment one, refer to Figure 1 The present application provides a soil erosion resistance prediction method based on surface parameters, which comprises the following steps:

[0061] Step S1: collecting surface parameters;

[0062] Step S2: parameter preprocessing;

[0063] Step S3: erosion response modeling;

[0064] Step S4: erosion resistance threshold tensor inversion;

[0065] Step S5: erosion resistance prediction.

[0066] By performing the above operation, for the technical problems of data link fragmentation, lack of event dynamic adaptation of empirical formula and large deviation of prediction results caused by inconsistency of cross-data source in the existing soil erosion resistance prediction method based on surface parameters, the present application creatively adopts a comprehensive erosion resistance prediction method combining erosion response modeling and erosion resistance threshold tensor inversion, establishes a closed-loop mechanism of "physical simulation-residual correction-tensor inversion-prediction grading", realizes dynamic, spatial and interpretable prediction of soil erosion resistance, and makes the prediction results have higher accuracy and migratability at the event level and the basin level.

[0067] Embodiment two, refer to Figure 1 and Figure 2 This embodiment is based on the above-mentioned embodiment, in step S1, the surface parameter collection is used to obtain original surface parameter data related to soil erosion resistance, specifically, through multi-source data collection, original surface parameter data is obtained;

[0068] The original surface parameter data specifically includes modeling original data and actual observed erosion data;

[0069] The modeling original data specifically includes topographic and geomorphic parameters, soil physicochemical parameters, vegetation coverage parameters, hydrological and meteorological parameters and human activity parameters;

[0070] Preferably, the topographic parameters, including slope (0°-90°), aspect (0°-360°), slope length (1m-500m), elevation (m), are used to characterize the surface slope morphology;

[0071] The soil physical and chemical parameters, including soil texture (sand, silt and clay ratio), organic matter content (0.0%-10.0%), bulk density (g / cm³), field capacity (%), pH value, salt content (g / kg), are used to reflect the soil resistance and vulnerability to erosion;

[0072] The vegetation coverage parameters, including normalized difference vegetation index (NDVI), leaf area index (LAI), vegetation coverage (0-100%), root depth (cm), are used to characterize the protective effect of vegetation on erosion;

[0073] The hydro-meteorological parameters, including annual rainfall (mm), rainfall intensity (mm / h), rainfall erosivity (MJ·mm / (ha·h·yr)), surface runoff (m³ / s), surface water content (%), are used to characterize the driving force of water erosion;

[0074] The human activity parameters, including land use type parameters (arable land, forest land, grassland and construction land), farmland tillage method parameters (no-tillage, rotary tillage and deep tillage), and road engineering disturbance range (area / length), are used to assess the impact of human activities on soil stability;

[0075] The actual observed erosion data, including measured gully erosion depth (cm), sheet erosion thickness (mm), sediment yield (t / ha) after rainfall events, are used for comparison and inversion verification with the modeling results;

[0076] The multi-source data collection specifically includes remote sensing data collection, unmanned aerial vehicle aerial survey parameter collection, ground monitoring parameter collection, laboratory soil detection collection and ground investigation and statistics collection.

[0077] Embodiment three, refer to Figure 1 、 Figure 2 and Figure 3 This embodiment is based on the above embodiment, in step S2, the parameter preprocessing is used to process the collected original surface parameter data into a standardized feature data set, specifically, according to the original surface parameter data, by sequentially performing geometric correction, atmospheric correction, data fusion, derivative feature extraction and feature standardization, a standardized surface parameter feature set is obtained, including the following steps:

[0078] Step S21: geometric correction, specifically, spatial position correction is performed on the remote sensing data and the unmanned aerial vehicle photogrammetry data, geometric distortion and terrain undulation effects are eliminated, and geometric correction data is obtained; preferably, the geometric correction is performed by establishing not less than 20 ground control points uniformly covering the study area, and a polynomial fitting function is used for spatial conversion, and the calculation formula is:

[0079] ;

[0080] In the formula, is the corrected pixel coordinate data, f(x, y) is a geometric correction operation function, (x, y) is the original pixel coordinate data before correction, n is the highest order of the fitted pixel in the horizontal direction, i is the horizontal order index, m is the highest order of the fitted pixel in the vertical direction, and j is the vertical order index. ij is a fitting coefficient, x i is the i-th power of the original coordinate in the horizontal direction, y i is the j-th power of the original coordinate in the vertical direction.

