Ecological restoration method and system for karst vulnerable areas based on metrological characteristics

Abstract

The present application relates to the field of data processing technology, and discloses a method and system for ecological restoration of karst fragile areas based on stoichiometric characteristics. The method comprises: detecting the carbon, nitrogen, phosphorus, potassium, calcium and magnesium contents of karst plants through a multispectral sensor, obtaining a stoichiometric characteristic matrix through standardization, inputting a deep convolutional neural network to identify plant conservative and resource acquisition adaptation strategies, quantitatively analyzing the weights of microtopography and microhabitat environmental impacts, optimizing the ratio of limiting nutrient factors based on the calcium and magnesium regulation mechanism, generating differentiated nutrient regulation schemes in combination with adaptation strategies, and forming an implementation plan for precise ecological restoration of karst fragile areas. The present application solves the technical problems in the ecological restoration of karst fragile areas, such as the inability to accurately identify nutrient limiting factors based on plant stoichiometric characteristics, the lack of quantitative analysis capabilities for environmental heterogeneity, and the difficulty in achieving personalized nutrient ratio optimization.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for ecological restoration of karst fragile areas based on metrological characteristics. Background Art

[0002] Karst landform is an important geological landscape type in the world. Its unique carbonate rock geological structure has formed a complex and diverse topography and ecological environment. Under such special geological conditions, vegetation growth faces severe challenges such as poor soil, severe nutrient loss, and high pH value. Traditional ecological restoration methods mainly rely on empirical plant selection and extensive nutrient supplementation, improving soil conditions by spreading chemical or organic fertilizers over large areas, and adopting conventional vegetation configuration patterns for ecological reconstruction. These methods have achieved certain results in ecological restoration in plain areas and provided basic technical support for ecosystem restoration.

[0003] However, existing ecological restoration technologies have exposed obvious limitations and shortcomings in their application in karst fragile areas. First, there is a lack of in-depth analysis of plant stoichiometric characteristics, which makes it impossible to accurately identify the actual nutrient requirements and limiting factors of different plants under specific environmental conditions, resulting in blind and inefficient nutrient supplementation. Secondly, the impact of the spatial heterogeneity of microtopography and microhabitats on plant adaptation strategies is ignored, and the use of a unified restoration model is difficult to adapt to the complex and changeable environmental conditions in the karst area. Thirdly, there is a lack of understanding of the action mechanisms of key regulatory elements such as Ca and Mg, which makes it impossible to fully utilize the rich calcium and magnesium resources in the karst area for precise regulation. Finally, there is a lack of intelligent decision-making support systems, which makes it impossible to achieve dynamic optimization and precise implementation of restoration plans.

[0004] Given the shortcomings of existing technologies and the actual needs of ecological restoration in fragile karst areas, there is an urgent need to address the technical issues of how to accurately identify plant stoichiometric characteristics and determine limiting nutrient factors, how to formulate differentiated nutrient regulation plans based on plant adaptation strategies and the degree of influence of environmental factors, and how to establish a precise ecological restoration decision-making system that integrates multi-source data and intelligent algorithms. Traditional restoration methods are unable to accurately quantify the coupling relationship between plant nutrient requirements and environmental conditions, making it difficult to achieve personalized nutrient ratios based on stoichiometric characteristics. Existing technologies lack the ability to quantitatively analyze the spatial heterogeneity of microtopography and microhabitats, and are unable to adjust restoration strategies based on environmental heterogeneity. Current ecological restoration lacks intelligent data processing and decision-making optimization mechanisms, making it difficult to generate precise restoration plans that coordinate multiple factors. Summary of the Invention

[0005] The present application provides a method and system for ecological restoration of karst fragile areas based on stoichiometric characteristics, which solves the technical problems in ecological restoration of karst fragile areas, such as the inability to accurately identify nutrient limiting factors based on plant stoichiometric characteristics, the lack of quantitative analysis capabilities for environmental heterogeneity, and the difficulty in achieving personalized nutrient ratio optimization.

