Water and soil loss intelligent deduction method and system based on AIGC technology

Abstract

The invention discloses an AIGC technology-based water and soil loss intelligent deduction method and system. The method comprises the steps of 1, collecting water and soil loss data of a target area; 2, preprocessing the water and soil loss data; 3, water and soil loss deduction data are generated through a water and soil loss generation model; 4, updating the water and soil loss scene generation module through a dynamic feedback mechanism; 5, deducing a soil erosion modulus in a future preset time period through the improved space-time diagram convolutional network; 6, executing multi-scene water and soil loss intelligent deduction, and generating water and soil loss deduction data under different scenes; and step 7, carrying out water and soil loss risk grade division, and carrying out uncertainty quantitative analysis by adopting Monte Carlo-based Dropout and Bootstrap sampling. According to the method, the credibility, the multi-scene adaptability and the spatial expression precision of the water and soil loss deduction result are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental monitoring and prediction, and particularly relates to a water and soil loss intelligent deduction method and system based on AIGC technology. BACKGROUND

[0002] With the wide deployment of remote sensing technology, unmanned aerial vehicle surveying and mapping platform and ground monitoring network in the field of soil and water conservation, water and soil loss monitoring and prediction based on multi-source observation data are attracting more and more attention. Existing researches usually rely on the combination of satellite remote sensing inversion, ground sensor observation and numerical simulation model to evaluate and deduce the water and soil loss process. However, in the actual application of complex terrain, extreme climate and data sparse area, the following problems still exist:

[0003] Firstly, although the traditional physical mechanism type erosion model has clear interpretability, it is highly dependent on continuous boundary conditions, refined parameter input and a large amount of measured data support. In the area where the observation condition is limited, it is difficult to ensure the robustness of the model, and the model has large calculation amount and slow response speed, which is difficult to meet the high-frequency dynamic prediction demand.

[0004] Secondly, the pure data-driven deep learning model can fit certain statistical rules under the condition of rich samples, but the generalization ability is weak, and it is easy to produce significant deviation for unobserved landform structure, rainfall mode or human disturbance scene. At the same time, such model usually lacks the expression ability of physical constraints of erosion process, which leads to the inconsistency between the deduced results and the real hydrodynamic process.

[0005] Thirdly, the existing researches usually separate the "data generation" and "state deduction", and lack unified modeling framework between data enhancement, missing data completion and erosion state prediction. The synthetic data generated by the generative model often cannot establish a closed-loop correction mechanism with the real observation, which leads to the deviation between the generated scene and the actual physical process, thereby affecting the credibility of the deduced results.

[0006] Fourthly, the multi-source observation data have significant heterogeneity in terms of spatial and temporal resolution, sampling density and semantic structure. Remote sensing image has large coverage but limited time resolution, unmanned aerial vehicle has high spatial resolution but long flight cycle, and ground sensor has high sampling frequency but limited spatial coverage. Different data sources are difficult to realize unified expression, which leads to the difficulty in balancing the macro trend description and the fine-scale erosion detail expression of the deduced results.

[0007] The above deficiencies make it difficult for the prior art to realize rapid, robust and high-credibility prediction of water and soil loss processes under the scenarios of sudden heavy rainfall, complex topography interference or engineering disturbance, and it is difficult to meet the urgent needs of forward-looking early warning and intelligent deduction in water and soil conservation management. Therefore, how to provide an intelligent deduction method and system for water and soil loss based on AIGC technology is a problem that needs to be solved by those skilled in the art. SUMMARY

[0008] One object of the present application is to provide an intelligent deduction method and system for water and soil loss based on AIGC technology. The present application comprehensively utilizes generative artificial intelligence model, spatio-temporal graph convolution network and multi-source observation data fusion technology to build an intelligent water and soil loss simulation system that can self-learn, self-correct and has cross-scenario deduction capability. By introducing multi-scale feature fusion, conditional generation model, dynamic feedback update mechanism and graph convolution deduction structure embedded with physical equation constraints, the present application can continuously output high-credibility water and soil loss deduction results under the conditions of sparse observation data, complex terrain or abnormal climate, effectively improving the deduction accuracy and stability, enhancing the model adaptability and physical consistency, and realizing high-credibility water and soil loss prediction under multiple scenarios.

[0009] According to an embodiment of the present application, an intelligent deduction method for water and soil loss based on AIGC technology comprises the following steps:

[0010] Step one: collect water and soil loss data of the target area;

[0011] Step two: perform spatio-temporal alignment and standardization processing on the water and soil loss data to obtain a multi-scale spatio-temporal feature vector sequence, and use a feature fusion network based on attention mechanism to weight and fuse the multi-scale spatio-temporal feature vector sequence to generate a spatio-temporal fusion feature tensor;

[0012] Step three: build a water and soil loss generation model based on AIGC technology, the water and soil loss generation model comprising a water and soil loss scene generation module and a discrimination module; input the spatio-temporal fusion feature tensor into the water and soil loss scene generation module to generate water and soil loss deduction data, and perform adversarial training on the real observation data and the water and soil loss deduction data through the discrimination module;

[0013] Step four: establish a dynamic feedback mechanism to update the water and soil loss scene generation module online to obtain an optimized water and soil loss generation model;

[0014] Step five: input the water and soil loss deduction data and the physical correction observation features into an improved spatio-temporal graph convolution network to deduce the soil erosion modulus in a future preset time period, and output a water and soil loss intensity layer and a spatial distribution layer, wherein the improved spatio-temporal graph convolution network introduces a physical constraint term;

[0015] Step six: introduce scenario driving factors and fuse with the spatio-temporal fusion feature tensor to input into the optimized soil erosion generation model, perform multi-scenario soil erosion intelligent deduction, generate soil erosion deduction data under different scenarios, and output soil erosion intensity layer sequence and soil erosion evolution trend graph;

[0016] Step seven: map the soil erosion intensity layer sequence to the geographic spatial layer sequence, and divide the soil erosion risk level; use Monte Carlo-based Dropout and Bootstrap sampling to perform uncertainty quantification analysis.

[0017] Optionally, the step one specifically includes:

[0018] The soil erosion data of the target region is collected by a satellite remote sensing platform, an unmanned aerial vehicle aerial survey system, and a ground sensor network, and the soil erosion data includes high-resolution multispectral images, laser radar point clouds, soil moisture, rainfall intensity, surface roughness, and vegetation coverage.

