A method for ecological restoration of steep rock slopes with large angles

Abstract

The present application relates to the field of ecological restoration technology, and in particular to a novel method for ecological restoration of steep rock slopes with large angles. The method first obtains high-temporal and spatial resolution monitoring data of the slope ecosystem through multi-source data fusion such as unmanned aerial vehicle remote sensing, ground three-dimensional laser scanning, and Internet of Things monitoring; then, an end-to-end deep learning model integrating spatial and temporal features is constructed to achieve adaptive learning and multi-scenario diagnosis of slope ecological degradation patterns; on this basis, intelligent decision-making, process optimization, visual management and control and other technologies are further coupled to construct an intelligent analysis platform and business collaboration mechanism for the entire process of ecological restoration, which significantly improves the intelligence and automation level of ecological restoration.

Description

Technical Field

[0001] The present application relates to the technical field of ecological restoration, and in particular to a method for ecological restoration of steep rock slopes with large angles. Background Art

[0002] A steep rock slope with a large angle refers to a mountain slope with a slope greater than 45°, exposed rocks on the slope surface, and sparse vegetation. This type of slope ecosystem is extremely fragile and prone to geological disasters such as soil erosion and landslides, which seriously threaten the safety of life and property of downstream residents. Slope ecological restoration is a process of restoring the structure and function of the damaged slope ecosystem and improving its ecological stability and environmental quality through a combination of engineering measures and biological measures. At present, the following technologies are mainly used for slope ecological restoration:

[0003] (1) Engineering measures. These mainly include the construction of retaining walls, lattice beams, anchor rods, anchor cables and other retaining structures, as well as drainage facilities such as drainage ditches and intercepting ditches. These measures can effectively prevent slope instability and landslides, but they are expensive, difficult to construct, and difficult to coordinate with the surrounding natural environment.

[0004] (2) Biological measures. These mainly include vegetation measures such as sowing grass seeds and planting trees and shrubs, as well as auxiliary measures such as laying vegetation nets and geotextile blankets. These measures can effectively prevent soil erosion by using the roots of vegetation to consolidate the soil and the stems and leaves to intercept precipitation. However, due to the limitations of site conditions, the survival rate and growth of vegetation are often unsatisfactory.

[0005] (3) Ecological restoration materials. These mainly include chemical materials such as soil conditioners, water retainers, and curing agents, as well as biological materials such as plant fiber blankets and degradable mulch films. These materials can improve the site conditions of the slope and create a favorable environment for vegetation growth, but they are relatively expensive and their long-term ecological effects need further demonstration.

[0006] (4) Micro-topography remodeling. This mainly includes micro-topography treatment measures such as slope shaping, stepped steps, and fish-scale pits. These measures can increase slope roughness, reduce surface runoff velocity, and promote soil and water conservation, but their scope of application is limited and may cause new ecological problems.

[0007] Although the above-mentioned slope ecological restoration technologies have achieved good application results to a certain extent, the following problems still exist:

[0008] (1) Lack of systematicness and specificity. Most existing technologies are based on experience and simple models, rarely considering the integrity and complexity of slope ecosystems, and are unable to meet the special needs of steep rock slopes with large angles.

[0009] (2) High cost and slow effectiveness. Engineering measures and ecological restoration materials are expensive, while biological measures often take a long time to achieve ecological benefits. These factors restrict the large-scale application of slope ecological restoration.

[0010] (3) Poor sustainability. Some chemical materials and artificial structures may have a negative impact on the environment, making it difficult to ensure the long-term effect of slope ecological restoration.

[0011] (4) Low level of intelligence. Existing technologies mainly rely on manual experience and simple models, lack data-driven intelligent decision-making capabilities, and cannot adapt to complex and changing slope environmental conditions.

[0012] In response to the above problems, the present invention proposes a method for ecological restoration of steep rock slopes with large angles. It automatically learns the ecological evolution law of slopes from massive monitoring data and intelligently generates the optimal restoration strategy. Specifically, it intends to solve the following technical problems:

[0013] (1) Mining the long-term dependency of slope ecological time series data and accurately predicting the trend of slope ecological evolution;

[0014] (2) Modeling the interaction between slope spatial structure and ecological processes to improve the generalization and robustness of the model;

[0015] (3) Propose a mechanism for the interactive fusion of spatiotemporal features to make full use of the spatiotemporal correlation information of slope ecosystems and improve the pertinence and effectiveness of restoration strategies;

[0016] (4) Develop an end-to-end deep learning ecological restoration framework to achieve full process automation from data input to strategy output, thereby improving the intelligence level of slope ecological restoration.