[0006] In the first aspect, the present application provides an ecological restoration method for karst fragile areas based on stoichiometric characteristics, and the ecological restoration method for karst fragile areas based on stoichiometric characteristics includes: detecting nutrient elements of plants in karst fragile areas through multispectral sensors to obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and obtaining a plant stoichiometric characteristic matrix through standardized processing; inputting the plant stoichiometric characteristic matrix into a deep convolutional neural network for pattern recognition processing, outputting conservative strategy labels and resource acquisition strategy labels, and obtaining plant adaptation strategy classification results; performing weighted quantitative analysis of environmental influencing factors based on microtopography parameters and microhabitat data, and obtaining microtopography influence weight values ​​and microhabitat influence weight values ​​through calculation; optimizing the ratio calculation of limiting nutrient factors based on the calcium and magnesium element regulation mechanism, and generating targeted nutrient regulation plans based on the plant adaptation strategy classification results to obtain an implementation plan for ecological restoration in karst fragile areas.

[0007] In a second aspect, the present application provides a karst fragile area ecological restoration system based on metrological characteristics, the karst fragile area ecological restoration system based on metrological characteristics comprising:

[0008] The detection module is used to detect nutrient elements in plants in karst fragile areas using a multispectral sensor, obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and obtain a plant stoichiometric characteristic matrix through standardization;

[0009] The recognition module is used to input the plant stoichiometric feature matrix into a deep convolutional neural network for pattern recognition processing, output conservative strategy labels and resource acquisition strategy labels, and obtain plant adaptation strategy classification results;

[0010] The quantification module is used to perform weighted quantitative analysis on environmental impact factors based on microtopography parameters and microhabitat data, and obtain the microtopography impact weight value and microhabitat impact weight value through calculation;

[0011] The ratio module is used to optimize the ratio calculation of limiting nutrient factors based on the calcium and magnesium element regulation mechanism, generate targeted nutrient regulation plans based on the plant adaptation strategy classification results, and obtain an implementation plan for ecological restoration in karst fragile areas.

[0012] In the third aspect, a karst fragile area ecological restoration device based on metrological characteristics is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the karst fragile area ecological restoration device based on metrological characteristics executes the above-mentioned karst fragile area ecological restoration method based on metrological characteristics.

[0013] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned method for ecological restoration of karst fragile areas based on metrological characteristics.

[0014] In the technical solution provided by the present application, the present invention uses a multispectral sensor to detect nutrient elements in plants in karst fragile areas to obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and obtains a plant stoichiometric feature matrix through standardized processing, thereby achieving accurate quantitative characterization of the intrinsic nutrient status of the plant and solving the problem of inaccurate nutrient demand identification caused by traditional methods relying on empirical judgment. At the same time, the plant stoichiometric feature matrix is ​​input into a deep convolutional neural network for pattern recognition processing to output conservative strategy labels and resource acquisition strategy labels to obtain plant adaptation strategy classification results. The multi-scale feature extraction capability of the deep convolutional neural network and the deep learning characteristics of the residual network structure enable the algorithm to automatically identify complex patterns and potential correlations in stoichiometric data. Compared with traditional statistical analysis methods, the convolutional neural network can capture the spatial correlation between stoichiometric ratios through local receptive fields and weight sharing mechanisms, while the residual connection and batch normalization technology ensure the effective training and feature transfer of the deep network, significantly improving the accuracy and robustness of plant adaptation strategy identification. Then, the environmental influencing factors are weighted and quantified according to microtopography parameters and microhabitat data. The microtopography influence weight value and microhabitat influence weight value are obtained by calculation, and the time series modeling of the long and short-term memory network is carried out. The algorithm's weighted allocation strategy, combining the ability to predict and assess the spatial and temporal variations of different environmental factors, overcomes the limitation of traditional methods in quantitatively assessing the impact of environmental heterogeneity. Ultimately, based on the calcium and magnesium regulation mechanisms, it optimizes the ratio of limiting nutrient factors and combines this with the plant adaptation strategy classification results to generate a targeted nutrient adjustment plan, resulting in an ecological restoration implementation plan for karst vulnerable areas. The adversarial training mechanism of the generative adversarial network and the nonlinear mapping capability of the multi-layer fully connected network enable the algorithm to learn the complex regulatory relationships between calcium and magnesium and limiting nutrient factors. Through the game process between the generator and the discriminator, the algorithm continuously optimizes the rationality of the ratio plan, achieving a fundamental shift from extensive, uniform fertilization to precise, personalized nutrient regulation. The overall solution deeply integrates the pattern recognition, feature learning, and optimization decision-making capabilities of artificial intelligence algorithms with specialized knowledge in karst ecological restoration. The core contribution of the algorithmic features to the solution lies in the automatic discovery of the inherent laws of stoichiometric characteristics and adaptation strategies through deep learning, the accurate quantification of environmental factor weights through a temporal network and attention mechanism, and the intelligent optimization of nutrient ratios through the generative adversarial network, thus constructing an intelligent ecological restoration technology system based on the integration of data-driven and knowledge-guided approaches. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is a schematic diagram of an embodiment of a method for ecological restoration of karst fragile areas based on metrological characteristics in an embodiment of the present application;