[0081] Preferably, the highest order n of the fitted pixel in the horizontal direction and the highest order m of the fitted pixel in the vertical direction are n = m = 1 when the image distortion is small and only translation, rotation and scale change exist, n = m = 2 when the image has strong perspective distortion or terrain undulation, and n = m = 3 in mountainous or severely distorted conditions, but the number of control points must be not less than the number of parameters (n + 1) (m + 1), and the control must be within 0.5 pixels according to the RMS evaluation;

[0082] Preferably, in this embodiment, n = m = 2;

[0083] Step S22: atmospheric correction, for eliminating the influence of atmospheric scattering and absorption on spectral reflectivity, specifically, real reflectivity optimization data of ground objects are obtained by reflectivity correction;

[0084] Preferably, the calculation formula of the reflectivity correction is:

[0085] ;

[0086] In the formula, is the reflectivity parameter in the real reflectivity optimization data of ground objects, L is the radiation brightness received by the sensor, d is the earth-sun distance correction factor, ESUN is the solar irradiance, is the solar elevation angle;

[0087] Step S23: data fusion, for fusing the remote sensing data, the unmanned aerial vehicle photogrammetry data, the ground sensor monitoring data and the laboratory detection data to obtain a fusion data set;

[0088] Step S24: Derivative feature extraction, for generating composite features from the fused dataset, including slope, aspect, curvature and contributing area based on digital elevation model, normalized difference vegetation index, soil adjusted vegetation index and moisture index based on spectral data, and rainfall erosivity factor, soil erodibility factor and structural stability index based on meteorological and soil parameters, to obtain derivative feature data;

[0089] Step S25: Feature standardization, for normalizing or standardizing features of different sources and dimensions to make them in a unified scale, specifically, standardizing the derivative feature data to obtain a standardized land surface parameter feature set;

[0090] The standardized land surface parameter feature set specifically includes topographic features, soil physicochemical features, vegetation cover features, water level and meteorological features, and human activity features.

[0091] Embodiment four, see Figure 1 、 Figure 2 and Figure 4 , based on the above embodiment, in step S3, the erosion response modeling is used to establish the physical relationship between the land surface parameters and the soil erosion process and simulate and quantify the possible erosion dynamic response of the land surface under different environmental conditions, specifically, according to the standardized land surface parameter feature set, an erosion response simulation method combining dynamic data assimilation and physical guided machine learning is used to model the erosion response, to obtain erosion response simulation data, including the following steps:

[0092] Step S31: Erosion response physical simulation, specifically, by integrating rainfall erosivity factor, soil erodibility factor, slope length factor, slope factor, water and soil conservation factor and dynamic vegetation cover factor, and introducing a simplified sediment continuity equation, the deposition and sediment movement process of soil erosion is carried out, the erosion response physical simulation is carried out, to obtain predicted soil erosion amount data, the calculation formula is:

[0093] ;

[0094] In the formula, A phys (t) is the predicted soil erosion amount data, R(t) is the rainfall erosivity factor, K is the soil erodibility factor, L is the slope length factor, S is the slope factor, P is the water and soil conservation factor, C dynamic (t) is the dynamic vegetation cover factor, The whole is a simplified sediment continuity equation, is a gradient operator, Q s (t) is the sediment transport rate parameter, is the surface runoff velocity vector;

[0095] Preferably, the rainfall erosivity factor is determined by calculating the total rainfall kinetic energy by summing up the rainfall intensity and kinetic energy of raindrops at each time interval during the rainfall process, and combining the maximum 30-minute rainfall intensity of the rainfall event to represent the scouring capacity of the rainfall on the soil surface.

[0096] The soil erodibility factor is determined according to the soil texture composition (specifically including the proportions of sand, silt and clay particles), organic matter content, soil structure grade and permeability grade, and is used to represent the sensitivity of the soil to erosion.

[0097] The slope length factor is determined according to the ratio of the actual slope length to the reference slope length, and is modified by a power term related to the slope; and the slope factor is determined according to the slope angle.