[0017] By solving the above-mentioned technical problems, the method proposed in the present invention is expected to significantly improve the efficiency and effect of ecological restoration of steep rock slopes with large angles. Summary of the invention

[0018] Based on the above problems, the present invention proposes a method for ecological restoration of steep rock slopes with large angles, which comprises the following steps:

[0019] Step 1: Multi-source data fusion to construct a large-angle steep rock slope ecological monitoring dataset:

[0020] Step 1.1 Use multiple platforms and multiple sensors to obtain high-resolution remote sensing images, three-dimensional point clouds, and multi-source heterogeneous data of geological radar images of steep rock slopes with large angles:

[0021] Use drones equipped with high-resolution visible light cameras and multi-spectral cameras to take aerial photos of the study area and obtain high-definition images; use ground-based three-dimensional laser scanners to obtain three-dimensional point cloud data of the study area; use geological radar to detect the internal geological structure of the study area; and deploy multi-parameter automatic monitoring equipment in the study area to continuously collect meteorological, soil, vegetation and other data.

[0022] Step 1.2 Preprocess the multi-source heterogeneous ecological monitoring data: perform quality control on the collected multi-source heterogeneous data, remove outliers and noise data, and perform spatiotemporal registration to unify the spatiotemporal benchmark; use digital photogrammetry technology, combined with high-resolution images and three-dimensional point cloud data, to generate digital surface models (DSM) and true orthogonal images (DOM) to obtain detailed terrain information; perform terrain attribute analysis on DSM to extract micro-topographic features such as slope gradient, slope aspect, and curvature.

[0023] Step 1.3 Extract slope ecological attributes:

[0024] Extract vegetation information from DOM and calculate the normalized vegetation index ( ), Enhanced Vegetation Index ( ) and other vegetation coverage indicators; interpolation analysis is performed on soil monitoring data to generate spatial distribution maps of soil moisture content, nutrients and other properties; statistical analysis is performed on meteorological monitoring data to calculate climate factors such as rainfall erosion and dryness.

[0025] Step 1.4: Construct the ecological status index:

[0026] Constructing a comprehensive ecological status index ( ), the calculation formula is:

[0027] ;

[0028] in, Indicates Ecological attribute indicators, Indicates The weight coefficient of each indicator is is the total number of indicators. The value range is 0-1, and the larger the value, the healthier the ecological situation.

[0029] Step 1.5 Generate sample data:

[0030] The study area was divided into OK A grid of columns, with properties for each grid including spatial coordinates , geological properties, Time series data, etc., to form model training and test samples.

[0031] Step 2: Mining the ecological time series evolution law of steep rock slopes with large angles based on bidirectional long short-term memory network:

[0032] Step 2.1 Forward calculation:

[0033] For The grid is Ecological status data of time steps , using the forward Extract its hidden state :

[0034] ;

[0035] in, is the forward hidden state of the previous time step, is the total time step.

[0036] Step 2.2 Backward calculation:

[0037] For The grid is Ecological status data of time steps , using the backward Extract its hidden state :

[0038] ;

[0039] in, is the backward hidden state at the next time step.

[0040] Step 2.3 Merge the bidirectional hidden states:

[0041] The forward hidden state and the backward hidden state Splice to get The grid is Time series characteristics of time steps :

[0042] ;

[0043] Step 3 is based on The influence of complex terrain structure of steep rock slopes with large angles on ecological processes by deep modeling of attention mechanism:

[0044] Step 3.1 spatial coordinate embedding:

[0045] The first The spatial coordinates of the grid Embedded into high-dimensional space to obtain coordinate embedding vector , and perform position encoding to obtain the position embedding vector , adding the two together as Initial hidden state of the input :

[0046] ;

[0047] Step 3.2 Calculate the self-attention weight:

[0048] exist In each layer of The attention weight between the grid and all other grids. Tier Hidden state of the grid Multiply by the parameter matrix Get the query matrix , multiplied by the parameter matrix Get the key matrix , multiplied by the parameter matrix Get the value matrix :

[0049] ;

[0050] Then use the query matrix and the bond matrix for all grids Calculate the attention weights and transform the value matrix Perform weighted summation and get Tier Multi-head self-attention output of a grid :

[0051] ;

[0052] ;

[0053] in, is the number of self-attention heads, is the hidden state dimension, is the total number of grids, For splicing operation, is the output layer parameter matrix.