[0017] Figure 2 This is a schematic diagram of an embodiment of a karst fragile area ecological restoration system based on metrological characteristics in an embodiment of the present application;

[0018] Figure 3 It is a schematic block diagram of the structure of the karst fragile area ecological restoration equipment based on measurement characteristics in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present application provide a method and system for ecological restoration of karst fragile areas based on metrological characteristics. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the karst fragile area ecological restoration method based on metrological characteristics includes:

[0021] Step S101: nutrient element detection of plants in the karst fragile area is performed using a multispectral sensor to obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and the plant stoichiometric characteristic matrix is ​​obtained through standardization processing;

[0022] Step S102: Input the plant stoichiometric feature matrix into a deep convolutional neural network for pattern recognition processing, output conservative strategy labels and resource acquisition strategy labels, and obtain plant adaptation strategy classification results;

[0023] Step S103: performing weighted quantitative analysis on environmental impact factors based on microtopography parameters and microhabitat data, and obtaining microtopography impact weight values ​​and microhabitat impact weight values ​​by calculation;

[0024] Step S104: Optimize the ratio of limiting nutrient factors according to the calcium and magnesium regulation mechanism, generate a targeted nutrient regulation plan based on the plant adaptation strategy classification results, and obtain an ecological restoration implementation plan for karst fragile areas.

[0025] It is understandable that the execution subject of this application can be a karst fragile area ecological restoration system based on measurement characteristics, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0026] Specifically, a multispectral sensor was used to obtain plant spectral reflectance data, and then these spectral data were input into the element content inversion algorithm for quantitative analysis. The algorithm calculated the original content values ​​of carbon, nitrogen, phosphorus, potassium, calcium and magnesium of a single plant based on the regression relationship between spectral bands and element contents. The stoichiometric ratios of the original element contents were then calculated, including fifteen stoichiometric ratio combinations such as carbon-nitrogen ratio, nitrogen-phosphorus ratio and calcium-magnesium ratio. These stoichiometric ratio data constituted a multidimensional data vector to form a plant individual stoichiometric data set. The data set was then subjected to outlier detection and missing value filling processing, and the processed data were reconstructed into a classified stoichiometric data matrix according to functional group categories. Finally, through zero-mean standardization and maximum-minimum value normalization processing, the sample feature vector was constructed in combination with the microtopography location label and the microhabitat type label to obtain the plant stoichiometric feature matrix.