[0019] Optionally, the step two specifically includes:

[0020] Atmospheric correction and geometric correction are performed on the high-resolution multispectral images to extract normalized vegetation index, slope, aspect, and soil brightness index;

[0021] The laser radar point cloud is reconstructed in three dimensions to generate a digital surface model, and the surface roughness feature and fine-scale vegetation coverage feature are calculated based on the moving window method;

[0022] The soil moisture, rainfall intensity, surface roughness, and vegetation coverage are subjected to outlier rejection, unit unification, and time stamp synchronization processing to obtain a standardized hydrological physical quantity sequence;

[0023] The normalized vegetation index, slope, aspect, and soil brightness index, surface roughness feature, fine-scale vegetation coverage feature, and standardized hydrological physical quantity sequence are resampled and interpolated according to spatial position and time label with spatial grid unit and time step as index to obtain a multi-scale feature vector sequence;

[0024] The multi-scale feature vector sequence is input into the feature fusion network based on the attention mechanism to assign attention weights to the feature vectors from different data sources and perform weighted fusion to generate a spatio-temporal fusion feature tensor.

[0025] Optionally, the step three specifically includes:

[0026] The soil erosion scenario generation module uses a conditional variational autoencoder structure to learn the latent distribution of the soil erosion process and generate soil erosion deduction data;

[0027] The encoder of the water and soil erosion scene generation module includes four down-sampling stages, each of which extracts local spatial features in a spatio-temporal fusion feature tensor through a 3*3 convolution layer to obtain a down-sampled feature map, and reduces the spatial dimension of the down-sampled feature map through a 2*2 max-pooling layer; the number of channels of the encoder is multiplied by a factor of two at each down-sampling stage, from an initial 64 to 128, 256 and 512, to enhance the semantic expression ability of deep features;

[0028] The decoder of the water and soil erosion scene generation module is provided with four up-sampling stages, each of which restores the spatial resolution step by step through a transpose convolution layer, and adopts a skip connection to concatenate the down-sampled feature map of each down-sampling stage of the encoder with the feature map of the corresponding up-sampling stage in the channel to generate an up-sampling fusion feature map and serve as an input feature map for the next up-sampling stage;

[0029] The discriminator module adopts a multi-scale convolutional neural network structure, including a first scale discrimination branch, a second scale discrimination branch and a third scale discrimination branch; each scale discrimination branch is sequentially stacked by a plurality of 3*3 convolution layers, batch normalization layers and LeakyReLU activation layers, and is provided with different down-sampling times within each scale discrimination branch;

[0030] Original resolution images, 2 times down-sampling images and 4 times down-sampling images of real observation data and water and soil erosion deduction data are respectively constructed;

[0031] The first scale discrimination branch receives the original resolution images of the real observation data and the original resolution images of the water and soil erosion deduction data respectively, outputs first scale real discrimination feature maps and first scale generated discrimination feature maps, and performs per-pixel discrimination through a 1*1 convolution layer to obtain first scale real response maps and first scale generated response maps;

[0032] The second scale discrimination branch receives the 2 times down-sampling images of the real observation data and the 2 times down-sampling images of the water and soil erosion deduction data respectively, outputs second scale real discrimination feature maps and second scale generated discrimination feature maps, and performs per-pixel discrimination through a 1*1 convolution layer to obtain second scale real response maps and second scale generated response maps;

[0033] The third scale discrimination branch receives the 4 times down-sampling images of the real observation data and the 4 times down-sampling images of the water and soil erosion deduction data respectively, outputs third scale real discrimination feature maps and third scale generated discrimination feature maps, and performs per-pixel discrimination through a 1*1 convolution layer to obtain third scale real response maps and third scale generated response maps;

[0034] The loss function of the adversarial training includes a least squares loss and a feature matching loss;

[0035] The least square loss is: the expectation of the square difference of the three scale real response maps and 1, the expectation of the square of the three scale generated response maps, and the addition after averaging respectively;

[0036] The feature matching loss is: the expectation of the element-wise absolute difference of the three scale real discriminative feature maps and the generated discriminative feature maps, and the addition.

[0037] Optionally, the dynamic feedback mechanism specifically includes: setting a sliding time window, continuously receiving real-time observation data, and calculating the root mean square error between the soil and water loss deduction data and the real-time observation data as a performance loss function; based on the performance loss function, the adaptive moment estimation optimization algorithm is used to update the parameters of the soil and water loss scene generation module, and the initial value of the learning rate is 0.001, and the learning rate is exponentially attenuated according to the training round or the error convergence condition.

[0038] Optionally, the step five specifically includes:

[0039] According to the soil and water loss data, a general soil loss equation is used to construct a physical corrected observation feature, and the physical corrected observation feature includes a rainfall erosivity factor, a soil erodibility factor and an elevation;

[0040] The soil and water loss deduction data is mapped to a graph node feature matrix, and a graph structure relationship matrix is constructed based on the topographic connectivity;

[0041] Based on the physical corrected observation feature, a physical constraint term is constructed, specifically:

[0042]

[0043] wherein, the physical constraint term is represented by, the soil erodibility factor is, the elevation is, the elevation gradient is represented by, the rainfall erosivity factor is, the constraint strength coefficient is, the square of the L2 norm is represented by;

[0044] The improved spatio-temporal graph convolutional network introduces a physical constraint term, specifically:

[0045] The physical constraint term and the graph node feature matrix are spliced in the feature dimension to form an extended node feature matrix; the graph structure relationship matrix is normalized to obtain a topographic connectivity normalized adjacency matrix, and the topographic connectivity normalized adjacency matrix and the extended node feature matrix are used to update the graph convolution to obtain an updated graph node feature matrix;

[0046] ​Based on the updated graph node feature matrix, time convolution is used for time modeling to obtain a predicted hidden feature matrix in a future preset time period; the future preset time period includes a plurality of prediction time steps, and each prediction time step corresponds to a prediction hidden feature vector;

[0047] For each prediction time step, the prediction hidden feature vector is generated into a soil erosion modulus through linear mapping, and the soil erosion modulus is rearranged into a grid form according to the corresponding relationship between the graph node and the spatial grid to obtain a soil and water loss intensity layer and a spatial distribution layer at the current prediction time step.