[0054] Step 3.3 Feedforward network calculation:

[0055] Output the multi-head self-attention Through two feed-forward layers, a linear transformation is first performed, and then Activation function to get the feedforward network output :

[0056] ;

[0057] in, and is the feed-forward layer parameter matrix, and is the bias term.

[0058] Step 3.4 residual connection and layer normalization:

[0059] The feedforward network output With input Add them together and then normalize them to get the first Tier The output hidden state of the grid :

[0060] ;

[0061] Step 3.5 Stacking piece:

[0062] Repeat steps 3.2-3.4 for a total of layer, get the Grid Final Output :

[0063] ;

[0064] Step 4: Fusion of spatiotemporal features:

[0065] Step 4.1: Splice spatiotemporal features:

[0066] Will Extracted time series features and Extracted spatial features Splicing to obtain spatiotemporal joint features :

[0067] ;

[0068] Step 4.2 Temporal-Spatial Attention Interaction:

[0069] Based on time series characteristics For query, spatial features Compute temporal-spatial attention for keys and values :

[0070] ;

[0071] in, are the first query matrix, the first key matrix and the first value matrix respectively, It is the joint feature dimension of time and space.

[0072] Step 4.3 Spatial-temporal attention interaction:

[0073] Based on spatial characteristics To query, time series features Compute spatial-temporal attention for keys and values :

[0074] ;

[0075] in, are the second query matrix, the second key matrix and the second value matrix respectively.

[0076] Step 4.4 Gated unit fusion:

[0077] Fusion of temporal-spatial attention using gated units and spatial-temporal attention , and obtain the final spatiotemporal interaction features :

[0078] ;

[0079] ;

[0080] in, is the fusion coefficient, for Activation function, For Hadamard, and is the gating unit parameter.

[0081] Step 5: Model training:

[0082] Step 5.1 Define the loss function:

[0083] Using the root mean square error ( ) as the loss function, measuring the difference between the model's predicted value and the true value:

[0084] ;

[0085] in, is the total number of grids, is the total time step, is the real ecological status value, is the predicted value.

[0086] Step 5.2 Exploitation Optimizer training model:

[0087] use The optimization algorithm minimizes the loss function and adaptively adjusts the learning rate. The hyperparameters are set to batch size 32, learning rate 0.001, and number of training rounds 200.

[0088] Step 5.3 Model evaluation and optimization:

[0089] Each training , evaluate the model performance on the validation set, calculate the loss value and evaluation index. If the validation loss does not decrease, the training is terminated early. During the training process, the gradient value is monitored and the gradient is clipped if it exceeds the threshold. The model parameters with the smallest validation loss are saved as the optimal model.

[0090] Step 6: Model inference:

[0091] Step 6.1 Preprocess the data to be predicted:

[0092] Preprocess the slope data to be predicted and extract Attributes such as time series features and spatial coordinates are converted into model input format.

[0093] Step 6.2 Load the model and predict:

[0094] Load the trained optimal model parameters, input the preprocessed data into the model, and obtain the ecological status prediction value through forward propagation :

[0095] ;

[0096] in, is the spatial-temporal interaction feature, and is the output layer parameter, for Activation function.

[0097] Step 6.3 Generate forecast report:

[0098] The predicted value The data are mapped back to the original scale to generate the predicted ecological status of each grid unit at different time steps. The spatiotemporal distribution characteristics of the predicted values ​​are statistically analyzed, and prediction charts are drawn to form a slope ecological status prediction report.

[0099] Step 6.4 Prediction result analysis:

[0100] Compare and analyze the predicted values ​​with the measured values ​​to evaluate the prediction accuracy and stability of the model. Analyze the changing trend of the predicted values ​​and key influencing factors to reveal the laws of slope ecological evolution. Combined with the prediction results, identify high-risk areas of ecological degradation and propose targeted measures for ecological restoration.

[0101] Step 6.5 Model update:

[0102] Regularly collect new monitoring data and iteratively train and optimize the model.