[0027] The plant stoichiometric feature matrix is ​​fed into a deep convolutional neural network as input data. The network first performs feature extraction through a multi-scale convolutional layer, in which a three-by-three convolution kernel is responsible for extracting local stoichiometric relationship features, and a seven-by-seven convolution kernel is responsible for extracting global nutrient distribution features. The two convolution kernels perform convolution operations on the stoichiometric feature matrix respectively to produce feature maps of different scales. The multi-dimensional feature map is then weightedly fused using an attention mechanism, and the local stoichiometric relationship features and the global nutrient distribution features are merged and calculated according to the weight coefficient to obtain a fused feature vector. The fused feature vector is input into a residual network structure for deep feature learning. The input features are added to the intermediate layer features through jump connections, and batch normalization operations are applied to standardize the output of each layer. Finally, a high-level semantic feature representation is obtained. A softmax classifier is calculated based on this feature representation, and the features are mapped to conservative strategy probability values ​​and resource acquisition strategy probability values, and the plant adaptation strategy classification results are output.

[0028] A weighted quantitative analysis of microtopography parameters and microhabitat data was conducted. First, data on slope, aspect, altitude, and soil pH parameters in karst fragile areas were collected, and the collected data were constructed into a microtopography parameter matrix. At the same time, the sunny slope, shady slope, semi-shady slope, and flat land types were numerically encoded to obtain microhabitat type vectors. Then, the microtopography parameter matrix and microhabitat type vectors were input into a long short-term memory network for temporal feature learning. The network uses a forget gate to control which historical information needs to be forgotten. The input gate determines the importance of the current input information, and the output gate controls the flow of output information. The three gating mechanisms work together to extract the temporal feature sequence of environmental factors. Then, a bidirectional loop calculation is performed on the temporal feature sequence. The forward hidden state and the backward hidden state are weightedly summed according to the time step to obtain a bidirectional environmental impact feature representation. Finally, the attention weight allocation mechanism is applied to the feature representation to evaluate its importance. The attention score of each environmental factor is calculated using the softmax function to obtain the microtopography impact weight value and the microhabitat impact weight value.

[0029] Based on the calcium and magnesium regulation mechanism, the optimal ratio of limiting nutrient factors was calculated. First, a limiting analysis of the nitrogen, phosphorus, and potassium contents in the plant stoichiometric characteristic matrix was performed. Nutrient-limiting factors were identified using a minimum determination algorithm. This algorithm compared the relative contents of nitrogen, phosphorus, and potassium to determine the limiting factor corresponding to the minimum. Calcium and magnesium contents were also extracted as regulatory variables to form a nutrient-limiting factor dataset. This dataset was then input into a generative adversarial network for ratio optimization. The generator network calculated the optimal ratio between calcium and magnesium and the limiting nutrient factors through multiple fully connected layers. The discriminator network evaluated the authenticity of the generated solution. The two networks were trained adversarially to obtain an initial nutrient ratio solution. Strategy matching was then performed on the initial ratio solution based on the plant adaptation strategy classification results. The conservative strategy label and resource acquisition strategy label were used as constraints to adjust the ratio parameters and obtain a strategy-matched nutrient ratio solution. Finally, the ratio solution was spatially differentiated based on the microtopography and microhabitat influence weights. Specific ratio coefficients were calculated for different terrain conditions and habitat types, forming an implementation plan for ecological restoration in karst vulnerable areas.

[0030] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0031] Spectral data of plants from different functional groups in karst fragile areas were collected, and the spectral reflectance data were input into the element content inversion algorithm for quantitative analysis to obtain the original content values ​​of carbon, nitrogen, phosphorus, potassium, calcium and magnesium for each plant.

[0032] The stoichiometric ratios of the elements were calculated based on the original element content values ​​to obtain the plant individual stoichiometric data set;

[0033] The plant individual stoichiometric data sets were processed for outlier detection and missing value filling, and the processed data were reconstructed into a matrix according to functional group categories to obtain a classified stoichiometric data matrix;

[0034] The classified stoichiometric data matrix was subjected to zero-mean standardization and maximum-minimum value normalization, and the sample feature vector was constructed by combining the microtopography location label and microhabitat type label to obtain the plant stoichiometric feature matrix.