[0098] The soil and water conservation factor is determined according to the agricultural or engineering soil conservation measures taken on the ground surface, and when measures such as contour tillage, vegetation cover, terracing or trenching are taken on the slope surface, the value range of the factor is set to [0.5, 0.9], indicating the reduction effect of the soil conservation measures on the erosion amount; and when no measures are taken, the value range is set to [0.9, 1).

[0099] Preferably, the calculation formula of the dynamic vegetation cover factor is:

[0100]

[0101] In the formula, C dynamic (t) is the dynamic vegetation cover factor, is a vegetation calibration coefficient, and the default value range is set to [0.8, 1.2], NDVI(t) is a normalized vegetation index based on spectral data, is a vegetation protection coefficient, and the default value range is set to [1.0, 2.0], is a leaf area coefficient, and the default value range is set to [0.3, 0.7], and LAI(t) is a leaf area index.

[0102] Step S32: constructing a physical correction prediction model, specifically, learning the error of the predicted soil erosion amount data by constructing a lightweight machine learning model, and correcting the predicted soil erosion amount data to obtain physical correction prediction erosion data.

[0103] Preferably, the lightweight machine learning model specifically adopts a LightGBM model.

[0104] The physical correction prediction erosion data is specifically physically corrected and predicted by introducing a physical consistency regularization term, and the calculation formula is:

[0105] ;​

[0106] wherein L is a physically corrected prediction loss function for representing introducing a physically consistent regularization term for physically corrected prediction, is a data fitting loss term, wherein, is a corrected erosion amount prediction value, A obs is an actually observed erosion amount, is a rainfall intensity monotonicity weight, is a rain intensity monotonic loss term, is a corresponding gradient to the maximum rain intensity, is a maximum rainfall intensity in a 30-minute time scale, ||·||1 is an L1 norm operator, is an over-correction penalty weight, is an over-correction penalty term, A phys is prediction soil erosion amount data;

[0107] Step S33: data smoothing fusion, specifically, by quality evaluation, weighted fusion and time dimension smoothing on the physically corrected prediction erosion data and the prediction soil erosion amount data, fusion prediction data is obtained;

[0108] The quality evaluation is used for calculating the quality score of the data results of physical simulation and machine learning prediction;

[0109] The weighted fusion is used for data weighted fusion of the physically corrected prediction erosion data and the prediction soil erosion amount data according to the quality evaluation result;

[0110] The time dimension smoothing is used for exponential average weighting in the time dimension;

[0111] Step S34: erosion response modeling, specifically, erosion response comprehensive modeling is performed according to the fusion prediction data, and erosion response simulation data is obtained;

[0112] The erosion response simulation data specifically includes erosion response simulation data and response posterior parameters.

[0113] By performing the above operation, in the existing erosion response modeling method, there are technical problems that the single factor product model is difficult to express the rainstorm pulse and the continuity of sediment transport, and the pure machine learning model lacks physical constraints and is easy to overfit. The scheme creatively adopts an erosion response simulation method combining dynamic data assimilation and physically guided machine learning to model the erosion response. By introducing a simplified sediment continuity equation and a dynamic vegetation coverage factor in the physical process simulation, combining a lightweight machine learning model for residual correction, and performing quality weighting and time smoothing fusion on the dual-source results, the erosion process under extreme rainfall events and intermittent runoff scenarios is more finely modeled, effectively improving the prediction ability of peak intensity and time evolution.

[0114] Embodiment five, refer to Figure 1 、 Figure 2 and Figure 5 , this embodiment is based on the above embodiment, in step S4, the erosion resistance threshold tensor inversion is used to inversely deduce the comprehensive erosion resistance threshold of the soil according to the erosion response modeling result and the actually observed erosion result, and an erosion resistance threshold tensor is constructed, specifically: according to the erosion response simulation data and the actually observed erosion data in the original ground surface parameter data, an explicit Bayesian neural network hybrid inversion method combining space-time optimization is used to perform erosion resistance threshold tensor inversion, to obtain an erosion resistance threshold tensor, including the following steps:

[0115] Step S41: erosion resistance threshold space-time tensor definition, specifically: defining the erosion resistance threshold tensor by constructing a four-dimensional data structure, the four-dimensional data structure includes soil horizontal coordinates, soil vertical coordinates, soil layer depth and time step, and is used to represent the comprehensive erosion resistance ability, to obtain an initial erosion resistance tensor;