[0103] The beneficial technical effects of the present invention are as follows: and The deep fusion of spatiotemporal features can better capture the temporal evolution law and spatial correlation structure of the slope ecosystem, and improve the diagnosis and early warning capabilities of ecological degradation; the designed spatiotemporal feature interactive fusion mechanism can enhance the model's ability to identify key driving factors, which helps to understand the mechanism of slope ecological degradation; the end-to-end deep learning architecture is adopted, which reduces the dependence on artificial feature engineering and has good environmental adaptability. The present invention provides a new technical solution for the ecological restoration of steep rock slopes with large angles, which has positive significance for promoting the intelligentization and refinement of slope ecological restoration. DETAILED DESCRIPTION

[0104] This paper takes the typical steep rock slopes with large angles in a large open-pit mine as the research object, and constructs a comprehensive and accurate slope ecological monitoring data set by fusion of multi-source data. First, the study area is photographed by drones equipped with high-resolution visible light cameras and multispectral cameras to obtain high-definition images, and the digital surface model (DSM) and true orthogonal image (DOM) with centimeter-level resolution are generated by oblique photogrammetry technology to accurately express the topographic and geomorphic characteristics of the slope. Secondly, the three-dimensional point cloud data of the slope is obtained by ground-based three-dimensional laser scanners, and is fused with DSM to construct a high-precision three-dimensional model of the slope. Thirdly, the internal structure of the slope is detected by geological radar to obtain geological attribute information such as slope lithology distribution and groundwater distribution. At the same time, multi-parameter automatic monitoring equipment is deployed at different locations on the slope to continuously collect time-series dynamic data of ecological elements such as meteorology (rainfall, temperature, humidity, etc.), soil (water content, temperature, pH value, etc.), vegetation (phenology, leaf area index, etc.), etc. The time granularity is at the hourly level, and the monitoring time span is 3 years. Finally, through data collection and manual investigation, the engineering property data of the slope (slope, slope aspect, slope height, etc.) and historical disaster data are obtained.

[0105] In the data preprocessing stage, firstly, the quality of multi-source heterogeneous data is controlled, outliers and noise are removed, and spatiotemporal registration is performed to unify the spatial reference system and time base of the slope three-dimensional model. Then, the terrain attribute analysis of DSM is performed to extract micro-topographic features such as slope gradient, slope aspect, and curvature; vegetation information is extracted from DOM, and the normalized vegetation index ( ), Enhanced Vegetation Index ( ) and other vegetation coverage indicators; interpolation analysis of soil monitoring data to generate spatial distribution maps of soil moisture, nutrients and other properties; statistical analysis of meteorological monitoring data to calculate climate erosion indicators such as rainfall erosion and dryness. On this basis, a comprehensive ecological status index ( ):

[0106] (1);

[0107] in, Represents various ecological attribute indicators (such as , soil moisture content, rainfall erosivity, etc.), after normalization of maximum and minimum values; It represents the weight coefficient of each indicator, which is determined by the hierarchical analysis method; is the total number of indicators. It is a dimensionless comprehensive index with a value between 0 and 1. The larger the value, the better the ecological condition.

[0108] Finally, the slope 3D model is divided into OK Column Total grid cells, each of which has properties including the spatial coordinates of its location , geological body, Time series data, etc., form model training and test samples. It should be noted that due to the complex slope terrain and strong vegetation heterogeneity, the division of grid units needs to comprehensively consider factors such as terrain undulation and vegetation patchiness, and try to achieve homogeneous ecological attributes and moderate spatial scale.

[0109] This paper constructs an end-to-end deep learning ecological restoration model, which consists of Time series feature extraction module and The system is composed of a spatial structure modeling module, and the seamless integration of the two modules is achieved through the interactive fusion mechanism of spatiotemporal features.

[0110] The ecological evolution process of steep rock slopes with large angles shows obvious temporal continuity and cumulative effects, and it is often necessary to analyze the development trend of the slope ecosystem over a long time scale. The model has a natural advantage in processing long time series data and can use information in both the forward and reverse time directions at the same time. The model extracts the forward hidden state separately and the reverse hidden state , and then concatenate the hidden states of the two directions to form the first grid cells at time The timing characteristics of :

[0111] (2);

[0112] (3);

[0113] (4);

[0114] in, Indicates grid cells at time of value, Represents a splicing operation, is the total time step. By introducing The model can explore the temporal dependencies of slope ecological evolution more comprehensively and deeply.

[0115] Another major feature of steep rock slopes with large angles is the complex spatial structure and dramatic terrain fluctuations. The ecological attributes and evolution laws of different regions are significantly different. ) Only local spatial relationships are considered, which makes it difficult to model the integrity and connectivity of slope ecosystems. The model is based on the self-attention mechanism and can adaptively learn the dependency between any two spatial positions, which is more suitable for modeling complex terrain structures of slopes. Mapping to high-dimensional embedding vector , and superimpose sine and cosine position codes , get the initial hidden state :

[0116] (5);

[0117] Then, by stacking layer blocks to update the hidden state, each layer includes a multi-head self-attention ( ) sublayer and a feed-forward network ( ) sublayers, which are used to capture long-range spatial dependencies and nonlinear feature transformations, respectively.

