A method for evaluating soil conservation function of a watershed based on deep learning

Abstract

The application discloses a kind of based on deep learning's watershed soil conservation function evaluation method, comprising: through water and soil conservation dynamic intelligent monitoring analysis platform acquisition watershed topography, vegetation cover, precipitation intensity and so on multidimensional core parameter, constructs original data set;The data set is input into multi-modal watershed erosion perception deep learning model to extract key feature vector, utilize the mapping relationship of space-time correlation soil holding prediction model to construct soil holding capacity and space-time variable, adopt runoff sediment coupling evaluation algorithm to complete runoff intensity and sediment transport amount collaborative calculation;Integrate core index in combination with model output and coupling result, construct comprehensive evaluation system, realize watershed different regional soil conservation function grade division and quantitative evaluation.This method is through multi-source data fusion, deep learning modeling and collaborative calculation, improves the comprehensiveness and accuracy of evaluation, provides scientific and reliable decision-making basis for watershed water and soil conservation planning, erosion control and ecological restoration.

Description

Technical Field

[0001] This invention relates to the field of watershed soil conservation technology, and in particular to a deep learning-based method for evaluating watershed soil conservation function. Background Technology

[0002] Watershed soil conservation function assessment is a core foundation for ecological protection and soil and water conservation, and its accuracy directly affects the formulation of watershed ecological restoration plans and the sustainable use of water and soil resources. Currently, global watersheds face multiple challenges, including complex topography, uneven spatial and temporal distribution of precipitation, and significant differences in vegetation cover. Traditional assessment methods, relying on single data sources and empirical analysis, are ill-suited to the complex watershed scenarios involving multiple coupled factors. With the escalating demands for ecological protection, there is an urgent need to develop technical methods that can integrate multi-source data, capture spatiotemporal dynamic changes, and achieve precise quantitative assessment, providing scientific support for watershed soil and water conservation planning, erosion control, and functional optimization. This has become a core research direction and practical application requirement in this field. Currently, watershed soil conservation function assessment is mainly completed through manual monitoring, inversion of single remote sensing data, or simple model calculations. Typically, limited basic data such as topography, precipitation, and vegetation are first collected, then simply processed and input into traditional empirical models. Interpolation analysis is then performed using a small amount of ground observation data to classify soil conservation function levels. Although some technologies have introduced basic monitoring equipment, the data collection dimensions are limited and the spatiotemporal coverage is incomplete. The evaluation process relies heavily on static parameters and linear calculation logic, lacking an effective characterization of the dynamic correlation of key processes such as watershed erosion, soil retention, and runoff sediment transport. Ultimately, the evaluation results are output through simple weighting or hierarchical statistics.

[0003] Existing technologies suffer from two major drawbacks: First, they lack data integration and perception capabilities, relying heavily on single-type data sources and failing to effectively integrate key watershed parameters across multiple dimensions and spatiotemporal scales. They also exhibit lag in perceiving dynamic changes in factors such as topography, vegetation, and precipitation, resulting in insufficient comprehensiveness and timeliness of the evaluation data. Second, the evaluation logic lacks adaptability to real-world scenarios, failing to fully consider the complex interrelationships and spatiotemporal heterogeneity of key processes within the watershed. Evaluation results are easily influenced by local parameters, making it difficult to accurately reflect the true differences in soil conservation functions across different regions. Furthermore, the lack of dynamic adjustment and precise calibration mechanisms limits the reliability and application value of the evaluation results. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a watershed soil conservation function evaluation method based on deep learning.

[0005] The technical solution adopted in this invention is a method for evaluating watershed soil conservation function based on deep learning, comprising the following steps: S1, collecting parameter data on watershed topography, vegetation cover, precipitation intensity, soil texture, land use type, slope and slope length through a dynamic intelligent monitoring and analysis platform for soil and water conservation, and constructing a multi-dimensional original dataset; S2, inputting the dataset into a multimodal watershed erosion perception deep learning model, and extracting watershed erosion feature vectors through feature fusion and hierarchical perception mechanisms; S3, based on the extracted feature vectors, constructing a mapping relationship between watershed soil holding capacity and spatiotemporal variables using a spatiotemporal correlation soil holding prediction model; S4, using a runoff-sediment coupling assessment algorithm to collaboratively calculate watershed runoff intensity and sediment transport, generating coupling assessment results; S5, combining the model output and coupling assessment results, integrating the indicators for watershed soil conservation function evaluation, and constructing a comprehensive evaluation system; S6, using this system to complete the classification and quantitative assessment of soil conservation function levels in different areas of the watershed, and outputting an evaluation report.

[0006] Furthermore, the modeling process of the multimodal watershed erosion perception deep learning model in S2 includes: selecting watershed vegetation cover, precipitation erosivity, soil erodibility, slope, and slope length as model input modal parameters; constructing a multi-source data input matrix; and performing preliminary feature extraction on each modal data through a convolutional neural network to obtain a single-modal feature matrix. ,in This is the preliminary characteristic matrix of the m-th mode. This is the raw data for the m-th mode. The convolution kernel weight matrix is... Here is the modal bias matrix. To output the bias term, an attention mechanism is used to calculate the weight coefficients of each modality feature, and the multimodal fusion feature is obtained through weighted fusion. ,in These are the weighting coefficients for the m-th mode. Here, n is the attention weight matrix, and n is the total number of modes. A deep residual network is constructed to extract the fusion features in an advanced manner. The gradient vanishing problem is alleviated by residual connections, and the propagation and retention of erosion features are enhanced. Simultaneously, parameters such as watershed soil porosity and erosion resistance age are introduced to optimize the network training process. Based on the extracted high-order feature vectors, they are mapped to the erosion intensity prediction space through fully connected layers to complete the model training and optimization, ensuring the model's adaptability and perception accuracy for different erosion scenarios in the watershed.

[0007] Furthermore, the modeling process of the spatiotemporal correlation soil retention prediction model in S3 includes: selecting spatiotemporal sequence data of watershed soil organic matter content, root density, soil clay content, and vegetation cover as input parameters; constructing a spatiotemporal data tensor; extracting temporal correlation features through a temporal convolutional network; and simultaneously capturing spatial neighborhood correlation information using a graph convolutional network; constructing a soil retention capacity prediction model based on spatiotemporal features; and introducing a spatiotemporal attention factor to establish a nonlinear mapping between features and soil retention capacity, expressed as: ,in This is a predicted value for soil sequestration. For spatial characteristic matrix, For graph convolution weights, For temporal convolution weights, This is the spatiotemporal adjustment coefficient. This is a spatiotemporal correlation function; combined with the topographic complexity and spatiotemporal distribution differences of precipitation in different regions of the watershed, the model prediction results are regionally calibrated, and the spatiotemporal sensitive parameters of the model are adjusted through iterative optimization to improve the spatiotemporal consistency and accuracy of soil retention capacity prediction under different watershed scenarios.

[0008] Furthermore, the calculation process of the runoff-sediment coupling assessment algorithm in S4 includes: selecting total watershed precipitation, runoff coefficient, surface roughness, soil infiltration rate, and gradient as runoff calculation parameters; constructing a runoff intensity calculation model; and performing distributed calculation of watershed runoff through hydrological response unit division; based on the runoff calculation results, and combining sediment particle size distribution, soil erosion modulus, and vegetation interception coefficient, constructing a sediment transport calculation module, whose runoff intensity calculation expression is as follows: ,in For runoff intensity, For the precipitation data of time period i, Let be the runoff coefficient for the i-th time period. Let be the length of the confluence during the i-th time period. Let j be the area of ​​the j-th hydrological unit. Let be the infiltration rate of the j-th unit. The runoff correction coefficient is used; a runoff-sediment coupling model is constructed, and a quantitative correlation between runoff intensity and sediment transport is established through the coupling coefficient. The coupling expression is as follows: ,in This refers to sediment transport. The coupling coefficient is... The content of sediment particles in Class I sediment is... The coefficient for particle transport of type I is used. The parameters of watershed topographic relief and vegetation buffer zone width are introduced to correct the coupling results. The influence of different land use types on runoff and sediment transport is considered. The coupling model parameters are optimized through iterative calculation to ensure the consistency between the calculation results of runoff intensity and sediment transport, and to provide accurate basic data on runoff and sediment for watershed soil conservation function evaluation.