[0048] Optionally, the step six specifically includes:

[0049] The scenario driving factor includes a climate scenario parameter, a land use change parameter, and an engineering intervention measure parameter;

[0050] The climate scenario parameter includes a temperature change amount and a rainfall change amount, the land use change parameter represents a change in the spatial proportion of different land types, and the engineering intervention measure parameter includes spatial configuration quantitative indicators of vegetation restoration, terrace construction, and sand trap arrangement; the scenario driving factor is reduced to 10 main components through principal component analysis;

[0051] The reduced scenario driving factor and the spatio-temporal fusion feature tensor are spliced in the feature dimension to form a scenario condition feature tensor, and the scenario condition feature tensor is input into the optimized soil and water loss scenario generation module to generate soil and water loss deduction data corresponding to the scenario;

[0052] The soil and water loss deduction data is rearranged according to the spatial grid unit and the time sequence to construct a soil and water loss intensity layer sequence, and the soil and water loss intensity layer sequence includes a plurality of prediction time steps, and each prediction time step corresponds to a soil and water loss intensity grid layer;

[0053] Based on the soil and water loss intensity layer sequence, a soil and water loss evolution trend index is calculated, the soil and water loss evolution trend index includes a spatial cumulative erosion amount, an erosion center offset trajectory, and a high-risk area expansion rate, and a soil and water loss evolution trend map corresponding to the scenario is generated;

[0054] The soil and water loss intensity layer sequence and the soil and water loss evolution trend map are visually displayed.

[0055] Optionally, the step seven specifically includes:

[0056] The soil and water loss intensity grid layer in the soil and water loss intensity layer sequence is mapped to a corresponding geographic coordinate system according to the spatial grid unit to generate a geographic spatial layer;

[0057] Set the soil erosion risk level threshold, divide the soil erosion intensity value corresponding to each spatial grid cell into soil erosion risk levels according to the soil erosion risk level threshold, and label the levels based on the soil erosion modulus value in each threshold interval to generate a soil erosion risk level map;

[0058] The soil erosion risk level includes extremely low risk, low risk, medium risk, high risk and extremely high risk.

[0059] In the reasoning process, the Dropout layer of the soil erosion generation model is kept in an activated state, N forward reasonings are performed on the same model input, and N sets of soil erosion deduction data are obtained; the mean and variance of the N sets of soil erosion deduction data on each spatial grid cell are calculated, and a confidence interval is constructed according to the mean and variance to generate a Monte Carlo Dropout confidence interval map;

[0060] Bootstrap resampling is performed on the soil erosion data to obtain a plurality of sampling data sets; deduction is performed based on the plurality of sampling data sets to obtain a plurality of sets of soil erosion deduction data, and the distribution of the plurality of sets of soil erosion deduction data on each spatial grid cell is counted to calculate quantiles and generate a Bootstrap confidence interval map;

[0061] The soil erosion risk level map, the Monte Carlo Dropout confidence interval map and the Bootstrap confidence interval map are spatially visualized and displayed.

[0062] According to an embodiment of the soil erosion intelligent deduction system based on AIGC technology, comprising:

[0063] The soil erosion data acquisition module is used for acquiring soil erosion data of a target region.

[0064] The spatiotemporal feature construction module is used for data preprocessing of the soil erosion data and generating a spatiotemporal fusion feature tensor through a feature fusion network based on an attention mechanism.

[0065] The AIGC soil erosion modeling module is used for constructing a soil erosion generation model and generating soil erosion deduction data.

[0066] The dynamic feedback updating module is used for online updating of the soil erosion generation model.

[0067] The physical constraint graph convolution deduction module is used for future period deduction of the soil erosion deduction data and physical correction observation features to generate a soil erosion modulus.

[0068] The multi-scenario deduction module is used for constructing scenario condition feature tensors and inputting the scenario condition feature tensors into the optimized soil erosion scenario generation module to generate soil erosion deduction data corresponding to the scenarios.

[0069] Risk and uncertainty analysis module: for soil erosion risk classification, and using Monte Carlo Dropout and Bootstrap sampling to quantitatively analyze the uncertainty of soil erosion deduction data.

[0070] The beneficial effects of the present application are:

[0071] Firstly, the present application realizes the unified expression of soil erosion multi-scale features by solving the problems of inconsistent spatio-temporal resolution of remote sensing images, unmanned aerial vehicle images and sensor data, and heterogeneous semantic information, etc., through the collaborative collection and spatio-temporal alignment of multi-source heterogeneous observation data, and constructing a spatio-temporal fusion feature tensor based on an attention mechanism-based feature fusion network.

[0072] Secondly, the present application uses a soil erosion generation model based on AIGC technology, uses the spatio-temporal fusion feature tensor for conditional generative deduction, and constructs an adversarial training system of real observation data and soil erosion deduction data through a discriminant module, which effectively improves the authenticity, spatial detail expression ability and complex topography adaptability of the soil erosion deduction data in data scarce or observation gap areas.

[0073] In addition, the present application realizes online updating of the soil erosion scene generation module by establishing a dynamic feedback mechanism, so that the soil erosion generation model can continuously correct the generation distribution according to real-time observation data, and improve the dynamic adaptability of the deduction model in sudden rainfall, rapid runoff change and other scenarios. At the same time, combined with the improved spatio-temporal graph convolutional network with physical constraint term, the soil erosion modulus generated by deduction can maintain the expression ability of deep model while meeting the physical law of general soil erosion equation, so as to ensure the stability, physical consistency and interpretability of the deduction result.

[0074] In summary, the present application realizes intelligent deduction of soil erosion in multiple scenarios through scenario driving factors, and uses Monte Carlo Dropout and Bootstrap sampling for uncertainty quantitative analysis, which can output high-credibility soil erosion intensity layer sequence and soil erosion evolution trend graph under different climate, land use and engineering intervention scenarios, realize high-time-efficiency, high-robustness and high-credibility intelligent deduction of soil erosion process under complex scenarios, and has significant engineering application value and popularization significance. BRIEF DESCRIPTION OF DRAWINGS

[0075] The accompanying drawings are included 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. In the drawings:

[0076] Figure 1 is a method schematic diagram of a soil erosion intelligent deduction method and system based on AIGC technology according to the present application.