[0118] The sublayer first calculates the Tier The attention weights between the unit and all units, and then the value vector Perform weighted summation to aggregate and update spatial information:

[0119] (6);

[0120] (7);

[0121] (8);

[0122] The sublayer undergoes two linear transformations and a Activation function improves the nonlinear expression ability of the model:

[0123] (9);

[0124] Finally, and The outputs of the sub-layers are added together and then normalized through the layer to get the updated hidden state :

[0125] (10);

[0126] Repeat the above process times, the final spatial feature representation of each grid cell can be obtained . Global spatial modeling is achieved through the self-attention mechanism, which can accurately characterize the spatial differentiation of slope ecological attributes.

[0127] The ecological pattern shaping and evolution process of steep rock slopes with large angles is the result of the intertwined coupling of time and space scales. Separate time series analysis or spatial modeling cannot fully explore the inherent laws of slope ecological data. In order to further enhance the semantic interaction of spatiotemporal features, this paper designs a spatiotemporal feature interaction fusion mechanism based on cross-attention. First, Extracted time series features and Extracted spatial features Splicing to form spatiotemporal joint features :

[0128] (11);

[0129] Then, through temporal-spatial attention and spatial-temporal attention The cross-attention calculation in two directions realizes the mutual influence and correction of spatiotemporal features. For example, the time series features Mapping to query matrix , the spatial features Mapping to key matrix Sum Matrix , calculate the attention weight by scaling the dot product, and then The weighted summation is performed to obtain the temporal-spatial interaction features.

[0130] (12);

[0131] in, , is the dimension of the spatiotemporal joint feature. Similarly, we can get the spatial-temporal attention Finally, the interactive features in the two directions are adaptively fused through the gating unit:

[0132] (13);

[0133] (14);

[0134] in, It is a fusion gate. for Activation function, represents the Hadamard product, and is the parameter of the gating unit. Through the gating unit, the importance of the interactive features in the two directions can be adaptively adjusted, and the complementarity of spatiotemporal features can be fully utilized to form a more accurate and comprehensive spatiotemporal representation. .

[0135] This paper adopts an end-to-end approach to train the slope ecological restoration model, with the goal of minimizing the error between the model prediction value and the actual restoration effect. We choose the mean square error ( ) as the loss function:

[0136] ;

[0137] in, Indicates grid cells at time The real ecological status score is obtained through expert scoring or reference plot evaluation; Represents the prediction score of the model, through the spatiotemporal interaction features calculate:

[0138] (15);

[0139] are the parameters of the output layer, for The activation function maps the output value to between 0 and 1, indicating the quality of the ecological situation. During the training process, we use The optimizer is used to minimize the loss function, and the learning rate is set to 0.001 and the batch size is set to 32. At the same time, gradient clipping (threshold is set to 5) is used to prevent gradient explosion, and an early stopping strategy (tolerance is set to 10 times) is used to prevent overfitting. The model is iterated 200 times on the training set. , each After completion, the model performance is evaluated on the validation set, and the model parameters with the smallest validation loss are saved.

[0140] Using the trained slope ecological restoration model, we can predict the future ecological evolution trend of any steep rock slope with a large angle and formulate the optimal restoration plan accordingly. Given a slope to be restored, we first preprocess the data to obtain the Then, input the time series data into Module, input spatial data Module, after the interaction of time and space, the final time and space representation features are obtained Finally, Mapped into ecological status prediction values ​​through a fully connected layer :

[0141] (16);

[0142] By setting different future time steps , the ecological evolution trend of the slope at different time scales can be obtained. According to the prediction results, the high-risk areas of slope ecological degradation are identified, and targeted restoration measures are formulated, including engineering measures (such as slope reduction and load reduction, anchor reinforcement, etc.) and biological measures (such as vegetation restoration, soil improvement, etc.), to optimize the spatial allocation of restoration resources. It is worth noting that due to the high nonlinearity and uncertainty of the slope ecosystem, the model prediction results can only be used as a reference for auxiliary decision-making. The formulation of the actual restoration plan also needs to fully consider engineering, economic, management and other factors, and dynamically adjust and continuously monitor to form a closed-loop optimization.