[0009] Furthermore, the dynamic intelligent monitoring and analysis platform for soil and water conservation includes a multispectral remote sensing monitoring module, a ground sensor network, an unmanned aerial vehicle (UAV) survey unit, and a data fusion processing center. The multispectral remote sensing monitoring module uses a high-resolution multispectral camera with a spectral resolution range of 400-900nm and a spatial resolution better than 5m. The ground sensor network deploys soil moisture sensors, precipitation sensors, and runoff sensors, with a sampling frequency of 15 minutes / time and a data transmission latency of no more than 3 seconds. The UAV survey unit is equipped with a lidar and a thermal infrared imager, with a survey coverage radius of no less than 10km and a data acquisition accuracy better than 0.1m. The data fusion processing center adopts a GPU cluster architecture, including 8 high-performance computing graphics cards, supporting parallel processing of multimodal data, with a data storage capacity of no less than 100TB and a computational latency of no more than 10 seconds for model training and inference.

[0010] Further, S2 includes the following sub-steps: S21, embedding watershed soil structure stability and vegetation root soil-fixing strength parameters into the multimodal input data of watershed vegetation cover, precipitation erosivity, and soil erodibility, constructing a high-dimensional feature space including spatial distribution characteristics and temporal dynamic change information, and organically integrating multi-dimensional parameters; S22, using deep separateable convolution to perform layered processing on the expanded high-dimensional features, extracting inter-channel correlation information through point convolution, and capturing spatial feature details through deep convolution, processing feature data while preserving erosion features to the maximum extent; S23, introducing a cross-modal attention mechanism, dynamically allocating the weight ratio of each modal feature by calculating the mutual information entropy between different modal features, and strengthening the mining of correlation features between vegetation cover change and precipitation intensity fluctuation modality; S24, embedding a BatchNorm layer to standardize feature data during network training, and randomly shielding some neuron connections through a Dropout layer, filtering effective feature information according to a preset ratio, ensuring the orderly progress of the feature extraction process.

[0011] Further, S3 includes the following sub-steps: S31, according to the watershed topographic unit division rules, the watershed is divided into several equally sized spatiotemporal grid units, and based on the geographic coordinate information and time series nodes of each grid unit, the input parameters of soil organic matter content, root density, and vegetation cover are accurately matched and directionally allocated; S32, based on the spatiotemporal feature similarity and correlation analysis results of each grid unit, a dynamic spatiotemporal attention weight matrix is ​​constructed, and by calculating the contribution coefficient of different spatiotemporal units to soil retention capacity, the proportion of each unit feature in the mapping relationship is adjusted in real time; S33, a long short-term memory network is used to perform time series modeling of soil retention capacity-related parameters of each grid unit, and the feature evolution law in the long-term time dimension is selectively memorized through a gating mechanism to capture the dynamic changes of soil retention capacity with time factors.

[0012] Further, S4 includes the following sub-steps: S41, analyzing the topographic features of the watershed using digital elevation model data, dividing it into several independent hydrological response units, and establishing a one-to-one correspondence between hydrological response units and sediment transport paths based on the distribution of the watershed river network and the survey results of sediment transport paths; S42, using an adaptive threshold segmentation algorithm to classify the runoff calculation results into levels, dynamically determining multi-level classification thresholds based on the mean and standard deviation of runoff intensity, and classifying runoff intensity into different levels; S43, dynamically adjusting the calculation parameters of sediment particle initiation velocity and transport resistance coefficient for different runoff intensity levels, combined with the corresponding differences in soil erosion modulus and vegetation interception coefficient parameters; S44, constructing a coupling error correction model based on the least squares method, and establishing an error correction function by analyzing the deviation distribution law between the calculated and measured values ​​of runoff and sediment transport; S45, introducing watershed land use type distribution data, and performing secondary correction on the calibrated coupling results based on the differences in surface roughness and vegetation cover for different land use types.

[0013] Further, S5 includes the following sub-steps: S51, starting from the soil conservation function, soil holding capacity, erosion inhibition efficiency, vegetation soil conservation contribution rate, and runoff regulation effect are selected as evaluation indicators, including multiple dimensions such as soil properties, vegetation role, and runoff regulation; S52, a hierarchical structure model is constructed using the analytic hierarchy process (AHP), and field experts are invited to conduct pairwise comparisons and scoring of the importance of each evaluation indicator, constructing a judgment matrix and performing consistency checks to determine the preliminary weights of each indicator; S53, the information entropy value and difference coefficient of each indicator are calculated using the entropy weight method, and the preliminary weights are objectively corrected to obtain a comprehensive weight that takes into account both subjective experience and objective data; S54, based on domestic and international watershed soil conservation function evaluation standards and combined with the actual situation of the study area, a multi-level evaluation indicator system is constructed, clarifying the quantitative range and division boundaries of the indicators corresponding to each level.

[0014] A deep learning-based method for evaluating watershed soil conservation function is proposed. This method is implemented through a deep learning-based watershed soil conservation function evaluation system, comprising: a watershed multi-source parameter intelligent acquisition unit, used to collect topographic, vegetation cover, and precipitation intensity parameters through multispectral remote sensing, ground sensors, and UAV surveys, and establish a data transmission link with a dynamic intelligent monitoring and analysis platform for soil and water conservation; a multimodal erosion feature extraction unit, which receives data output from the acquisition unit, performs feature extraction using a multimodal watershed erosion perception deep learning model, and outputs feature vectors to subsequent units; and a spatiotemporal correlation consolidation prediction unit, based on... The system employs a feature vector model to construct a spatiotemporal mapping relationship, calculates soil holding capacity using a spatiotemporal correlation soil holding prediction model, and outputs prediction results. A runoff-sediment coupling calculation unit integrates prediction results with watershed parameters, uses a runoff-sediment coupling assessment algorithm to complete collaborative calculations, and outputs coupling results. A comprehensive evaluation system construction unit receives coupling results, integrates evaluation indicators to construct a system, and completes grade classification and quantitative assessment. An evaluation result output and visualization unit receives comprehensive evaluation results, presents them in chart and report formats, supports data export and real-time display, and connects all units via a high-speed data bus for real-time data transmission and collaborative work.