[0035] Specifically, spectral data collection for plants of different functional groups in karst fragile areas is carried out by obtaining spectral reflectance information of plant leaves within a specific band range through multispectral sensors. Plants of different functional groups refer to plant groups with different ecological functions, such as trees, shrubs and herbs. Spectral reflectance data reflects the light energy reflection characteristics of plant leaves at different wavelengths. There is a specific mathematical relationship between these reflectance values ​​and the element content inside the plant. The element content inversion algorithm is based on the regression model between spectral bands and element concentrations. The algorithm converts spectral reflectance values ​​into specific element content values ​​through a pre-established calibration equation, thereby obtaining the original content values ​​of the six key elements of carbon, nitrogen, phosphorus, potassium, calcium and magnesium for a single plant. These content values ​​are expressed in milligrams per gram of dry weight. Calculating stoichiometric ratios based on the obtained original element content values ​​involves performing ratio operations on the contents of different elements. Stoichiometric ratio refers to the relative proportional relationship between the contents of two or more elements, specifically including fifteen different element ratio combinations such as carbon-nitrogen ratio, carbon-phosphorus ratio, nitrogen-phosphorus ratio, calcium-magnesium ratio, and potassium-calcium ratio. These ratios are obtained through simple division operations. For example, the carbon-nitrogen ratio is equal to the carbon content divided by the nitrogen content. All stoichiometric ratio numerical combinations of each plant individual form a multidimensional data vector. The set of multidimensional data vectors of all plant individuals constitutes a plant individual stoichiometric dataset, which contains complete element ratio information for each plant.

[0036] Outlier detection uses statistical methods to identify values ​​in the data set that deviate from the normal range. The quartile method or standard deviation method is usually used to determine the threshold range of outliers. Data points outside this range are marked as outliers and removed from the data set. Missing value filling processing uses mean interpolation or regression interpolation methods to estimate and fill in missing data that may occur during data collection. The processed data are grouped and organized according to functional group categories such as trees, shrubs, and herbs. The stoichiometric data of each functional group are arranged in rows to form a submatrix, and the submatrices of all functional groups are merged in columns to form a classified stoichiometric data matrix. The rows of the matrix represent different plant individuals, and the columns represent different stoichiometric ratio indicators. Zero-mean normalization is to subtract the mean of each indicator from the value of the classified stoichiometric data matrix and then divide it by the standard deviation, so that the mean of each indicator becomes zero and the standard deviation is one, eliminating the differences in dimensions and numerical ranges between different indicators. Maximum and minimum normalization is to subtract the minimum value of each indicator from the value of the indicator and then divide it by the difference between the maximum and minimum values, so that all values ​​are compressed into the range of zero to one. Micro-topography location labels include coding identifiers of terrain location types such as uphill, mid-slope, downhill, and depression. Micro-habitat type labels include coding identifiers of habitat conditions such as sunny slope, shady slope, semi-shady slope, and flat land. The standardized and normalized stoichiometric data are merged with the corresponding terrain location labels and habitat type labels by column to construct a sample feature vector containing stoichiometric characteristics and environmental label information. The feature vectors of all samples are arranged by row to form a plant stoichiometric feature matrix, which contains both the intrinsic stoichiometric characteristics of the plant and information on external environmental conditions.

[0037] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0038] The plant stoichiometric feature matrix is ​​input into the multi-scale convolution layer for feature extraction. The local stoichiometric relationship features are extracted by a three-by-three convolution kernel, and the global nutrient distribution features are extracted by a seven-by-seven convolution kernel, thus obtaining a multi-dimensional feature map.

[0039] The multi-dimensional feature map is subjected to weighted fusion processing using an attention mechanism, and the local stoichiometric relationship features and the global nutrient distribution features are combined and calculated according to the weight coefficient to obtain a fused feature vector;

[0040] The fused feature vector is input into the residual network structure for deep feature learning. The gradient propagation process is optimized through skip connections and batch normalization operations to obtain high-level semantic feature representation.