[0143] In order to comprehensively evaluate the performance of the intelligent analysis model for ecological restoration of steep rock slopes with large angles proposed in this paper, we selected the following three representative existing methods for comparison: (1) Empirical ecological restoration methods ( ), mainly based on expert experience and historical cases, to formulate general measures and strategies for slope ecological restoration; (2) process-based ecological restoration methods ( ), focusing on the physical process of slope ecological degradation, simulating slope ecological evolution by building process models such as soil erosion and vegetation degradation, and guiding restoration decisions; (3) Ecological restoration methods based on machine learning ( ), using traditional machine learning models such as support vector machines and random forests to explore the statistical correlation between ecological factors and ecological degradation and predict the risk of slope ecological degradation.

[0144] We selected 10 representative steep rock slopes with large angles in the study area as experimental samples, and compared and analyzed the performance of each model in the following five aspects: ecological degradation diagnosis accuracy, ecological degradation early warning timeliness, key driving factor identification, ecological restoration measures interpretability, and model generalization ability. At the same time, in order to ensure the reliability of the results, we adopted a 5-fold cross-validation method, repeated the experiment 5 times, and took the average value as the final evaluation indicator.

[0145] The accuracy of ecological degradation diagnosis measures the ability of the model to distinguish the ecological degradation state of the slope, and the F1 value is used as the evaluation index. The F1 value is the harmonic mean of the precision and recall rate, which comprehensively considers the precision and recall rate of the model. The larger the value, the better the model. The experimental results are as follows:

[0146] ;

[0147] It can be seen that the F1 value of the proposed model reaches 0.91, which is significantly higher than that of the other three models. This is mainly due to the stronger feature extraction and expression capabilities of the deep learning model, which can fully explore the internal connection of slope ecological factors and accurately describe the state of ecological degradation. and None of them considered the complex interactions among ecological factors and relied too much on expert experience and theoretical assumptions, making it difficult to adapt to the diversity and uncertainty of slope ecosystems, resulting in low diagnostic accuracy. Although the nonlinear relationships among ecological factors have been considered to some extent, the spatiotemporal dependency information has not been fully utilized, and the spatiotemporal modeling capabilities are limited.

[0148] The timeliness of ecological degradation early warning measures the ability of the model to warn of ecological degradation risks in advance, using the early warning time ( ) as the evaluation index. It refers to the difference between the time when the model warns of ecological degradation and the time when ecological degradation actually occurs. The larger the value, the better the timeliness of the warning. The experimental results are as follows:

[0149] ;

[0150] It can be seen that the warning time of the model of the present invention can reach 9.3, which is much higher than other models. With the introduction of the network, the model can fully learn the long-term temporal characteristics of ecological elements and possess the ability of "long-term memory", thus being more sensitive to the "danger signals" of ecological degradation. and Most of them are based on "static" data modeling, lack the description of the dynamic evolution trends of ecological elements, and the warning timeliness is insufficient. It is essentially a delayed response with the worst warning effectiveness.

[0151] The identification of key driving factors measures the ability of the model to reveal the formation mechanism of ecological degradation, using the proportion of variance explained by key factors ( ) as the evaluation index. It refers to the proportion of the variance explained by the key driving factors identified by the model for ecological degradation to the total variance. The higher the value, the deeper the model's understanding of the mechanism of ecological degradation. The experimental results are as follows:

[0152] ;

[0153] It can be seen that the model of the present invention The key factors identified can explain 86.3% of the variance of ecological degradation, revealing the main driving mechanism of slope ecological degradation. The self-attention mechanism deeply models the spatial correlation of slope ecological factors and adaptively extracts the key factors that play a decisive role in degradation from the complex ecological factors. It can also identify key factors, but is limited by the model’s expressive power and lacks the ability to depict complex interactions. Although it can reveal some driving mechanisms, some important factors are easily overlooked due to the limitations of theoretical basis. It is difficult to quantitatively analyze the effects of key factors.

[0154] The interpretability of ecological restoration measures is measured by the model to assess the rationality and feasibility of the restoration measures proposed by the restoration measures expert scoring ( ) as the evaluation index. Experts in the field of ecological restoration were invited to score the restoration measures generated by each model. The scores ranged from 0 to 5. The higher the score, the more scientific and reasonable the restoration measures. The experimental results are as follows:

[0155] ;

[0156] It can be seen that the model of the present invention The score is 4.6, and the repair measures are the most reasonable and credible. Experts generally believe that the model of the present invention can propose targeted repair measures based on diagnosing the causes of degradation and predicting degradation trends, taking into account both scientificity and feasibility. This is mainly due to the fact that the model of the present invention can make full use of the results of diagnosis and early warning, and combine historical repair cases, and optimize the repair plan based on the degradation mechanism through transfer learning and other methods. Other models either only focus on degradation diagnosis and ignore repair feedback, or the repair plan is out of touch with the diagnosis results, and lacks scientificity and pertinence.