[0015] Beneficial Effects: This invention proposes a deep learning-based method for evaluating watershed soil conservation function. By integrating the advantages of multispectral remote sensing, ground sensor networks, and UAV surveys through a dynamic intelligent monitoring and analysis platform for soil and water conservation, it achieves comprehensive collection and efficient fusion of multi-source parameters such as topography, vegetation cover, and precipitation intensity. This solves the problems of insufficient data integration capabilities and incomplete spatiotemporal coverage in traditional technologies, providing timely and complete basic data support for evaluation. Through feature fusion and hierarchical perception mechanisms of a multimodal watershed erosion perception deep learning model, key features of watershed erosion are accurately extracted. Combined with the ability of a spatiotemporally correlated soil retention prediction model to capture spatiotemporal dynamic changes, and the collaborative computational advantages of a runoff-sediment coupling assessment algorithm, this method fully characterizes the complex correlations and spatiotemporal heterogeneity of various key processes within the watershed, overcoming the limitations of traditional technology evaluation logic and its lack of adaptability to real-world scenarios. Meanwhile, the comprehensive evaluation system, through the integration of multiple indicators and weight optimization, combined with hierarchical quantitative standards, has achieved accurate classification and quantitative assessment of soil conservation functions in different regions. Furthermore, the collaborative working mechanism of each unit in the system ensures the efficient advancement of the evaluation process. The final accurate evaluation report provides a scientific and reliable basis for watershed soil and water conservation planning, erosion control, and ecological restoration, significantly improving the accuracy, comprehensiveness, and application value of watershed soil conservation function evaluation. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0017] Figure 2This is a flowchart of method step S2 of the present invention; Figure 3 This is a flowchart of method step S3 of the present invention; Figure 4 This is a flowchart of method step S4 of the present invention; Figure 5 This is a flowchart of step S5 of the method of the present invention; Detailed Implementation It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a deep learning-based method for evaluating watershed soil conservation function includes the following steps: S1, collecting parameter data on watershed topography, vegetation cover, precipitation intensity, soil texture, land use type, slope and slope length through a dynamic intelligent monitoring and analysis platform for soil and water conservation, and constructing a multi-dimensional raw dataset; S2, inputting the dataset into a multimodal watershed erosion perception deep learning model, and extracting watershed erosion feature vectors through feature fusion and hierarchical perception mechanisms; S3, based on the extracted feature vectors, constructing a mapping relationship between watershed soil holding capacity and spatiotemporal variables using a spatiotemporal correlation soil retention prediction model; S4, using a runoff-sediment coupling assessment algorithm to collaboratively calculate watershed runoff intensity and sediment transport, generating coupled assessment results; S5, combining the model output and coupled assessment results, integrating indicators for watershed soil conservation function evaluation, and constructing a comprehensive evaluation system; S6, using this system to complete the classification and quantitative assessment of soil conservation function levels in different areas of the watershed, and outputting an evaluation report.

[0019] Step S1 involves collecting multi-dimensional core parameters through a dynamic intelligent monitoring and analysis platform for soil and water conservation. This platform integrates a multispectral remote sensing monitoring module, a ground sensor network, an unmanned aerial vehicle (UAV) survey unit, and a data fusion processing center. The multispectral remote sensing monitoring module uses a high-resolution multispectral camera with a spectral resolution of 400-900nm and a spatial resolution better than 5m to continuously capture spatial distribution information of watershed vegetation cover and topography. The ground sensor network deploys soil moisture sensors, precipitation sensors, and runoff sensors at a sampling frequency of 15 minutes per sampling, simultaneously collecting parameters such as soil texture and precipitation intensity, with data transmission latency controlled within 3 seconds. The UAV survey unit is equipped with a lidar and a thermal infrared imager, supplementing the collection of key data such as land use type, slope, and slope length with a survey coverage radius of no less than 10km and a data acquisition accuracy better than 0.1m. All collected data is transmitted in real time to a fusion processing center with a GPU cluster architecture. This center is equipped with eight high-performance computing graphics cards, supports multimodal data parallel processing, and has a storage capacity of no less than 100TB. Through data cleaning and correlation matching, a multi-dimensional raw dataset is constructed, including six major categories of parameters: topography, vegetation cover, precipitation intensity, soil texture, land use type, slope and slope length. This provides comprehensive, accurate, and timely data support for subsequent model calculations and evaluation analyses, ensuring the reliability and scientific validity of the calculation results in each subsequent step.

[0020] Step S2 inputs the constructed multi-dimensional raw dataset into the multimodal watershed erosion perception deep learning model. First, watershed vegetation cover, precipitation erosivity, soil erodibility, slope, and slope length are selected as core input modal parameters, with auxiliary parameters such as soil structure stability and vegetation root system soil-fixing strength embedded to construct a high-dimensional feature space including spatial distribution characteristics and temporal dynamic change information. The model uses a convolutional neural network to perform preliminary feature extraction on each modal data. Point convolution with deep separateable convolutions is used to extract inter-channel correlation information, and deep convolutions are used to capture spatial feature details, optimizing data processing efficiency while preserving key erosion features to the maximum extent. Subsequently, a cross-modal attention mechanism is introduced, dynamically allocating weight ratios by calculating the mutual information entropy between different modal features to strengthen the mining of key correlation features such as vegetation cover change and precipitation intensity fluctuations. Then, residual connections in a deep residual network are used to alleviate the gradient vanishing problem, and soil porosity and erosion resistance parameters are simultaneously incorporated to optimize network training. During network training, a BatchNorm layer is embedded for data standardization. A Dropout layer randomly masks some neuron connections at a preset ratio to filter effective features. Finally, a fully connected layer maps the extracted high-order feature vectors to the erosion intensity prediction space to complete model training optimization, ensuring the model's adaptability to different erosion scenarios and providing accurate feature inputs for subsequent soil retention capacity prediction.

[0021] Step S3, based on the key feature vectors extracted in Step S2, constructs a mapping relationship using a spatiotemporal correlation soil retention prediction model. First, according to the watershed topographic unit division rules, the watershed is divided into several equally sized spatiotemporal grid units. Based on the geographic coordinates and time series nodes of each grid unit, input parameters such as soil organic matter content, root density, soil clay content, and vegetation cover are precisely matched and directionally allocated. Then, a dynamic spatiotemporal attention weight matrix is ​​constructed. By calculating the contribution coefficient of different spatiotemporal units to soil retention capacity, the proportion of each unit's features in the mapping relationship is adjusted in real time. Simultaneously, a temporal convolutional network is used to extract temporal-dimensional correlation features, a graph convolutional network is used to capture spatial neighborhood correlation information, and a long short-term memory network gating mechanism is combined to selectively memorize the long-term temporal feature evolution patterns, accurately capturing the dynamic changes in soil retention capacity over time. The model introduces a spatiotemporal attention factor to establish a nonlinear mapping between features and soil holding capacity. It combines the topographic complexity and spatiotemporal distribution differences of precipitation in different regions of the watershed to perform regional calibration of the prediction results. Through multiple rounds of iterative optimization, the model's spatiotemporal sensitive parameters are adjusted, which effectively improves the spatiotemporal consistency and accuracy of soil holding capacity prediction under different watershed scenarios, and provides scientific basic data on soil holding capacity for subsequent runoff and sediment coupling calculations.

[0022] Step S4 employs a runoff-sediment coupling assessment algorithm for collaborative calculation. First, total precipitation, runoff coefficient, surface roughness, soil infiltration rate, and gradient are selected as core parameters for runoff calculation. The watershed topographic features are analyzed using digital elevation model data, and several independent hydrological response units are defined to construct a runoff intensity calculation model for distributed computation. Based on the runoff calculation results, a sediment transport calculation module is constructed by combining sediment particle size distribution, soil erosion modulus, and vegetation interception coefficient. An adaptive threshold segmentation algorithm is used to dynamically determine multi-level segmentation thresholds based on the mean and standard deviation of runoff intensity, classifying runoff intensity into different levels. For the parameter differences corresponding to each level of runoff intensity, calculation parameters such as sediment particle initiation velocity and transport resistance coefficient are dynamically adjusted to construct a runoff-sediment coupling relationship model. A quantitative correlation between the two is established through a coupling coefficient. Data on watershed topographic relief, vegetation buffer zone width, and land use type distribution are introduced to initially correct the coupling results. Subsequently, a coupling error correction model based on the least squares method was constructed. The deviation distribution law between the calculated and measured values ​​was analyzed to establish an error correction function. The preliminary correction results were then calibrated a second time. Through multiple rounds of iterative optimization of the coupling model parameters, the consistency between the calculated runoff intensity and sediment transport was ensured, and accurate coupling evaluation results were generated, providing core data support for the construction of a comprehensive evaluation system.