[0077] Figure 2 is a water and soil loss generation model structure flow chart in the water and soil loss intelligent deduction method and system based on AIGC technology proposed by the application;

[0078] Figure 3 is an improved spatiotemporal graph convolution network structure flow chart in the water and soil loss intelligent deduction method and system based on AIGC technology proposed by the application;

[0079] Figure 4 is a multi-scenario water and soil loss intelligent deduction flow chart in the water and soil loss intelligent deduction method and system based on AIGC technology proposed by the application. DETAILED DESCRIPTION

[0080] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show only the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0081] REFERENCE Figures 1-4 A water and soil loss intelligent deduction method based on AIGC technology, comprising the following steps:

[0082] Step one: collecting water and soil loss data of a target area;

[0083] Step two: performing spatiotemporal alignment and standardization processing on the water and soil loss data, obtaining a multi-scale spatiotemporal feature vector sequence, and using a feature fusion network based on an attention mechanism to perform weighted fusion on the multi-scale spatiotemporal feature vector sequence to generate a spatiotemporal fusion feature tensor;

[0084] Step three: constructing a water and soil loss generation model based on AIGC technology, the water and soil loss generation model comprising a water and soil loss scenario generation module and a discrimination module; inputting the spatiotemporal fusion feature tensor into the water and soil loss scenario generation module to generate water and soil loss deduction data, and performing adversarial training on real observation data and the water and soil loss deduction data through the discrimination module;

[0085] Step four: establishing a dynamic feedback mechanism to perform online updating on the water and soil loss scenario generation module to obtain an optimized water and soil loss generation model;

[0086] Step five: inputting the water and soil loss deduction data and physical correction observation features into an improved spatiotemporal graph convolution network to deduce and generate soil erosion modulus in a future preset time period, and outputting a water and soil loss intensity layer and a spatial distribution layer, wherein the improved spatiotemporal graph convolution network introduces a physical constraint term;

[0087] Step six: introduce scenario driving factors, and fuse with the spatio-temporal fusion feature tensor to input into the optimized soil erosion generation model, perform multi-scenario soil erosion intelligent deduction, generate soil erosion deduction data under different scenarios, and output soil erosion intensity layer sequence and soil erosion evolution trend graph;

[0088] Step seven: map the soil erosion intensity layer sequence to the geographic spatial layer sequence, and divide the soil erosion risk level; adopt Monte Carlo-based Dropout and Bootstrap sampling to perform uncertainty quantification analysis.

[0089] In the present application, step five quantitatively deduces the future soil erosion trend under a single scenario, and the core is to input the soil erosion deduction data and physical correction observation features into the improved spatio-temporal graph convolution network with a physical constraint term, generate soil erosion modulus in a future preset period by combining spatial topological structure and physical priori, output refined soil erosion intensity layer and spatial distribution layer, and realize high-precision simulation of short-term evolution trend. Step six focuses on multi-scenario simulation, introduces scenario driving factors and fuses with the spatio-temporal fusion feature tensor to input into the optimized soil erosion generation model, performs multi-scenario simulation, generates soil erosion deduction data under multiple conditions, and outputs soil erosion intensity layer sequence and evolution trend graph covering multiple hypothetical scenarios, aiming to reveal the influence of different external condition changes on the soil erosion process.

[0090] In the present embodiment, the step one specifically comprises:

[0091] The soil erosion data of the target region is collected by a satellite remote sensing platform, an unmanned aerial vehicle aerial survey system and a ground sensor network, and the soil erosion data includes high-resolution multispectral images, laser radar point clouds, soil moisture, rainfall intensity, ground roughness and vegetation coverage.

[0092] In the present embodiment, the step two specifically comprises:

[0093] Atmospheric correction and geometric correction are performed on the high-resolution multispectral images, and normalized vegetation index, slope, slope direction and soil brightness index are extracted;

[0094] The laser radar point cloud is reconstructed in three dimensions to generate a digital surface model, and the ground roughness feature and the fine-scale vegetation coverage feature are calculated based on the moving window method;

[0095] The soil moisture, rainfall intensity, ground roughness and vegetation coverage are subjected to outlier rejection, unit unification and time stamp synchronization processing to obtain a standardized hydrological physical quantity sequence;

[0096] The normalized vegetation index, the slope, the aspect and the soil brightness index, the surface roughness feature, the fine-scale vegetation coverage feature and the standardized hydrological physical quantity sequence are resampled and interpolated according to the spatial position and the time label to obtain a multi-scale feature vector sequence, with the spatial grid unit and the time step as indexes;

[0097] The multi-scale feature vector sequence is input into a feature fusion network based on an attention mechanism, attention weights are assigned to the feature vectors from different data sources and weighted fusion is performed, and a spatio-temporal fusion feature tensor is generated.

[0098] In the embodiment, the step three specifically comprises:

[0099] The soil erosion scene generation module adopts a conditional variational autoencoder structure, learns the latent distribution of the soil erosion process and generates soil erosion deduction data;

[0100] The encoder of the soil erosion scene generation module comprises four down-sampling stages, each down-sampling stage extracts local spatial features in the spatio-temporal fusion feature tensor through a 3x3 convolution layer to obtain a down-sampling feature map, and reduces the spatial dimension of the down-sampling feature map through a 2x2 max-pooling layer; the number of channels of the encoder is multiplied by a factor of two at each down-sampling stage, from an initial 64 to 128, 256 and 512, for enhancing the semantic expression capability of deep features;

[0101] The decoder of the soil erosion scene generation module is provided with four up-sampling stages, each up-sampling stage gradually recovers the spatial resolution through a transposed convolution layer, and adopts a skip connection to concatenate the down-sampling feature map of each down-sampling stage of the encoder with the feature map of the corresponding up-sampling stage in the channel to generate an up-sampling fusion feature map, which is used as the input feature map of the next up-sampling stage;

[0102] The discrimination module adopts a multi-scale convolutional neural network structure, comprising a first scale discrimination branch, a second scale discrimination branch and a third scale discrimination branch; each scale discrimination branch is sequentially stacked by a plurality of 3x3 convolution layers, batch normalization layers and LeakyReLU activation layers, and is provided with different down-sampling times within each scale discrimination branch;

[0103] Original resolution images, 2x down-sampling images and 4x down-sampling images of the real observation data and the soil erosion deduction data are respectively constructed;

[0104] The first scale discrimination branch receives the original resolution images of the real observation data and the original resolution images of the soil erosion deduction data respectively, outputs first scale real discrimination feature maps and first scale generated discrimination feature maps, and performs per-pixel discrimination through a 1x1 convolution layer to obtain first scale real response maps and first scale generated response maps;

[0105] The second scale discrimination branch respectively receives 2 times down-sampling images of real observation data and 2 times down-sampling images of soil and water loss deduction data, outputs second scale real discrimination feature maps and second scale generated discrimination feature maps, and adopts a 1*1 convolution layer to perform per-pixel discrimination to obtain second scale real response maps and second scale generated response maps.