[0116] Step S42: constructing a simulated observation likelihood function, specifically: constructing a simulated observation likelihood function to calculate the probability of observing actual erosion data under the condition of a given initial erosion resistance tensor, to obtain an observation likelihood probability;

[0117] The calculation formula of the simulated observation likelihood function is:

[0118] ;

[0119] In the formula, is the observation likelihood probability, D is the actual observation data, used to represent the actual observed erosion data, is the initial erosion resistance tensor, is a multivariate Gaussian distribution probability distribution function, wherein, is a forward operator;

[0120] Step S43: constructing a variational inference neural network, specifically, constructing an explicit Bayesian neural network and a variational inference framework to approximately infer the observation likelihood probability, to obtain the variational inference neural network;

[0121] Preferably, the step of constructing an explicit Bayesian neural network and a variational inference framework to approximately infer the observation likelihood probability comprises Bayesian network setting, variational family setting, variational objective setting, spatiotemporal regularization term construction, and hybrid gate inversion.

[0122] The Bayesian network setting specifically represents the initial etch resistance tensor as a learnable weight in the Bayesian neural network according to the observation likelihood probability, and sets a prior distribution, preferably a Gaussian distribution with a mean of zero and an adjustable variance, to represent the parameter uncertainty without observation constraints.

[0123] The variational family setting specifically selects a Gaussian distribution family as a variational approximation form, uses mean parameters and variance parameters to characterize the posterior distribution of the etch resistance threshold tensor in the parameter space, and facilitates the use of reparameterization sampling methods to realize derivable inference in calculation.

[0124] The variational objective setting specifically sets the maximization of the evidence lower bound under the observation data as an optimization target.

[0125] The spatiotemporal regularization term construction specifically introduces a spatial smoothing term and a temporal smoothing term into the objective function to suppress non-physical oscillation of the etch resistance threshold tensor in the spatial and temporal dimensions, and to ensure the continuity and reasonableness of the prediction results. The spatial smoothing term is obtained by constraining the second-order difference of the etch resistance threshold tensor in the horizontal, vertical, and depth directions, and the temporal smoothing term is obtained by constraining the difference of adjacent time step etch resistance threshold tensors.

[0126] The calculation formula of the spatiotemporal regularization term is:

[0127] ;

[0128] In the formula, is the spatiotemporal regularization term, is a spatial regularization weight parameter, and the default value is set to 0.002, T is the total time step and takes a value of 12, t is the time index, L s is a spatial difference operator for calculating the second-order difference or neighborhood change rate of the etch resistance threshold tensor in the horizontal, vertical, and depth dimensions, is the threshold tensor after calculation of the initial etch resistance tensor, is a temporal regularization weight parameter, and the default value is set to 0.002, L tis the time difference operator used to compute the difference between adjacent time steps, suppresses unreasonable jumps in the time dimension, and ||·||2 is the L2 norm operator;

[0129] The mixed gate inversion is specifically dynamic weighting between data-driven variational inference and a regular term of physical constraints by introducing a gate mechanism in the network structure, adaptive balance in different scenarios is realized, and fitting accuracy and physical interpretability of the inversion result can be considered.

[0130] Preferably, Table 1 is a reference level and parameter example table of the variational inference neural network, as shown in the table, the variational inference neural network in the embodiment adopts a hierarchical structure design, a four-dimensional grid organization consistent with step S41 is adopted at the input end, and a tensor with a shape of (time step, depth layer, vertical, horizontal, channel) is constructed; the minimum implementation contains 8 channels, which are observation erosion, fused prediction, quality score, maximum 30-minute rain intensity (or rainfall erosion), slope, convergence area (landform convergence agent, logarithm is taken and then normalized), soil erodibility and erosion threshold initial field. The input data are extracted local features through 3*3*3 three-dimensional convolution, and then the time sequence dependence is modeled along the time dimension through a single-layer GRU (hidden unit 128); the network adopts a Gaussian prior with a mean of 0 and a variance of 0.5 and a full-factor Gaussian variational approximation, and the evidence lower bound is maximized through reparameterization sampling (sample number 3). In the training target, a joint loss of observation negative log-likelihood and a space-time regular term is adopted, and the space and time regular coefficients are both set to 0.002; at the same time, a mixed gate mechanism is introduced in the network structure, and dynamic weighting (average about 6:4 during training period) between the data fitting term and the physical regular term is performed through Sigmoid activation, so as to ensure the continuity and physical rationality of the prediction result. The observation noise adopts a homoscedastic Gaussian model (variance 0.01); the optimizer is Adam (learning rate 0.0002), the batch size is 2, the training is 80 rounds, and the verification early stop is 10 rounds; the above configuration balances between memory occupation, convergence performance and physical interpretability, and can realize robust inversion of the erosion threshold spatio-temporal tensor;

[0131] Table 1 Reference level and parameter example table of the variational inference neural network.