[0023] Step S5 combines the soil retention capacity prediction results from Step S3 with the runoff-sediment coupling assessment results from Step S4 to integrate the core indicators of watershed soil conservation function evaluation and construct a comprehensive evaluation system. First, starting from the core functions of soil conservation, four major categories of core indicators are selected: soil retention capacity, erosion inhibition efficiency, vegetation soil fixation contribution rate, and runoff regulation effect. These indicators include multiple key dimensions such as soil properties, vegetation role, and runoff regulation, ensuring the comprehensiveness of the evaluation system. Then, a hierarchical structure model is constructed using the analytic hierarchy process (AHP). Three to five senior experts in the field are invited to conduct pairwise comparisons and scoring of the importance of each evaluation indicator, constructing a judgment matrix and determining the preliminary weights of each indicator through a consistency test (test coefficient less than 0.1). Simultaneously, the information entropy value and difference coefficient of each indicator are calculated using the entropy weight method to objectively correct the preliminary weights, resulting in a comprehensive weight that balances subjective experience and objective data, avoiding the limitations of a single weight determination method. Finally, based on domestic and international watershed soil conservation function evaluation standards, and combined with the actual situation of the study area such as topography, climate conditions, and vegetation cover, a multi-level evaluation index system including five levels was constructed. The quantitative range and division boundary of the indicators corresponding to each level were clarified, and clear and operable evaluation rules were formulated to ensure the scientificity, objectivity and practicality of the evaluation system, and to provide a standardized basis for subsequent level division and quantitative assessment.

[0024] Step S6 utilizes the comprehensive evaluation system constructed in Step S5 to conduct a classification and quantitative assessment of soil conservation function levels in different areas of the watershed and output an accurate evaluation report. First, the prediction results from Step S3 and the coupled evaluation results from Step S4 are quantitatively converted according to the evaluation system requirements. Based on the comprehensive weight and quantitative value of each indicator, a weighted summation is used to calculate the comprehensive score of soil conservation function for each evaluation unit. According to the five-level classification criteria set in the evaluation system, the comprehensive score is mapped to the corresponding functional level, with Level 1 being the optimal level and Level 5 the lowest level, clearly defining the current status of soil conservation function in different areas. During the assessment process, the spatiotemporal heterogeneity of different areas of the watershed is fully considered, taking into account factors such as topographic complexity, differences in precipitation distribution, and land use types. The classification results are validated for rationality and locally adjusted to ensure a high degree of consistency between the classification and the actual situation. The final accurate evaluation report includes core contents such as the overall soil conservation function status analysis of the watershed, distribution maps of different regional levels, identification of core influencing factors at each level, and assessment of the potential for functional improvement. The report is presented in the form of charts and text, supports data export and real-time display, and provides scientific and accurate decision-making basis for the formulation of watershed soil and water conservation plans, the layout of erosion control projects, and the optimization of ecological restoration schemes, helping to improve the pertinence and effectiveness of watershed soil and water conservation work.

[0025] Preferably, the modeling process of the multimodal watershed erosion perception deep learning model in S2 includes: selecting watershed vegetation cover, precipitation erosivity, soil erodibility, slope, and slope length as model input modal parameters; constructing a multi-source data input matrix; and performing preliminary feature extraction on each modal data using a convolutional neural network to obtain a single-modal feature matrix. ,in This is the preliminary characteristic matrix of the m-th mode. This is the raw data for the m-th mode. The convolution kernel weight matrix is... Here is the modal bias matrix. To output the bias term, an attention mechanism is used to calculate the weight coefficients of each modality feature, and the multimodal fusion feature is obtained through weighted fusion. ,in These are the weighting coefficients for the m-th mode. Here, n is the attention weight matrix, and n is the total number of modes. A deep residual network is constructed to extract the fusion features in an advanced manner. The gradient vanishing problem is alleviated by residual connections, and the propagation and retention of erosion features are enhanced. Simultaneously, parameters such as watershed soil porosity and erosion resistance age are introduced to optimize the network training process. Based on the extracted high-order feature vectors, they are mapped to the erosion intensity prediction space through fully connected layers to complete the model training and optimization, ensuring the model's adaptability and perception accuracy for different erosion scenarios in the watershed.

[0026] Specifically, the multimodal watershed erosion perception deep learning model selects watershed vegetation cover, precipitation erosivity, soil erodibility, slope, and slope length as core input modal parameters. These parameters cover key dimensions of topography, vegetation, and soil. A multi-source data input matrix is ​​constructed, and a convolutional neural network is used for initial feature extraction. The derivation of its single-modal feature matrix formula is based on the essence of convolution operations, performing a tensor product operation between the original data and the convolution kernel weight matrix. The modal bias matrix and output bias term are then superimposed to achieve the mapping from original data to features. The convolution kernel weight matrix is ​​randomly initialized and optimized through training. The initial values ​​of the modal bias matrix and output bias term are set to 0.01 and adjusted iteratively through backpropagation. An attention mechanism is used to calculate weight coefficients. The weighted fusion formula incorporates the Softmax normalization result of the attention weight matrix to dynamically allocate the importance of each modal feature. The weight coefficients are randomly generated through a uniform distribution. The total number of modalities is determined to be 5-8 categories based on the input parameter types. This formula aims to highlight the contribution of key modalities and suppress redundant information. A deep residual network was constructed to mitigate gradient vanishing by passing features from previous layers through residual connections. Soil porosity (0.3-0.6) and erosion resistance duration (5-30 years) parameters were introduced to optimize training. The network learning rate was initially set to 0.001 and adjusted exponentially. High-order feature vectors were mapped to the erosion intensity prediction space through fully connected layers. The number of neurons in the fully connected layers was 2-3 times the dimension of the input features. Parameters were optimized through 500-1000 iterations of training to ensure the model adapts to different watershed erosion scenarios, achieving efficient extraction and accurate perception of key features, and providing high-quality feature input for subsequent soil retention capacity prediction.

[0027] Preferably, the modeling process of the spatiotemporal correlation soil retention prediction model in S3 includes: selecting spatiotemporal sequence data of watershed soil organic matter content, root density, soil clay content, and vegetation cover as input parameters; constructing a spatiotemporal data tensor; extracting temporal dimension correlation features through a temporal convolutional network; and simultaneously capturing spatial neighborhood correlation information using a graph convolutional network; constructing a soil retention capacity prediction model based on spatiotemporal features; and introducing a spatiotemporal attention factor to establish a nonlinear mapping between features and soil retention capacity, expressed as: ,in This is a predicted value for soil sequestration. For spatial characteristic matrix, For graph convolution weights, For temporal convolution weights, This is the spatiotemporal adjustment coefficient. This is a spatiotemporal correlation function; combined with the topographic complexity and spatiotemporal distribution differences of precipitation in different regions of the watershed, the model prediction results are regionally calibrated, and the spatiotemporal sensitive parameters of the model are adjusted through iterative optimization to improve the spatiotemporal consistency and accuracy of soil retention capacity prediction under different watershed scenarios.