[0106] The third scale discrimination branch respectively receives 4 times down-sampling images of real observation data and 4 times down-sampling images of soil and water loss deduction data, outputs third scale real discrimination feature maps and third scale generated discrimination feature maps, and adopts a 1*1 convolution layer to perform per-pixel discrimination to obtain third scale real response maps and third scale generated response maps.

[0107] The loss function of the adversarial training includes a least square loss and a feature matching loss.

[0108] The least square loss is to calculate the expectation of the square difference of the three scale real response maps and 1, the expectation of the square of the three scale generated response maps, and then add them after averaging.

[0109] The feature matching loss is to calculate the expectation of the absolute difference of the three scale real discrimination feature maps and the generated discrimination feature maps, and then add them.

[0110] In the embodiment, the dynamic feedback mechanism specifically includes: setting a sliding time window, continuously receiving real-time observation data, and calculating the root mean square error between the soil and water loss deduction data and the real-time observation data as a performance loss function; based on the performance loss function, the parameters of the soil and water loss scene generation module are updated using a self-adaptive matrix estimation optimization algorithm, and the initial value of the learning rate is 0.001, and is exponentially attenuated with the training round or error convergence.

[0111] In the embodiment, the step five specifically includes:

[0112] According to the soil and water loss data, a general soil loss equation is used to construct a physical corrected observation feature, and the physical corrected observation feature includes a rainfall erosivity factor, a soil erodibility factor and an elevation;

[0113] The soil and water loss deduction data is mapped into a graph node feature matrix, and a graph structure relationship matrix is constructed based on the topographic connectivity;

[0114] Based on the physical corrected observation feature, a physical constraint term is constructed, specifically:

[0115] ;

[0116] wherein, represents the physical constraint term, is the soil erodibility factor, is an elevation, represents an elevation gradient, is a rainfall erosivity factor, is a constraint intensity coefficient;

[0117] The improved spatio-temporal graph convolution network introduces a physical constraint term to guide the improved spatio-temporal graph convolution network to update the graph node features, so that the prediction result is more in line with the true soil and water loss mechanism; the physical constraint term is used to construct a physical residual signal in the graph convolution operation, and the physical residual signal is taken as an additional input for updating the graph node features, so that the graph node features are constrained by the coupling relationship between the terrain gradient, soil erodibility and rainfall erosivity during spatial propagation, specifically:

[0118] The physical constraint term and the graph node feature matrix are spliced in the feature dimension to form an extended node feature matrix; the graph structure relationship matrix is normalized to obtain a terrain connectivity normalized adjacency matrix, and graph convolution updating is performed based on the terrain connectivity normalized adjacency matrix and the extended node feature matrix to obtain an updated graph node feature matrix;

[0119] Based on the updated graph node feature matrix, time convolution is used for time modeling to obtain a prediction hidden feature matrix in a future preset time period; the future preset time period includes a plurality of prediction time steps, and each prediction time step corresponds to a prediction hidden feature vector;

[0120] For each prediction time step, the prediction hidden feature vector is generated into soil erosion modulus through linear mapping, and the soil erosion modulus is rearranged into a grid form according to the corresponding relationship between the graph node and the spatial grid to obtain a soil and water loss intensity layer and a spatial distribution layer at the current prediction time step.

[0121] In this embodiment, step six specifically includes:

[0122] The scenario driving factor includes climate scenario parameters, land use change parameters and engineering intervention measure parameters;

[0123] The climate scenario parameters include temperature change and rainfall change, the land use change parameters represent the change of the spatial proportion of different land types, and the engineering intervention measure parameters include spatial configuration quantitative indicators of vegetation restoration, terrace construction and sand trap arrangement; the scenario driving factor is reduced to 10 main components through principal component analysis;

[0124] The reduced scenario driving factor and the spatio-temporal fusion feature tensor are spliced in the feature dimension to form a scenario condition feature tensor, and the scenario condition feature tensor is input into the optimized soil and water loss scenario generation module to generate soil and water loss deduction data corresponding to the scenario;

[0125] The soil erosion deduction data is rearranged according to spatial grid units and time series to construct a soil erosion intensity layer sequence, the soil erosion intensity layer sequence comprising a plurality of prediction time steps, each prediction time step corresponding to a soil erosion intensity grid layer;

[0126] Based on the soil erosion intensity layer sequence, a soil erosion evolution trend index is calculated, the soil erosion evolution trend index including spatial cumulative erosion amount, erosion center shift trajectory and high-risk area expansion rate, and a soil erosion evolution trend map corresponding to the scenario is generated;

[0127] The soil erosion intensity layer sequence and the soil erosion evolution trend map are visually displayed.

[0128] In this embodiment, the step seven specifically includes:

[0129] The soil erosion intensity grid layer in the soil erosion intensity layer sequence is mapped to the corresponding geographic coordinate system according to the spatial grid unit to generate a geographic space layer;

[0130] A soil erosion risk level threshold is set, the soil erosion intensity value corresponding to each spatial grid unit is divided into soil erosion risk levels according to the soil erosion risk level threshold, and the soil erosion risk level is labeled based on the soil erosion modulus value in each threshold interval to generate a soil erosion risk level map; the soil erosion risk level includes extremely low risk, low risk, medium risk, high risk and extremely high risk,

[0131] During the reasoning process, the Dropout layer of the soil erosion generation model is kept in an active state, N forward reasoning is performed on the same model input to obtain N groups of soil erosion deduction data; the mean and variance of the N groups of soil erosion deduction data on each spatial grid unit are calculated, and a confidence interval is constructed according to the mean and variance to generate a Monte Carlo Dropout confidence interval map;

[0132] The soil erosion data is Bootstrap resampled to obtain a plurality of sampling data sets; based on the plurality of sampling data sets, deduction is performed respectively to obtain a plurality of groups of soil erosion deduction data, and the distribution of the plurality of groups of soil erosion deduction data on each spatial grid unit is counted to calculate quantiles and generate a Bootstrap confidence interval map;

[0133] The soil erosion risk level map, the Monte Carlo Dropout confidence interval map and the Bootstrap confidence interval map are spatially visually displayed.