[0132]

[0133] Step S44: erosion threshold tensor inversion, specifically, the variational inference neural network is used as an inverter to jointly minimize the observation negative log-likelihood and the space-time regular term, and the erosion threshold tensor is inversely solved to obtain the erosion threshold tensor;

[0134] Preferably, the calculation formula of the minimum observation negative log-likelihood and the space-time regular term is:

[0135] ;

[0136] wherein, is a joint loss function for representing the loss composed of minimizing the observation negative log-likelihood and the spatio-temporal regularization term, is a fitting error operator, is a variational posterior distribution, which is specifically set as a full-factorized Gaussian distribution, is an observation probability term, which is specifically calculated by introducing the observation likelihood probability of the model parameter vector, is a variational distribution parameter vector of the variational inference neural network, and KL is a KL divergence operator, is a prior distribution, is a random parameter of the Bayesian neural network.

[0137] By performing the above operation, in order to solve the technical problems of threshold staticization, inability to represent the dynamic changes of soil depth and time, and lack of pixel-level uncertainty description in the existing anti-erosion threshold tensor inversion method, the scheme creatively adopts an explicit Bayesian neural network hybrid inversion method combined with spatio-temporal optimization to perform anti-erosion threshold tensor inversion. By constructing a four-dimensional anti-erosion threshold tensor (spatial coordinates, soil depth, time step), and jointly minimizing the observation negative log-likelihood and the spatio-temporal regularization term, the dynamic inversion and uncertainty quantification of the soil anti-erosion ability threshold are realized, so as to output the anti-erosion threshold field with a confidence interval, and improve the interpretability of the model and the reliability of the engineering application.

[0138] In the embodiment, the anti-erosion ability prediction is used for comprehensively predicting and grading the soil anti-erosion ability according to the anti-erosion threshold tensor. Figure 1 and Figure 6 The embodiment is based on the above-mentioned embodiment. In step S5, the anti-erosion ability prediction is used for comprehensively predicting and grading the soil anti-erosion ability according to the anti-erosion threshold tensor. Specifically, the anti-erosion ability prediction is performed by using a dual-branch spatio-temporal multi-task fusion prediction model improved by introducing the anti-erosion threshold tensor, to obtain soil anti-erosion ability comprehensive prediction data, including the following steps.

[0139] Step S51: constructing a dynamic tensor feature extraction branch, specifically by sequentially constructing a three-dimensional convolution feature extraction layer and a time dynamic feature extraction layer, and constructing a dynamic tensor feature extraction branch by spatial dimension compression;

[0140] Step S52: constructing a static feature encoding branch, specifically by constructing two fully connected layers to process the static input vector, and constructing a static feature encoding branch;

[0141] Step S53: constructing a feature fusion layer, specifically, constructing the feature fusion layer by performing feature splicing on the dynamic feature vector output by the dynamic tensor feature extraction branch and the static feature vector output by the static feature encoding branch, and outputting a fused feature;

[0142] Step S54: constructing a multi-task output layer, specifically, constructing the multi-task output layer by sequentially constructing a prediction of an erosion resistance value regression task output layer and a prediction of an erosion resistance level classification task output layer according to the fused feature;

[0143] Step S55: constructing an erosion resistance threshold improvement loss, specifically, constructing the erosion resistance threshold improvement loss by introducing a weighted combination loss of improved regression and classification losses, to obtain an erosion resistance threshold improvement loss function;

[0144] Step S56: predicting an erosion resistance, specifically, training an erosion resistance prediction model according to the dynamic tensor feature extraction branch, the static feature encoding branch, the feature fusion layer, the multi-task output layer, and the erosion resistance threshold improvement loss function, obtaining the erosion resistance prediction model, and predicting the erosion resistance according to the erosion resistance threshold tensor by using the erosion resistance prediction, to obtain comprehensive prediction data of soil erosion resistance;

[0145] The comprehensive prediction data of soil erosion resistance specifically includes an erosion resistance prediction value and an erosion resistance prediction level.