[0028] Specifically, the spatiotemporal correlation soil retention prediction model selects spatiotemporal sequence data of soil organic matter content (10-50), root density (0.5-3.0), soil clay content (10-40), and vegetation cover (0.2-0.9) as input parameters. These parameters directly determine the spatiotemporal changes in soil retention capacity. By constructing a spatiotemporal data tensor, a temporal convolutional network (kernel size 3-5) is used to extract temporal trend correlation features, and a graph convolutional network (neighborhood radius 2-3 times that of watershed grid cells) is used to capture spatial neighborhood correlations, achieving spatiotemporal feature separation and extraction. Based on the extracted spatiotemporal features, a prediction model is constructed. The core formula derivation integrates the output results of the temporal convolutional network and the graph convolutional network. The spatiotemporal correlation function calculation results are integrated through a spatiotemporal adjustment coefficient (initially 0.7, dynamically adjusted according to training error). This function is determined by fitting the nonlinear relationship between soil retention capacity and spatiotemporal variables. The graph convolution weights and temporal convolution weights are randomly initialized and optimized through backpropagation. The formula construction aims to explore the synergistic influence of spatiotemporal features on soil retention capacity and achieve accurate mapping. By combining the complexity of the watershed topography (graded from 1 to 5 according to slope) and the differences in the spatiotemporal distribution of precipitation (graded from 0.3 to 1.2 according to the precipitation variation coefficient), the prediction results are regionally calibrated. After 300-500 rounds of iterative optimization, the spatiotemporal sensitivity parameters are adjusted (the initial time sensitivity coefficient is 0.6, the initial spatial sensitivity coefficient is 0.4, and it is adjusted according to regional errors) to improve the adaptability of the model under different watershed scenarios, ensure the spatiotemporal continuity and consistency of the soil retention capacity prediction results, and provide accurate data support for the evaluation of watershed soil conservation function.

[0029] Preferably, the calculation process of the runoff-sediment coupling assessment algorithm in S4 includes: selecting total watershed precipitation, runoff coefficient, surface roughness, soil infiltration rate, and gradient as runoff calculation parameters; constructing a runoff intensity calculation model; and performing distributed calculation of watershed runoff through hydrological response unit division; based on the runoff calculation results, constructing a sediment transport calculation module by combining sediment particle size distribution, soil erosion modulus, and vegetation interception coefficient, with the runoff intensity calculation expression being: ,in For runoff intensity, For the precipitation data of time period i, Let be the runoff coefficient for the i-th time period. Let be the length of the confluence during the i-th time period. Let j be the area of ​​the j-th hydrological unit. Let be the infiltration rate of the j-th unit. The runoff correction coefficient is used; a runoff-sediment coupling model is constructed, and a quantitative correlation between runoff intensity and sediment transport is established through the coupling coefficient. The coupling expression is as follows: ,in This refers to sediment transport. The coupling coefficient is... The content of sediment particles in Class I sediment is... The coefficient for particle transport of type I is used. The parameters of watershed topographic relief and vegetation buffer zone width are introduced to correct the coupling results. The influence of different land use types on runoff and sediment transport is considered. The coupling model parameters are optimized through iterative calculation to ensure the consistency between the calculation results of runoff intensity and sediment transport, and to provide accurate basic data on runoff and sediment for watershed soil conservation function evaluation.

[0030] Specifically, the runoff and sediment coupling assessment algorithm selects total precipitation (500-2000), runoff coefficient (0.1-0.6), surface roughness (0.01-0.1), soil infiltration rate (5-30), and slope (0.001-0.1) as the core parameters for runoff calculation. It analyzes the terrain features through a digital elevation model, divides the area into hydrological response units of 1-5 square kilometers, and constructs a model to achieve distributed computing, comprehensively covering the key influencing factors of runoff formation. Based on runoff calculations, a sediment transport calculation module was constructed, incorporating sediment particle size distribution (3-5 grades), soil erosion modulus (500-5000), and vegetation interception coefficient (0.2-0.8). The runoff intensity calculation formula was derived by integrating the ratios of precipitation, runoff coefficient, runoff length to hydrological unit area, and infiltration rate for each time period, multiplied by an initial 0.85 and calibrated according to measured data. The formula construction considered the comprehensive influence of the spatiotemporal distribution of precipitation and underlying surface conditions, and the parameter values ​​were determined statistically through watershed measured data. A coupling relationship model was constructed, with the coupling expression derived by summing the products of runoff intensity and coupling coefficient (initial 0.5, adjusted according to runoff-sediment correlation), and the contents of various sediment particles and their corresponding transport coefficients (0.3-0.9), quantifying the nonlinear correlation between the two. Step R4 incorporates parameters such as topographic relief (graded according to elevation difference, values ​​1-4) and vegetation buffer zone width (10-50) to correct the coupling results. It also considers the influence of land use type (assigned values ​​1-6) and adjusts the land use correction coefficient by 0.6-1.3 to enhance relevance. After 400-600 iterations to optimize the coupling model parameters, the mean square error is used as the loss function (target loss value less than 0.01) to ensure consistency between runoff intensity and sediment transport calculations, providing accurate and reliable basic runoff and sediment data for watershed soil conservation function evaluation.

[0031] Preferably, the dynamic intelligent monitoring and analysis platform for soil and water conservation includes a multispectral remote sensing monitoring module, a ground sensor network, an unmanned aerial vehicle (UAV) survey unit, and a data fusion processing center. The multispectral remote sensing monitoring module uses a high-resolution multispectral camera with a spectral resolution range of 400-900nm and a spatial resolution better than 5m. The ground sensor network deploys soil moisture sensors, precipitation sensors, and runoff sensors, with a sampling frequency of 15 minutes / time and a data transmission delay of no more than 3 seconds. The UAV survey unit is equipped with a lidar and a thermal infrared imager, with a survey coverage radius of no less than 10km and a data acquisition accuracy better than 0.1m. The data fusion processing center adopts a GPU cluster architecture, including 8 high-performance computing graphics cards, supporting parallel processing of multimodal data, with a data storage capacity of no less than 100TB, and a computational delay of no more than 10 seconds for model training and inference.

[0032] Preferred, such as Figure 2 As shown, S2 includes the following sub-steps: S21, for multimodal input data of watershed vegetation cover, precipitation erosivity, and soil erodibility, embed parameters of watershed soil structure stability and vegetation root soil-fixing strength to construct a high-dimensional feature space including spatial distribution characteristics and temporal dynamic change information, and organically integrate multi-dimensional parameters; S22, use deep separateable convolution to perform layered processing on the expanded high-dimensional features, extract inter-channel correlation information through point convolution, and capture spatial feature details through deep convolution, processing feature data while preserving erosion features to the maximum extent; S23, introduce a cross-modal attention mechanism, dynamically allocate the weight ratio of each modal feature by calculating the mutual information entropy between different modal features, and strengthen the mining of correlation features between vegetation cover change and precipitation intensity fluctuation modality; S24, embed a BatchNorm layer to standardize feature data during network training, and randomly shield some neuron connections through a Dropout layer to filter effective feature information according to a preset ratio, ensuring the orderly progress of the feature extraction process.

[0033] Specifically, the feature extraction process in step S2 achieves efficient feature mining and optimization of multimodal data through four sub-steps. S21 expands the multimodal input data, including watershed vegetation cover, precipitation erosivity, and soil erodibility, embedding watershed soil structure stability parameters ranging from 0.6 to 0.9 and vegetation root soil-fixing strength parameters ranging from 0.3 to 0.8. This constructs a high-dimensional feature space that simultaneously includes spatial distribution characteristics and temporal dynamic change information, achieving the organic integration of multi-dimensional parameters such as topography, vegetation, and soil, laying a comprehensive data foundation for subsequent feature extraction. S22 uses deep separateable convolution to perform layered processing on the expanded high-dimensional features. Point convolution extracts the correlation information between different parameter channels, while deep convolution accurately captures spatial feature details. While preserving key erosion features to the maximum extent, the data dimension is compressed to 30%-50% of the original dimension, effectively reducing computational load. S23 introduces a cross-modal attention mechanism, which dynamically allocates the weight ratio of each modality feature by calculating the mutual information entropy between different modal features. The weight coefficient ranges from 0.1 to 0.9, strengthening the mining of correlation features between key modalities such as vegetation cover change and precipitation intensity fluctuations, and improving the targeting of features. S24 embeds a BatchNorm layer during network training to standardize the feature data, adjusting the mean to 0 and the variance to 1. At the same time, a Dropout layer randomly masks some neuron connections at a ratio of 20%-30%, filtering effective feature information, avoiding model overfitting, and ensuring the orderly progress and stability of the feature extraction process.