[0157] The generalization ability of the model measures the ability of the model to adapt to different environments and different working conditions. The generalization error of the model is used. ) as the evaluation index. is the mean of the prediction error of the model on different slope samples. The smaller the error, the stronger the generalization ability of the model. The experimental results are as follows:

[0158] ;

[0159] It can be seen that the generalization error of the model of the present invention is the smallest, which is only 0.12, indicating that the model can better adapt to the task of slope ecological restoration under different environmental conditions. The reason for this is that, on the one hand, thanks to the powerful feature extraction ability of the deep learning model, it can adaptively learn the ecological degradation patterns of different slopes; on the other hand, through the expansion of training samples and regularization technology, the robustness and environmental adaptability of the model are improved. In contrast, Relying on experience summary, lacking environmental adaptability, and having the worst generalization performance; Although it can take environmental heterogeneity into account, it is limited in its adaptability to the environment due to theoretical simplification and parameter calibration. Although the generalization of the model has been improved through feature engineering, the environmental adaptability of the model has not been enhanced from a mechanistic perspective.

[0160] In summary, through experimental comparisons on a large number of real samples, it is fully demonstrated that the model of the present invention has significant advantages in terms of ecological degradation diagnosis accuracy, early warning timeliness, driving mechanism analysis, restoration plan optimization, and environmental adaptability. On the one hand, the model of the present invention can fully utilize spatiotemporal characteristics, accurately characterize the degradation process, early warning of degradation risks, explore the formation mechanism, and optimize restoration measures; on the other hand, the model of the present invention reduces the dependence on artificial feature engineering, has the generalization ability to adapt to environmental changes, and is easier to implement in engineering. At present, slope ecological restoration urgently needs to shift from traditional experience-driven to data-driven, from static analysis to dynamic prediction, from single-disciplinary research to collaborative cross-disciplinary, and from passive restoration to active prevention and control. The outstanding performance of the model of the present invention is precisely due to the unique advantages of deep learning models in massive heterogeneous data processing, complex spatiotemporal feature extraction, multimodal feature fusion, end-to-end modeling optimization, etc., representing the development direction of slope ecological restoration models.