[0146] Preferably, Table 2 is an example table of parameters of the erosion resistance prediction model. As shown in the table, the erosion resistance prediction model constructed in this embodiment adopts a double-branch spatio-temporal multi-task structure. The dynamic tensor branch is provided with two layers of three-dimensional convolution (convolution kernel 3x3x3, channel number 32→64), and a single layer of GRU128 unit is introduced for time series modeling. Then, spatial dimension compression is realized through global average pooling. The static feature branch adopts a two-layer fully connected structure, and the node numbers are 128 and 64 in sequence. The dynamic and static features are spliced in the fusion layer, and are regularized through a fully connected layer (128 nodes) with Dropout 0.2. The output part includes a regression head (fully connected 128→1) and a classification head (fully connected 128→5 classes), and is trained through a multi-task weighted loss. The regression and classification weight ratio is 0.7:0.3. L2 weight decay 5e−4 and gradient clipping threshold 0.8 are adopted in the training process to ensure the stability of the model. The Adam optimizer is selected as the optimizer, the learning rate is 0.0003, the batch size is 16, the model is trained for 100 rounds, and the verification early stops for 15 rounds, so that a relatively stable erosion resistance prediction result is obtained.

[0147] Table 2: Example table of parameters of the erosion resistance prediction model.

[0148]

[0149] By performing the above operation, for the technical problems of inconsistency between regression prediction and grade classification, error amplification in threshold adjacent area, low prediction accuracy of severe erosion grade in existing anti-erosion ability prediction methods, the scheme creatively adopts an improved double-branch space-time multi-task fusion prediction model introducing an anti-erosion threshold tensor, performs anti-erosion ability prediction, constructs a dynamic tensor feature branch and a static feature encoding branch, simultaneously performs anti-erosion ability value regression and grade classification prediction in a multi-task output layer, and combines an improved loss function introducing threshold consistency constraint to realize consistency enhancement of regression results and grade labels, significantly reduce prediction misjudgment rate at the threshold boundary, and improve the recognition ability of severe erosion area.

[0150] Embodiment seven, refer to Figure 1 and Figure 2 Based on the above-mentioned embodiments, the application provides a soil anti-erosion ability prediction system based on surface parameters, which comprises a surface parameter acquisition module, a parameter preprocessing module, an erosion response modeling module, an anti-erosion threshold tensor inversion module and an anti-erosion ability prediction module.

[0151] The surface parameter acquisition module is used for surface parameter acquisition, and through surface parameter acquisition, original surface parameter data is obtained, and the original surface parameter data is sent to the parameter preprocessing module and the anti-erosion threshold tensor inversion module.

[0152] The parameter preprocessing module is used for parameter preprocessing, and through parameter preprocessing, a standardized surface parameter feature set is obtained, and the standardized surface parameter feature set is sent to the erosion response modeling module.

[0153] The erosion response modeling module is used for erosion response modeling, and through erosion response modeling, erosion response simulation data is obtained, and the erosion response simulation data is sent to the anti-erosion threshold tensor inversion module.

[0154] The anti-erosion threshold tensor inversion module is used for anti-erosion threshold tensor inversion, and through anti-erosion threshold tensor inversion, an anti-erosion threshold tensor is obtained, and the anti-erosion threshold tensor is sent to the anti-erosion ability prediction module.

[0155] The anti-erosion ability prediction module is used for anti-erosion ability prediction, and through anti-erosion ability prediction, soil anti-erosion ability comprehensive prediction data is obtained.

[0156] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, since the scope of the present application will be limited to the appended claims. It will be understood that where the application is defined by a plurality of separate claims, each of the claims can be combined with any of the other claims.

[0157] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made therein without departing from the spirit and scope of the application.

[0158] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution, which should belong to the protection scope of the application.