[0034] Preferred, such as Figure 3 As shown, S3 includes the following sub-steps: S31, according to the watershed topographic unit division rules, the watershed is divided into several equally sized spatiotemporal grid units. Based on the geographic coordinate information and time series nodes of each grid unit, the input parameters of soil organic matter content, root density, and vegetation cover are accurately matched and directionally allocated; S32, based on the spatiotemporal feature similarity and correlation analysis results of each grid unit, a dynamic spatiotemporal attention weight matrix is ​​constructed. By calculating the contribution coefficient of different spatiotemporal units to soil retention capacity, the proportion of each unit feature in the mapping relationship is adjusted in real time; S33, a long short-term memory network is used to perform time series modeling of soil retention capacity-related parameters of each grid unit. Through a gating mechanism, the feature evolution law in the long-term time dimension is selectively memorized to capture the dynamic changes of soil retention capacity with time factors.

[0035] Specifically, the mapping relationship construction process in step S3 achieves precise correlation between soil retention capacity and spatiotemporal variables through three sub-steps. S31 divides the watershed into equal-sized spatiotemporal grid units with sides ranging from 100 to 500 meters, according to the watershed topographic unit division rules. Based on the latitude and longitude coordinates of each grid unit and time series nodes on a daily basis, input parameters such as soil organic matter content (10-50), root density (0.5-3.0), soil clay content (10-40), and vegetation cover (0.2-0.9) are precisely matched and directionally allocated, ensuring a one-to-one correspondence between parameters and spatiotemporal units. S32, based on the spatiotemporal feature similarity and correlation analysis results of each grid unit, uses the Pearson correlation coefficient to calculate the correlation strength, constructing a dynamic spatiotemporal attention weight matrix. The weight values ​​are dynamically adjusted between 0 and 1. By calculating the contribution coefficient of different spatiotemporal units to soil retention capacity, the proportion of each unit's features in the mapping relationship is adjusted in real time, highlighting the influence of key spatiotemporal units. S33 employs a long short-term memory network to model the time series parameters related to soil retention capacity in each grid unit. The number of neurons in the hidden layer of the network is set to 64-128. Through the gating mechanism of input gate, forget gate and output gate, it selectively memorizes the feature evolution patterns in the past 30-90 days, accurately captures the dynamic changes of soil retention capacity with time factors such as seasonal changes and precipitation fluctuations, and provides technical support for building a stable and reliable mapping relationship.

[0036] Preferred, such as Figure 4 As shown, S4 includes the following sub-steps: S41, analyzing the topographic features of the watershed using digital elevation model data, dividing it into several independent hydrological response units, and establishing a one-to-one correspondence between hydrological response units and sediment transport paths based on the distribution of the watershed river network and the survey results of sediment transport paths; S42, using an adaptive threshold segmentation algorithm to classify the runoff calculation results into levels, dynamically determining multi-level classification thresholds based on the mean and standard deviation of runoff intensity, and classifying runoff intensity into different levels; S43, dynamically adjusting the calculation parameters of sediment particle initiation velocity and transport resistance coefficient for different runoff intensity levels, combined with the corresponding differences in soil erosion modulus and vegetation interception coefficient parameters; S44, constructing a coupling error correction model based on the least squares method, and establishing an error correction function by analyzing the deviation distribution law between the calculated and measured values ​​of runoff and sediment transport; S45, introducing watershed land use type distribution data, and performing secondary correction on the calibrated coupling results based on the differences in surface roughness and vegetation cover for different land use types.

[0037] Specifically, the collaborative calculation process in step S4 achieves precise coupling between runoff intensity and sediment transport through five sub-steps. S41 analyzes the watershed topographic features using digital elevation model data with a resolution of 30 meters, dividing the area into several independent hydrological response units ranging from 1 to 5 square kilometers. Combining the watershed river network distribution vector data with field survey results of sediment transport paths, a one-to-one correspondence between hydrological response units and sediment transport paths is established, clarifying the boundary between runoff and sediment transport correlation. S42 employs an adaptive threshold segmentation algorithm to classify the runoff calculation results into levels. Based on the mean and standard deviation of runoff intensity, a 3-5 level classification threshold is dynamically determined, dividing runoff intensity into different levels such as weak, moderate, and strong, providing a grading basis for sediment transport calculations. S43 For different runoff intensity levels, combined with the corresponding soil erosion modulus of 500-5000 and vegetation interception coefficient of 0.2-0.8, calculation parameters such as sediment particle initiation velocity (0.1-0.5) and transport resistance coefficient (0.3-0.9) are dynamically adjusted to achieve accurate tiered calculations. S44 Constructs a coupling error correction model based on the least squares method. By analyzing the deviation distribution patterns between calculated and measured runoff and sediment transport values ​​over the past 1-3 years, an error correction function is established, with correction coefficients ranging from 0.8 to 1.2. S45 Incorporating watershed land use type distribution data, and considering the differences in surface roughness (0.01-0.1) and vegetation cover (0.2-0.9) among different land use types such as cultivated land, forest land, and grassland, a secondary correction is performed on the calibrated coupling results to ensure a high degree of consistency between the collaborative calculation results and the actual situation.

[0038] Preferred, such as Figure 5 As shown, S5 includes the following sub-steps: S51, starting from the soil conservation function, soil holding capacity, erosion inhibition efficiency, vegetation soil conservation contribution rate, and runoff regulation effect are selected as evaluation indicators, including multiple dimensions such as soil properties, vegetation role, and runoff regulation; S52, a hierarchical structure model is constructed using the analytic hierarchy process (AHP), and field experts are invited to conduct pairwise comparisons and scoring of the importance of each evaluation indicator, construct a judgment matrix, and conduct consistency checks to determine the preliminary weights of each indicator; S53, the information entropy value and difference coefficient of each indicator are calculated using the entropy weight method, and the preliminary weights are objectively corrected to obtain a comprehensive weight that takes into account both subjective experience and objective data; S54, based on domestic and international watershed soil conservation function evaluation standards and combined with the actual situation of the study area, a multi-level evaluation indicator system is constructed, and the quantitative range and division boundary of the indicators corresponding to each level are clarified.

[0039] Specifically, the comprehensive evaluation system construction process in step S5 is achieved through four sub-steps to scientifically improve the evaluation system. S51 starts with the core function of soil conservation, selecting soil holding capacity (0-5000), erosion inhibition efficiency (0-1), vegetation soil fixation contribution rate (0-0.8), and runoff regulation effect (0-0.7) as core evaluation indicators, including multiple dimensions such as soil properties, vegetation role, and runoff regulation, to ensure the comprehensiveness of the evaluation. S52 uses the analytic hierarchy process (AHP) to construct a hierarchical structure model, inviting 3-5 senior experts in watershed ecology to conduct pairwise comparisons and scores of the importance of each evaluation indicator, constructing a judgment matrix and performing a consistency test, with the test coefficient controlled within 0.1, determining the preliminary weights of each indicator, with a total weight of 1. S53 uses the entropy weight method to calculate the information entropy value and difference coefficient of each indicator. The information entropy value ranges from 0 to 1. The larger the difference coefficient, the higher the indicator's distinguishability. The initial weights are objectively corrected according to the entropy weight ratio of 0.3-0.5 to obtain a comprehensive weight that takes into account both subjective experience and objective data, thus improving the rationality of the weights. S54 Based on domestic and international watershed soil conservation function evaluation standards such as the "Soil Conservation Function Evaluation Specification," and combined with the actual conditions of the study area such as topography, climate conditions, and vegetation cover, a multi-level evaluation index system including five levels is constructed. The quantitative range and boundary of the indicators corresponding to each level are clearly defined, with level one being the optimal level and level five being the lowest. Clear and operable evaluation rules are formulated to ensure the objectivity and practicality of the evaluation system.