[0134] An intelligent soil erosion deduction system based on AIGC technology, comprising:

[0135] A soil erosion data acquisition module for acquiring soil erosion data of a target area;

[0136] spatiotemporal feature construction module: for data preprocessing of soil and water loss data and generating spatiotemporal fusion feature tensor through attention mechanism-based feature fusion network;

[0137] AIGC soil and water loss modeling module: for constructing a soil and water loss generation model to generate soil and water loss deduction data;

[0138] dynamic feedback updating module: for online updating of the soil and water loss generation model;

[0139] physical constraint graph convolution deduction module: for future period deduction of soil and water loss deduction data and physical correction observation features to generate soil erosion modulus;

[0140] multi-scenario deduction module: for constructing scenario condition feature tensor and inputting into the optimized soil and water loss scenario generation module to generate soil and water loss deduction data corresponding to the scenario;

[0141] risk and uncertainty analysis module: for soil and water loss risk grade division, and uncertainty quantification analysis of soil and water loss deduction data by using Monte Carlo Dropout and Bootstrap sampling.

[0142] Embodiment 1

[0143] In order to verify the feasibility of the application in implementation, the method of the application is applied to a typical watershed in a hilly area. The terrain of the watershed is large in relief, the slope is mainly distributed between 15° and 35°, the duration of the rainstorm season is long, and the vegetation coverage height depends on artificial restoration. In this area, there are problems such as sparse observation data all year round, long period of unmanned aerial survey, and significant cloud interference on satellite remote sensing. Traditional soil and water loss monitoring mode generally faces difficulties such as large deduction deviation, slow response, and inability to depict fine-scale erosion details in such scenarios. The application mainly evaluates its performance under complex topographic conditions, insufficient observation data, and obvious climate change scenarios.

[0144] In actual deployment, 10-meter resolution multispectral images are obtained through a satellite remote sensing platform, and more than 20 high-resolution images from unmanned aerial surveys in the past three years are collected, while real-time observation data such as rainfall intensity, soil moisture, and runoff are provided by ground sensors. Due to the low temporal resolution of satellite data, the discontinuous coverage of unmanned aerial data, and the large spatiotemporal scale difference between multi-source data, the application first uses an attention mechanism fusion network to unify these data into a spatiotemporal fusion feature tensor, which can express both spatial texture features and temporal dynamic change trends. In contrast, traditional methods usually use only a single data source or simple alignment, making it difficult to accurately reflect the erosion state during rainstorm processes or human disturbances.

[0145] In this scenario, the soil erosion generation model constructed based on AIGC technology is used to complete the erosion scene data in the missing area. For example, in the canyon area that cannot be covered by unmanned aerial vehicle photogrammetry, the AIGC generation model can generate soil erosion deduction images consistent with the real terrain, and the authenticity is verified by the discrimination module, so that the slope direction texture and erosion gully morphology of the generated scene are highly consistent with the actual observation. The generation model introduces a multi-scale discrimination structure in the adversarial training, which can capture multi-level information from micro-scale gully erosion texture to macro-scale slope erosion distribution, greatly improving the data consistency and detail fidelity.

[0146] To adapt to extreme rainfall events, the invention maintains a dynamic feedback mechanism in actual operation, enabling the model to automatically update parameters after receiving real-time sensor observation data. For example, during a 112mm / h short-term rainstorm event, the soil moisture at the top of the slope was monitored to rise to saturation within 25 minutes, and the model immediately incorporated this change into the parameter update, enabling the generated scene and subsequent deduction results to match the real erosion dynamics during the rainstorm. Traditional static models often take hours to complete re-simulation, making it difficult to meet the rapid warning needs. To ensure the physical credibility of the deduction results, the invention inputs the physical correction observation features and soil erosion deduction data into a spatio-temporal graph convolution network with a physical constraint term, enabling the model to automatically follow the soil erosion mechanism when predicting future erosion trends. For example, when the rainfall intensity increases sharply, the model will automatically increase the influence of the physical constraint term, causing the prediction curve to show a sharp rise consistent with the measured data, avoiding the smoothing prediction problem that pure deep learning models are prone to.

[0147] To further evaluate the adaptability of the invention under future change scenarios, the system constructs a multi-scenario deduction mechanism, sets different climate change parameters, land use change parameters, and engineering measure parameters, and simulates the soil erosion modulus change trend under three typical scenarios of continuous rainfall increase, farmland expansion, and artificial vegetation restoration in the next five years. The results show that the invention can stably output consistent physical logic erosion evolution process under various scenarios, enabling users to identify high-risk areas in advance and optimize the layout of engineering measures.

[0148] To verify the improvement effect of the invention in actual application, 120 typical slopes in the same watershed are selected as test samples, and the invention is compared with the USLE-R physical model and the CNN-based deep learning method. The comparison indicators include average error of erosion modulus, spatial matching degree, fine-scale gully erosion recognition accuracy, deduction response time, stability under data missing conditions, and uncertainty interval width. The experimental results are shown in Table 1.

[0149] Table 1 Performance comparison results of the invention and the comparison schemes in soil erosion deduction task

[0150]

[0151] From the data in Table 1, it can be seen that the present application is obviously better than the comparative schemes in multiple comparison indicators. In terms of erosion modulus prediction accuracy, the average error of the present application is only 12.4 t / ha·yr, which is reduced by 55.4% compared with the USLE-R physical model and reduced by 36.7% compared with the deep learning scheme based on CNN; in terms of spatial matching degree, the IoU of the present application reaches 0.87, which can more accurately restore the spatial pattern of slope erosion; in terms of fine-scale gully erosion recognition performance, the present application is more than 20 percentage points higher than the USLE-R physical model. In addition, in terms of response speed of deduction, the present application relies on the soil and water loss generation model and the improved spatio-temporal graph convolution network, and only 3.6 seconds are needed to complete a complete deduction, which greatly shortens the calculation time. In the data missing scene, the present application automatically fills the gap area through the soil and water loss generation model, so that the stability is improved to 94.5%; the uncertainty interval width is significantly reduced, indicating that the output of the soil and water loss generation model is more stable and reliable.