Claims

1. A method for ecological restoration of steep rock slopes with large angles, characterized in that: The steps include: Step 1: Multi-source data fusion to construct a large-angle steep rock slope ecological monitoring data set, including: Step 1.1 Use multiple platforms and multiple sensors to obtain high-resolution remote sensing images, three-dimensional point clouds, and multi-source heterogeneous data of geological radar images of steep rock slopes with large angles: Use drones equipped with high-resolution visible light cameras and multispectral cameras to take aerial photos of the study area and obtain high-definition images; use ground-based 3D laser scanners to obtain 3D point cloud data of the study area; use geological radar to detect the internal geological structure of the study area; deploy multi-parameter automatic monitoring equipment in the study area to continuously collect meteorological, soil, and vegetation data; Step 1.2: Preprocess the multi-source heterogeneous ecological monitoring data: Conduct quality control on the collected multi-source heterogeneous data, remove outliers and noise data, and perform spatiotemporal registration to unify the spatiotemporal benchmark; use digital photogrammetry technology, combine high-resolution images and three-dimensional point cloud data, generate digital surface models and true orthogonal images, and obtain detailed terrain information; perform terrain attribute analysis on DSM to extract micro-topographic features such as slope gradient, slope aspect, and curvature; Step 1.3 Extract slope ecological attributes: Extract vegetation information from DOM and calculate vegetation coverage index of normalized vegetation index and enhanced vegetation index; perform interpolation analysis on soil monitoring data to generate spatial distribution maps of soil moisture content and nutrient properties; perform statistical analysis on meteorological monitoring data to calculate rainfall erosivity and dryness climate factors; Step 2 is to mine the ecological temporal evolution law of steep rock slopes with large angles based on the bidirectional long short-term memory network, which includes: Step 2.1 Forward calculation: For The grid is Ecological status data of time steps , using the forward Extract its hidden state : ； in, is the forward hidden state of the previous time step, is the total time step; Step 2.2 Backward calculation: For The grid is Ecological status data of time steps , using the backward Extract its hidden state : ； in, is the backward hidden state of the next time step; Step 2.3 Merge the bidirectional hidden states: The forward hidden state and the backward hidden state Splice to get The grid in Time series characteristics of time steps : ； Step 3 is based on The attention mechanism deeply models the impact of complex terrain structure of steep rock slopes with large angles on ecological processes, including: Step 3.1 spatial coordinate embedding: The first The spatial coordinates of the grid Embed into high-dimensional space to obtain coordinate embedding vector , and perform position encoding to obtain the position embedding vector , adding the two together as Initial hidden state of the input : ； Step 3.2 Calculate the self-attention weight: exist In each layer of The attention weight between the grid and all other grids; Tier Hidden state of the grid Multiply by the parameter matrix Get the query matrix , multiplied by the parameter matrix Get the key matrix , multiplied by the parameter matrix Get the value matrix : ； Then use the query matrix and the bond matrix for all grids Calculate the attention weights and transform the value matrix Perform weighted summation and get Tier Multi-head self-attention output of a grid : ； ； in, is the number of self-attention heads, is the hidden state dimension, is the total number of grids, For splicing operation, is the output layer parameter matrix; Step 3.3 Feedforward network calculation: Output the multi-head self-attention Through two feed-forward layers, a linear transformation is first performed, and then Activation function to get the feedforward network output : ； in, represents the input of the feedforward network, and is the feed-forward layer parameter matrix, and is the bias term; Step 3.4 residual connection and layer normalization: The feedforward network output With input Add them together and then normalize them to get the first Tier The output hidden state of the grid : ； Step 3.5 Stacking piece: Repeat steps 3.2-3.4 for a total of Second, get the Grid Final Output ; Step 4 integrates spatiotemporal features, including: Step 4.1: Splice spatiotemporal features: Will Extracted time series features and Extracted spatial features Splicing to obtain spatiotemporal joint features : ； Step 4.2 Temporal-Spatial Attention Interaction: Based on time series characteristics For query, spatial features Compute temporal-spatial attention for keys and values : ； in, are query matrix, key matrix and value matrix respectively, is the spatiotemporal joint feature dimension; Step 4.3 Spatial-temporal attention interaction: Based on spatial characteristics To query, time series features Compute spatial-temporal attention for keys and values : ； in, They are query matrix, key matrix and value matrix respectively; Step 4.4 Gated unit fusion: Fusion of temporal-spatial attention using gated units and spatial-temporal attention , and obtain the final spatiotemporal interaction features : ； ； in, is the fusion coefficient, for Activation function, For Hadamard, and is the gating unit parameter; Step 5: Model training; Step 6: Model reasoning, including: Step 6.1 Preprocess the data to be predicted: Preprocess the slope data to be predicted and extract The time series features and spatial coordinate attributes are converted into model input format; Step 6.2 Load the model and predict: Load the trained optimal model parameters, input the preprocessed data into the model, and obtain the ecological status prediction value through forward propagation : ； in, is the spatial-temporal interaction feature, and is the output layer parameter, for Activation function; Step 6.3 Generate forecast report: The predicted value Reverse mapping back to the original scale to generate the predicted ecological status of each grid unit at different time steps; statistically analyze the spatiotemporal distribution characteristics of the predicted values, draw prediction charts, and form a slope ecological status prediction report; Step 6.4 Prediction result analysis: Compare and analyze the predicted values ​​with the measured values ​​to evaluate the prediction accuracy and stability of the model; analyze the changing trends and key influencing factors of the predicted values ​​to reveal the laws of slope ecological evolution; identify high-risk areas of ecological degradation based on the prediction results and propose targeted measures for ecological restoration; Step 6.5 Model update: Regularly collect new monitoring data and iteratively train and optimize the model.

2. The method according to claim 1, characterized in that: The step 5 specifically includes: Step 5.1 Define the loss function: The root mean square error is used as the loss function to measure the difference between the model prediction value and the true value: ， in, is the total number of grids, is the total time step, is the real ecological status value, is the predicted value; Step 5.2 Exploitation Optimizer training model: use The optimization algorithm minimizes the loss function and adaptively adjusts the learning rate; hyperparameters are set, the batch size is 32, the learning rate is 0.001, and the number of training rounds is 200; Step 5.3 Model evaluation and optimization: Each training , evaluate the model performance on the validation set, calculate the loss value and evaluation index; if 10 consecutive If the validation loss does not decrease, the training is terminated early. During the training process, the gradient value is monitored, and if the gradient exceeds the threshold, the gradient is clipped. The model parameters with the smallest validation loss are saved as the optimal model.

Images