[0040] A deep learning-based method for evaluating watershed soil conservation function is proposed. This method is implemented through a deep learning-based watershed soil conservation function evaluation system, comprising: a watershed multi-source parameter intelligent acquisition unit, used to collect topographic, vegetation cover, and precipitation intensity parameters through multispectral remote sensing, ground sensors, and UAV surveys, and establish a data transmission link with a dynamic intelligent monitoring and analysis platform for soil and water conservation; a multimodal erosion feature extraction unit, which receives data output from the acquisition unit, performs feature extraction using a multimodal watershed erosion perception deep learning model, and outputs feature vectors to subsequent units; and a spatiotemporal correlation consolidation prediction unit, based on... The system employs a feature vector model to construct a spatiotemporal mapping relationship, calculates soil holding capacity using a spatiotemporal correlation soil holding prediction model, and outputs prediction results. A runoff-sediment coupling calculation unit integrates prediction results with watershed parameters, uses a runoff-sediment coupling assessment algorithm to complete collaborative calculations, and outputs coupling results. A comprehensive evaluation system construction unit receives coupling results, integrates evaluation indicators to construct a system, and completes grade classification and quantitative assessment. An evaluation result output and visualization unit receives comprehensive evaluation results, presents them in chart and report formats, supports data export and real-time display, and connects all units via a high-speed data bus for real-time data transmission and collaborative work.

[0041] A deep learning-based method for evaluating watershed soil conservation function integrates a dynamic intelligent monitoring and analysis platform for soil and water conservation using multispectral remote sensing, ground sensor networks, and UAV surveys. This platform achieves comprehensive collection and efficient fusion of multi-dimensional parameters such as watershed topography, vegetation, and precipitation, significantly improving the spatiotemporal coverage and accuracy of the data. Through the hierarchical perception and feature fusion capabilities of a multimodal watershed erosion perception deep learning model, the complex correlations between different parameters are accurately uncovered, capturing key characteristics of watershed erosion. A spatiotemporal correlation soil holding capacity prediction model and a runoff-sediment coupling assessment algorithm respectively achieve accurate prediction of the spatiotemporal dynamic changes in soil holding capacity and collaborative calculation of runoff and sediment transport processes, providing core scientific data support for the evaluation. The comprehensive evaluation system, through multi-index integration and weight optimization, ensures the comprehensiveness and objectivity of the evaluation results.

[0042] This method addresses the shortcomings of traditional technologies in data integration and spatiotemporal coverage. By collaboratively acquiring data from multiple monitoring devices and fusing the data, it achieves complete acquisition of parameters across multiple dimensions and spatiotemporal scales, improving the timeliness and completeness of the data. Furthermore, it addresses the lack of adaptability between traditional evaluation logic and real-world scenarios. Through deep learning models for extracting complex correlation features, constructing dynamic spatiotemporal mapping relationships, and coupling calculations of runoff and sediment processes, it fully characterizes the spatiotemporal heterogeneity and complex correlations of key processes within the watershed. This avoids the limitations of single static parameters and linear calculations. Combined with regional calibration and parameter optimization mechanisms, it significantly improves the adaptability of evaluation results to real-world scenarios. The resulting accurate evaluation report provides a reliable basis for decision-making in watershed soil and water conservation.

[0043] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for evaluating watershed soil conservation function, characterized in that, Includes the following steps: S1. Collect parameter data on watershed topography, vegetation cover, precipitation intensity, soil texture, land use type, slope and slope length through the dynamic intelligent monitoring and analysis platform for soil and water conservation, and construct a multi-dimensional raw dataset; S2. Input the dataset into a multimodal watershed erosion perception deep learning model, and extract watershed erosion feature vectors through feature fusion and hierarchical perception mechanisms; S3. Based on the extracted feature vectors, construct the mapping relationship between watershed soil holding capacity and spatiotemporal variables using a spatiotemporal correlation soil holding prediction model; S4. Use a runoff-sediment coupling assessment algorithm to collaboratively calculate watershed runoff intensity and sediment transport, and generate coupling assessment results; S5. Combine the model output and coupling assessment results, integrate the indicators for watershed soil conservation function evaluation, and construct a comprehensive evaluation system; S6. Use this system to complete the classification and quantitative assessment of soil conservation function levels in different areas of the watershed, and output an evaluation report; The modeling process of the multimodal watershed erosion perception deep learning model in S2 includes: selecting watershed vegetation cover, precipitation erosivity, soil erodibility, slope, and slope length as model input modal parameters; constructing a multi-source data input matrix; and using a convolutional neural network to perform preliminary feature extraction on each modal data to obtain a single-modal feature matrix. ,in This is the preliminary characteristic matrix of the m-th mode. This is the raw data for the m-th mode. The convolution kernel weight matrix is... Here is the modal bias matrix. To output the bias term, an attention mechanism is used to calculate the weight coefficients of each modality feature, and the multimodal fusion feature is obtained through weighted fusion. ,in These are the weighting coefficients for the m-th mode. Here, n is the attention weight matrix, and n is the total number of modes. A deep residual network is constructed to extract the fused features in an advanced manner. The gradient vanishing problem is alleviated by residual connections, and the propagation and retention of erosion features are enhanced. Simultaneously, parameters such as watershed soil porosity and erosion resistance age are introduced to optimize the network training process. Based on the extracted high-order feature vectors, they are mapped to the erosion intensity prediction space through fully connected layers to complete the model training and optimization, ensuring the model's adaptability and perception accuracy for different erosion scenarios in the watershed. The modeling process of the spatiotemporal correlation soil retention prediction model in S3 includes: selecting spatiotemporal sequence data of watershed soil organic matter content, root density, soil clay content, and vegetation cover as input parameters; constructing a spatiotemporal data tensor; extracting temporal correlation features through a temporal convolutional network; and simultaneously capturing spatial neighborhood correlation information using a graph convolutional network. Based on the spatiotemporal features, a soil retention capacity prediction model is constructed, and a spatiotemporal attention factor is introduced to establish a nonlinear mapping between features and soil retention capacity, expressed as follows: ,in This is a predicted value for soil sequestration. For spatial characteristic matrix, For graph convolution weights, For temporal convolution weights, This is the spatiotemporal adjustment coefficient. This is a spatiotemporal correlation function; combined with the topographic complexity and spatiotemporal distribution differences of precipitation in different regions of the watershed, the model prediction results are regionally calibrated, and the spatiotemporal sensitive parameters of the model are adjusted through iterative optimization to improve the spatiotemporal consistency and accuracy of soil retention capacity prediction under different watershed scenarios.