[0152] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for intelligent simulation of soil erosion based on AIGC technology, characterized in that, Includes the following steps: Step 1: Collect soil and water loss data for the target area; Step 2: Perform spatiotemporal alignment and standardization on the soil erosion data to obtain a multi-scale spatiotemporal feature vector sequence, and use an attention-based feature fusion network to perform weighted fusion of the multi-scale spatiotemporal feature vector sequence to generate a spatiotemporal fusion feature tensor. Step 3: Construct a soil erosion generation model based on AIGC technology. The soil erosion generation model includes a soil erosion scene generation module and a discrimination module. Input the spatiotemporal fusion feature tensor into the soil erosion scene generation module to generate soil erosion simulation data, and conduct adversarial training between the real observation data and the soil erosion simulation data through the discrimination module. Step 4: Establish a dynamic feedback mechanism to update the soil erosion scenario generation module online and obtain an optimized soil erosion generation model; Step 5: Input the soil erosion projection data and the physically corrected observation features into the improved spatiotemporal graph convolutional network to generate the soil erosion modulus for the future preset time period, and output the soil erosion intensity layer and spatial distribution layer. The improved spatiotemporal graph convolutional network introduces physical constraint terms. Step 6: Introduce scenario-driven factors and fuse them with spatiotemporal fusion feature tensors. Input the fusion into the optimized soil and water loss generation model, perform multi-scenario intelligent soil and water loss simulation, generate soil and water loss simulation data under different scenarios, and output soil and water loss intensity layer sequence and soil and water loss evolution trend map. Step 7: Map the soil erosion intensity layer sequence to a geospatial layer sequence and classify the soil erosion risk level; use Monte Carlo-based Dropout and Bootstrap sampling to perform uncertainty quantification analysis.

2. The intelligent soil erosion simulation method based on AIGC technology according to claim 1, characterized in that, Step one specifically includes: Soil and water loss data of the target area are collected collaboratively by satellite remote sensing platform, UAV aerial survey system and ground sensor network. The soil and water loss data includes high-resolution multispectral imagery, lidar point cloud, soil moisture, rainfall intensity, surface roughness and vegetation coverage.

3. The intelligent simulation method for soil erosion based on AIGC technology according to claim 1, characterized in that, Step two specifically includes: Atmospheric and geometric corrections were performed on the high-resolution multispectral images to extract the normalized vegetation index, slope, aspect, and soil brightness index. The lidar point cloud is reconstructed in three dimensions to generate a digital surface model, and the surface roughness features and fine-scale vegetation cover features are calculated based on the moving window method. Outlier removal, unit standardization, and timestamp synchronization were performed on soil moisture, rainfall intensity, surface roughness, and vegetation cover to obtain a standardized hydrophysical quantity sequence. Using spatial grid cells and time steps as indices, normalized vegetation index, slope, aspect and soil brightness index, surface roughness characteristics, fine-scale vegetation cover characteristics and standardized hydrophysical quantity sequences are resampled and interpolated according to spatial location and time label to obtain multi-scale feature vector sequences. A multi-scale feature vector sequence is input into an attention-based feature fusion network. Attention weights are assigned to feature vectors from different data sources and weighted fusion is performed to generate a spatiotemporal fusion feature tensor.

4. The intelligent simulation method for soil erosion based on AIGC technology according to claim 1, characterized in that, Step three specifically includes: The soil erosion scenario generation module adopts a conditional variational autoencoder structure, which learns the potential distribution of the soil erosion process and generates soil erosion inference data. The encoder of the soil erosion scene generation module includes four downsampling stages. Each downsampling stage extracts local spatial features from the spatiotemporal fusion feature tensor through a 3×3 convolutional layer to obtain a downsampled feature map. The spatial dimension of the downsampled feature map is reduced through a 2×2 max pooling layer. The number of channels of the encoder increases exponentially with each downsampling stage, from the initial 64 to 128, 256 and 512, to enhance the semantic expressive power of deep features. The decoder of the soil erosion scene generation module is set with four upsampling stages. In each upsampling stage, the spatial resolution is gradually restored through transposed convolutional layers. Skip connections are used to concatenate the downsampled feature maps of each downsampling stage of the encoder with the feature maps of the corresponding upsampling stages to generate upsampled fusion feature maps, which are then used as input feature maps for the next upsampling stage. The discrimination module adopts a multi-scale convolutional neural network structure, including a first-scale discrimination branch, a second-scale discrimination branch, and a third-scale discrimination branch; each scale discrimination branch is composed of several 3×3 convolutional layers, batch normalization layers, and LeakyReLU activation layers stacked sequentially, and different downsampling times are set in each scale discrimination branch; Original resolution images, 2x downsampled images, and 4x downsampled images were constructed for real observation data and soil erosion projection data, respectively. The first scale discrimination branch receives the original resolution image of the real observation data and the original resolution image of the soil erosion inference data respectively, outputs the first scale real discrimination feature map and the first scale generated discrimination feature map, and uses a 1×1 convolutional layer to perform pixel-by-pixel discrimination to obtain the first scale real response map and the first scale generated response map. The second-scale discrimination branch receives a 2x downsampled image of the real observation data and a 2x downsampled image of the soil erosion inference data, respectively, and outputs a second-scale real discrimination feature map and a second-scale generated discrimination feature map. It then uses a 1×1 convolutional layer to perform pixel-by-pixel discrimination to obtain a second-scale real response map and a second-scale generated response map. The third-scale discrimination branch receives a 4x downsampled image of the real observation data and a 4x downsampled image of the soil erosion inference data, respectively, and outputs a third-scale real discrimination feature map and a third-scale generated discrimination feature map. It then uses a 1×1 convolutional layer to perform pixel-by-pixel discrimination to obtain a third-scale real response map and a third-scale generated response map. The loss function for adversarial training includes least squares loss and feature matching loss; The least squares loss is calculated by: calculating the expected value of the squared difference between the true response map at each of the three scales and 1, and the expected value of the squared difference of the generated response map at each of the three scales, and then averaging and summing them. The feature matching loss is the expectation of the element-wise absolute difference between the real discriminant feature maps and the generated discriminant feature maps at three scales, and then summed.