Claims

1. A method for predicting soil erosion resistance based on surface parameters, characterized in that: The method includes the following steps: Step S1: Surface parameter acquisition. Raw surface parameter data are obtained through multi-source data acquisition. Step S2: Parameter preprocessing. Based on the original surface parameter data, geometric correction, atmospheric correction, data fusion, derived feature extraction, and feature standardization are performed sequentially to obtain a standardized surface parameter feature set. Step S3: Erosion response modeling. Based on the standardized surface parameter feature set, an erosion response simulation method combining dynamic data assimilation and physical-guided machine learning is used to model the erosion response and obtain erosion response simulation data. This includes the following steps: Step S31: Physical simulation of erosion response; Step S32: Construction of a physical correction prediction model; Step S33: Data smoothing and fusion; Step S34: Erosion response modeling. Step S4: Erosion resistance threshold tensor inversion. Based on the erosion response simulation data and the actual observed erosion data in the original surface parameter data, the erosion resistance threshold tensor is inverted using a hybrid inversion method combining spatiotemporal optimization and explicit Bayesian neural networks to obtain the erosion resistance threshold tensor. Step S5: Erosion resistance prediction. Based on the erosion resistance threshold tensor, an improved dual-branch spatiotemporal multi-task fusion prediction model with the introduction of the erosion resistance threshold tensor is used to predict the erosion resistance and obtain comprehensive prediction data of soil erosion resistance.

2. The method for predicting soil erosion resistance based on surface parameters according to claim 1, characterized in that: In step S1, the original surface parameter data specifically includes modeling raw data and actual observed erosion data; the modeling raw data specifically includes topographic parameters, soil physicochemical parameters, vegetation cover parameters, hydrological and meteorological parameters, and human activity parameters.

3. The method for predicting soil erosion resistance based on surface parameters according to claim 2, characterized in that: In step S2, the standardized surface parameter feature set specifically includes topographic and geomorphological features, soil physicochemical features, vegetation cover features, hydrological and meteorological features, and human activity features.

4. The method for predicting soil erosion resistance based on surface parameters according to claim 3, characterized in that: In step S3, the erosion response modeling is used to establish the physical relationship between surface parameters and soil erosion processes, and to simulate and quantify the possible dynamic erosion responses of the surface under different environmental conditions. Specifically, based on the standardized surface parameter feature set, an erosion response simulation method combining dynamic data assimilation and physical-guided machine learning is used to perform erosion response modeling and obtain erosion response simulation data, including the following steps: Step S31: Physical simulation of erosion response. Specifically, by integrating rainfall erosion resistance factor, soil erodibility factor, slope length factor, slope factor, soil and water conservation factor and dynamic vegetation cover factor, and introducing a simplified sediment continuity equation, the deposition and sediment movement processes of soil erosion are simulated to obtain predicted soil erosion data. Step S32: Construct a physical correction prediction model, specifically by constructing a lightweight machine learning model to learn the error of the predicted soil erosion data and correct the predicted soil erosion data to obtain physically corrected predicted erosion data. Step S33: Data smoothing and fusion, specifically, by performing quality assessment, weighted fusion and time dimension smoothing on the physical correction predicted erosion data and the predicted soil erosion data, to obtain fused predicted data; Step S34: Erosion response modeling, specifically, based on the fused prediction data, erosion response comprehensive modeling is performed to obtain erosion response simulation data.

5. The method for predicting soil erosion resistance based on surface parameters according to claim 4, characterized in that: In step S3, the erosion response simulation data specifically includes erosion response simulation data and response posterior parameters.

6. The method for predicting soil erosion resistance based on surface parameters according to claim 5, characterized in that: In step S4, the erosion resistance threshold tensor inversion is used to infer the comprehensive erosion resistance threshold of the soil based on the erosion response modeling results and the actual observed erosion results, and to construct the erosion resistance threshold tensor. Specifically, based on the erosion response simulation data and the actual observed erosion data in the original surface parameter data, a hybrid inversion method combining spatiotemporal optimization and explicit Bayesian neural networks is used to perform erosion resistance threshold tensor inversion to obtain the erosion resistance threshold tensor, including the following steps: Step S41: Define the spatiotemporal tensor of the erosion resistance threshold. Specifically, the erosion resistance threshold tensor is defined by constructing a four-dimensional data structure. The four-dimensional data structure includes soil horizontal coordinates, soil vertical coordinates, soil depth, and time step, and is used to characterize the comprehensive erosion resistance capability to obtain the initial erosion resistance tensor. Step S42: Construct the simulated observation likelihood function, specifically by constructing the simulated observation likelihood function and calculating the probability of observing actual erosion data under the given initial erosion resistance tensor, thus obtaining the observation likelihood probability; Step S43: Construct a variational inference neural network, specifically by constructing an explicit Bayesian neural network and a variational inference framework to make an approximate inference of the observation likelihood probability, thereby obtaining the variational inference neural network; Step S44: Inversion of the corrosion resistance threshold tensor. Specifically, the corrosion resistance threshold tensor is obtained by using the variational inference neural network as an inversion device, and by jointly minimizing the observation negative log-likelihood and the spatiotemporal regularization term.