2. The watershed soil conservation function evaluation method based on deep learning according to claim 1, characterized in that, The calculation process of the runoff-sediment coupling assessment algorithm in S4 includes: selecting total watershed precipitation, runoff coefficient, surface roughness, soil infiltration rate, and gradient as runoff calculation parameters; constructing a runoff intensity calculation model; and performing distributed calculation of watershed runoff through hydrological response unit division; based on the runoff calculation results, and combining sediment particle size distribution, soil erosion modulus, and vegetation interception coefficient, constructing a sediment transport calculation module, whose runoff intensity calculation expression is as follows: ,in For runoff intensity, For the precipitation data of time period i, Let be the runoff coefficient for the i-th time period. Let be the length of the confluence during the i-th time period. Let j be the area of ​​the j-th hydrological unit. Let be the infiltration rate of the j-th unit. The runoff correction coefficient is used; a runoff-sediment coupling model is constructed, and a quantitative correlation between runoff intensity and sediment transport is established through the coupling coefficient. The coupling expression is as follows: ,in This refers to sediment transport. The coupling coefficient is... The content of sediment particles in Class I sediment is... The coefficient for particle transport of type I is used. The parameters of watershed topographic relief and vegetation buffer zone width are introduced to correct the coupling results. The influence of different land use types on runoff and sediment transport is considered. The coupling model parameters are optimized through iterative calculation to ensure the consistency between the calculation results of runoff intensity and sediment transport, and to provide accurate basic data on runoff and sediment for watershed soil conservation function evaluation.

3. The watershed soil conservation function evaluation method based on deep learning according to claim 1, characterized in that, The dynamic intelligent monitoring and analysis platform for soil and water conservation includes a multispectral remote sensing monitoring module, a ground sensor network, an unmanned aerial vehicle (UAV) survey unit, and a data fusion processing center. The multispectral remote sensing monitoring module uses a high-resolution multispectral camera with a spectral resolution range of 400-900nm and a spatial resolution better than 5m. The ground sensor network deploys soil moisture sensors, precipitation sensors, and runoff sensors, with a sampling frequency of 15 minutes / time and a data transmission latency of no more than 3 seconds. The UAV survey unit is equipped with a lidar and a thermal infrared imager, with a survey coverage radius of no less than 10km and a data acquisition accuracy better than 0.1m. The data fusion processing center adopts a GPU cluster architecture, including 8 high-performance computing graphics cards, supporting parallel processing of multimodal data, with a data storage capacity of no less than 100TB and a computational latency of no more than 10 seconds for model training and inference.

4. The watershed soil conservation function evaluation method based on deep learning according to claim 1, characterized in that, S2 includes the following sub-steps: S21, for multimodal input data of watershed vegetation cover, precipitation erosivity, and soil erodibility, embed parameters of watershed soil structure stability and vegetation root soil-fixing strength to construct a high-dimensional feature space including spatial distribution characteristics and temporal dynamic change information, and organically integrate multi-dimensional parameters; S22, use deep separateable convolution to perform layered processing on the expanded high-dimensional features, extract inter-channel correlation information through point convolution, and capture spatial feature details through deep convolution, processing feature data while preserving erosion features to the maximum extent; S23, introduce a cross-modal attention mechanism, dynamically allocate the weight ratio of each modal feature by calculating the mutual information entropy between different modal features, and strengthen the mining of correlation features between vegetation cover change and precipitation intensity fluctuation modality; S24, embed a BatchNorm layer to standardize feature data during network training, and randomly shield some neuron connections through a Dropout layer to filter effective feature information according to a preset ratio, ensuring the orderly progress of the feature extraction process.

5. The watershed soil conservation function evaluation method based on deep learning according to claim 1, characterized in that, S3 includes the following sub-steps: S31, according to the watershed topographic unit division rules, the watershed is divided into several equally sized spatiotemporal grid units. Based on the geographic coordinate information and time series nodes of each grid unit, the input parameters of soil organic matter content, root density, and vegetation cover are accurately matched and directionally allocated; S32, based on the spatiotemporal feature similarity and correlation analysis results of each grid unit, a dynamic spatiotemporal attention weight matrix is ​​constructed. By calculating the contribution coefficient of different spatiotemporal units to soil retention capacity, the proportion of each unit feature in the mapping relationship is adjusted in real time; S33, a long short-term memory network is used to perform time series modeling of soil retention capacity-related parameters of each grid unit. Through a gating mechanism, the feature evolution law in the long-term time dimension is selectively memorized to capture the dynamic changes of soil retention capacity with time factors.

6. The watershed soil conservation function evaluation method based on deep learning according to claim 1, characterized in that, S4 includes the following steps: S41, analyzing the topographic features of the watershed using digital elevation model data, dividing it into several independent hydrological response units, and establishing a one-to-one correspondence between hydrological response units and sediment transport paths based on the distribution of the watershed river network and the survey results of sediment transport paths; S42, using an adaptive threshold segmentation algorithm to classify the runoff calculation results into levels, dynamically determining multi-level classification thresholds based on the mean and standard deviation of runoff intensity, and classifying runoff intensity into different levels; S43, dynamically adjusting the calculation parameters of sediment particle initiation velocity and transport resistance coefficient for different runoff intensity levels, combined with the corresponding differences in soil erosion modulus and vegetation interception coefficient parameters; S44, constructing a coupling error correction model based on the least squares method, and establishing an error correction function by analyzing the deviation distribution law between the calculated and measured values ​​of runoff and sediment transport; S45, introducing watershed land use type distribution data, and performing secondary correction on the calibrated coupling results based on the differences in surface roughness and vegetation cover for different land use types.

7. The watershed soil conservation function evaluation method based on deep learning according to claim 1, characterized in that, S5 includes the following sub-steps: S51, starting from soil conservation function, soil holding capacity, erosion inhibition efficiency, vegetation soil conservation contribution rate, and runoff regulation effect are selected as evaluation indicators, including multiple dimensions such as soil properties, vegetation role, and runoff regulation; S52, a hierarchical structure model is constructed using the analytic hierarchy process (AHP), and field experts are invited to conduct pairwise comparisons and scoring of the importance of each evaluation indicator, construct a judgment matrix, and conduct consistency checks to determine the preliminary weights of each indicator; S53, the information entropy value and difference coefficient of each indicator are calculated using the entropy weight method, and the preliminary weights are objectively corrected to obtain a comprehensive weight that takes into account both subjective experience and objective data; S54, based on domestic and international watershed soil conservation function evaluation standards and combined with the actual situation of the study area, a multi-level evaluation indicator system is constructed, and the quantitative range and division boundary of the indicators corresponding to each level are clarified.

8. A watershed soil conservation function evaluation method based on deep learning according to any one of claims 1-7, characterized in that, This method is implemented through a deep learning-based watershed soil conservation function evaluation system, comprising: a watershed multi-source parameter intelligent acquisition unit, used to collect topographic, vegetation cover, and precipitation intensity parameters through multispectral remote sensing, ground sensors, and UAV surveys, and establish a data transmission link with the dynamic intelligent monitoring and analysis platform for soil and water conservation; a multimodal erosion feature extraction unit, which receives the data output from the acquisition unit, completes feature extraction through a multimodal watershed erosion perception deep learning model, and outputs feature vectors to subsequent units; a spatiotemporal correlation soil retention prediction unit, which constructs a spatiotemporal mapping relationship based on feature vectors, calculates soil retention capacity through a spatiotemporal correlation soil retention prediction model, and outputs prediction results; a runoff and sediment coupling calculation unit, which integrates prediction results with watershed parameters, completes collaborative calculation using a runoff and sediment coupling evaluation algorithm, and outputs coupling results; a comprehensive evaluation system construction unit, which receives coupling results, integrates evaluation indicators to construct a system, and completes grade classification and quantitative evaluation; and an evaluation result output and visualization unit, which receives comprehensive evaluation results, presents them in the form of charts and reports, supports data export and real-time display, and all units are connected through a high-speed data bus for real-time data transmission and collaborative work.

Images