5. The intelligent simulation method for soil erosion based on AIGC technology according to claim 1, characterized in that, The dynamic feedback mechanism specifically includes: setting a sliding time window, continuously receiving real-time observation data, and calculating the root mean square error between the soil erosion simulation data and the real-time observation data as a performance loss function; based on the performance loss function, using an adaptive moment estimation optimization algorithm to update the parameters of the soil erosion scenario generation module, with an initial learning rate of 0.001, which decays exponentially with the training rounds or error convergence.

6. The intelligent simulation method for soil erosion based on AIGC technology according to claim 1, characterized in that, Step five specifically includes: Based on soil erosion data, a physical correction observation feature was constructed using a general soil loss equation. The physical correction observation feature includes rainfall erosivity factor, soil erodibility factor, and elevation. The soil erosion simulation data is mapped into a graph node feature matrix, and a graph structure relationship matrix is ​​constructed based on terrain connectivity. Based on physically corrected observation characteristics, physical constraint terms are constructed, specifically as follows: ； in, Represents physical constraint terms. As a soil erodibility factor, For elevation, Indicates elevation gradient, As the erosivity factor of rainfall, The constraint strength coefficient, Represents the square of the L2 norm; The improved spatiotemporal graph convolutional network introduces physical constraint terms, specifically: The physical constraint terms are concatenated with the graph node feature matrix in the feature dimension to form an extended node feature matrix; the graph structure relation matrix is ​​normalized to obtain the terrain connectivity normalized adjacency matrix; and graph convolution is performed based on the terrain connectivity normalized adjacency matrix and the extended node feature matrix to obtain the updated graph node feature matrix. Based on the updated graph node feature matrix, temporal convolution is used for time modeling to obtain the predicted hidden feature matrix for a future preset time period; the future preset time period includes several prediction time steps, and each prediction time step corresponds to a predicted hidden feature vector. For each prediction time step, the predicted hidden feature vector is linearly mapped to generate the soil erosion modulus, and the soil erosion modulus is rearranged into a raster form according to the correspondence between graph nodes and spatial grids to obtain the soil erosion intensity layer and spatial distribution layer for the current prediction time step.

7. The intelligent simulation method for soil erosion based on AIGC technology according to claim 1, characterized in that, Step six specifically includes: The scenario-driving factors include climate scenario parameters, land use change parameters, and engineering intervention parameters. The climate scenario parameters include temperature and rainfall changes; the land use change parameters represent the spatial proportion changes of different land types; the engineering intervention parameters include spatial configuration quantitative indicators for vegetation restoration, terrace construction, and silt-trapping dam deployment; and the scenario driving factors are reduced to 10 principal components through principal component analysis. The scenario-driven factors and the spatiotemporal fusion feature tensor after dimensionality reduction are concatenated in the feature dimension to form the scenario condition feature tensor. The scenario condition feature tensor is then input into the optimized soil and water loss scenario generation module to generate soil and water loss inference data for the corresponding scenario. The soil erosion projection data is rearranged according to spatial grid units and time series to construct a soil erosion intensity layer sequence. The soil erosion intensity layer sequence contains multiple prediction time steps, and each prediction time step corresponds to a soil erosion intensity raster layer. Based on the soil erosion intensity layer sequence, the soil erosion evolution trend index is calculated. The soil erosion evolution trend index includes the spatial cumulative erosion, the erosion center offset trajectory and the expansion rate of high-risk areas, and a soil erosion evolution trend map corresponding to the scenario is generated. The sequence of soil erosion intensity layers and the trend map of soil erosion evolution are visualized.

8. The intelligent simulation method for soil erosion based on AIGC technology according to claim 1, characterized in that, Step seven specifically includes: The soil erosion intensity raster layer in the soil erosion intensity layer sequence is mapped to the corresponding geographic coordinate system according to the spatial grid unit to generate a geospatial layer. Set a threshold for the risk level of soil and water loss, classify the soil and water loss risk level according to the threshold for the soil and water loss intensity value corresponding to each spatial grid unit, and label the level within each threshold range based on the soil erosion modulus value to generate a soil and water loss risk level map. The soil and water loss risk levels include extremely low risk, low risk, medium risk, high risk and extremely high risk; During the inference process, the Dropout layer of the soil erosion generation model is kept active. N forward inferences are performed on the same model input to obtain N sets of soil erosion simulation data. The mean and variance of the N sets of soil erosion simulation data are calculated on each spatial grid cell, and confidence intervals are constructed based on the mean and variance to generate a Monte Carlo Dropout confidence interval map. Bootstrap resampling is performed on the soil erosion data to obtain several sampled datasets; inference is performed on the several sampled datasets to obtain several sets of soil erosion inference data, and the distribution of the several sets of soil erosion inference data on each spatial grid cell is statistically analyzed, quantiles are calculated, and Bootstrap confidence interval plots are generated. Spatial visualization of soil erosion risk level map, Monte Carlo Dropout confidence interval map, and Bootstrap confidence interval map.

9. A soil erosion intelligent simulation system based on AIGC technology, executing the soil erosion intelligent simulation method based on AIGC technology as described in any one of claims 1 to 8, characterized in that, include: Soil and water loss data acquisition module: used to collect soil and water loss data in the target area; Spatiotemporal feature construction module: used to preprocess soil erosion data and generate spatiotemporal fusion feature tensors through a feature fusion network based on an attention mechanism; AIGC Soil and Water Loss Modeling Module: Used to build soil and water loss generation models and generate soil and water loss projection data; Dynamic feedback update module: used to update the soil and water loss generation model online; Physical constraint graph convolutional extrapolation module: used to extrapolate future time periods from soil erosion extrapolation data and physically corrected observation features, and generate soil erosion modulus; Multi-scenario simulation module: used to construct scenario condition feature tensors and input them into the optimized soil and water loss scenario generation module to generate soil and water loss simulation data for the corresponding scenarios; Risk and Uncertainty Analysis Module: Used to classify soil erosion risk levels and employ Monte Carlo Dropout and Bootstrap sampling to perform uncertainty quantification analysis on soil erosion projection data.

Images