7. The method for predicting soil erosion resistance based on surface parameters according to claim 6, characterized in that: In step S5, the erosion resistance prediction is used to comprehensively predict and classify the soil erosion resistance based on the erosion resistance threshold tensor. Specifically, based on the erosion resistance threshold tensor, a dual-branch spatiotemporal multi-task fusion prediction model improved by introducing the erosion resistance threshold tensor is used to predict the erosion resistance, obtaining comprehensive soil erosion resistance prediction data, including the following steps: Step S51: Construct a dynamic tensor feature extraction branch, specifically by sequentially constructing a three-dimensional convolutional feature extraction layer and a temporal dynamic feature extraction layer, and by compressing the spatial dimension to obtain the dynamic tensor feature extraction branch; Step S52: Construct a static feature encoding branch, specifically by constructing two fully connected layers to process the static input vector and obtain the static feature encoding branch; Step S53: Construct a feature fusion layer, specifically by concatenating the dynamic feature vector output from the dynamic tensor feature extraction branch and the static feature vector output from the static feature encoding branch to construct a feature fusion layer and output fused features; Step S54: Construct a multi-task output layer. Specifically, based on the fusion features, a multi-task output layer is constructed by sequentially constructing a regression task output layer for predicting anti-erosion capability values ​​and a classification task output layer for predicting anti-erosion capability levels. Step S55: Construct the improved loss for corrosion resistance threshold. Specifically, the improved loss for corrosion resistance threshold is constructed by introducing a weighted combination loss of improved regression and classification loss, resulting in the improved loss function for corrosion resistance threshold. Step S56: Erosion resistance prediction, specifically, based on the dynamic tensor feature extraction branch, the static feature encoding branch, the feature fusion layer, the multi-task output layer and the erosion resistance threshold improved loss function, the erosion resistance prediction model is trained to obtain the erosion resistance prediction model, and by using the erosion resistance prediction, based on the erosion resistance threshold tensor, the erosion resistance prediction is performed to obtain comprehensive prediction data of soil erosion resistance. The comprehensive prediction data of soil erosion resistance specifically includes the predicted value of erosion resistance and the predicted level of erosion resistance.

8. A soil erosion resistance prediction system based on surface parameters, used to implement the soil erosion resistance prediction method based on surface parameters as described in any one of claims 1-7, characterized in that: It includes a surface parameter acquisition module, a parameter preprocessing module, an erosion response modeling module, an erosion resistance threshold tensor inversion module, and an erosion resistance prediction module.

9. A soil erosion resistance prediction system based on surface parameters according to claim 8, characterized in that: The surface parameter acquisition module is used for surface parameter acquisition. Through surface parameter acquisition, raw surface parameter data is obtained, and the raw surface parameter data is sent to the parameter preprocessing module and the erosion threshold tensor inversion module. The parameter preprocessing module is used for parameter preprocessing. Through parameter preprocessing, a standardized surface parameter feature set is obtained, and the standardized surface parameter feature set is sent to the erosion response modeling module. The erosion response modeling module is used for erosion response modeling. Through erosion response modeling, erosion response simulation data is obtained, and the erosion response simulation data is sent to the corrosion resistance threshold tensor inversion module. The corrosion resistance threshold tensor inversion module is used for corrosion resistance threshold tensor inversion. Through corrosion resistance threshold tensor inversion, a corrosion resistance threshold tensor is obtained, and the corrosion resistance threshold tensor is sent to the corrosion resistance prediction module. The erosion resistance prediction module is used to predict erosion resistance and obtain comprehensive prediction data of soil erosion resistance through erosion resistance prediction.

Images