[0041] The softmax classifier is calculated based on the high-level semantic feature representation, and the feature representation is mapped into the probability values ​​of the conservative strategy and the resource acquisition strategy to obtain the plant adaptation strategy classification results.

[0042] Specifically, the multi-scale convolutional layer feature extraction process is to send the plant stoichiometric feature matrix as input data into two parallel convolution branches for feature extraction at different scales. The three-by-three convolution kernel captures the local correlation between adjacent elements by performing a sliding window operation on the stoichiometric feature matrix. The convolution kernel covers a data area of ​​three rows and three columns each time, and obtains an output value by summing the dot multiplication operation between the weight matrix and the input data. This small-size convolution kernel mainly identifies fine-grained correlations between stoichiometric ratios, such as the direct correlation between the carbon-nitrogen ratio and the nitrogen-phosphorus ratio. The seven-by-seven convolution kernel covers a larger receptive field range and processes a data area of ​​seven rows and seven columns each time. It can capture a wider range of global distribution patterns and long-distance dependencies between nutrient elements, such as the complex regulatory relationship between calcium, magnesium and other elements. The two convolution kernels of different scales perform convolution operations on the stoichiometric feature matrix respectively to generate corresponding local stoichiometric relationship feature maps and global nutrient distribution feature maps. These feature maps are stacked according to the channel dimension to form a multidimensional feature map. The attention mechanism weighted fusion processing optimizes the feature combination by calculating the importance weights of different feature channels. First, a global average pooling operation is performed on the multidimensional feature map to compress the two-dimensional feature map of each feature channel into a scalar value. These scalar values ​​are then input into a two-layer fully connected network for nonlinear transformation. The first-layer fully connected network compresses the input dimension to one-sixteenth of the original. After processing by the ReLU activation function, it is input into the second-layer fully connected network to restore the original number of channels. Finally, the output value is limited to between zero and one through the sigmoid activation function as the attention weight of each channel. The local stoichiometric relationship features and the global nutrient distribution features are respectively multiplied element-by-element with the corresponding attention weights. Then, the weighted features are added according to the channel dimension to obtain a fused feature vector, which integrates important information of features at different scales.

[0043] The deep feature learning processing of the residual network structure adopts the skip connection mechanism to solve the gradient disappearance problem in deep network training. The fused feature vector is first input into the first residual block, which contains two convolutional layers. The first convolutional layer performs convolution operation on the input features to obtain intermediate features. The second convolutional layer continues to convolve the intermediate features, and then the output of the second convolutional layer is added element-by-element to the original input of the residual block through a skip connection. This skip connection allows the gradient to be directly back-propagated from the output layer to the input layer. The batch normalization operation is performed after each convolutional layer. The features are standardized by calculating the mean and variance of the current batch data, so that the input distribution of each layer remains stable. After cascading multiple residual blocks, a high-level semantic feature representation is obtained, which contains deep abstract information of plant stoichiometric characteristics.

[0044] The softmax classifier calculation process maps the high-level semantic feature representation into two output nodes through the fully connected layer, corresponding to the conservative strategy and the resource acquisition strategy respectively. The fully connected layer multiplies the feature vector by the weight matrix and adds the bias vector to obtain two original score values. The softmax function performs an exponential operation on these two score values ​​and then normalizes them so that the sum of the two output values ​​is equal to one. The specific calculation process is to take the exponent of each score value and divide it by the sum of the sum of the exponential sums of all score values ​​to obtain the probability value of the conservative strategy and the probability value of the resource acquisition strategy. The strategy category with the larger probability value is used as the final adaptation strategy classification result of the plant. The entire classification process makes decisions based on the deep learning representation of the plant's stoichiometric characteristics.

[0045] In a specific embodiment, the process of executing the step of inputting the fused feature vector into the residual network structure for deep feature learning processing may specifically include the following steps:

[0046] The fused feature vector is input into the first residual block for nonlinear transformation processing, the first intermediate feature is calculated through the convolution layer and activation function, and the first intermediate feature is skip-connected and added with the fused feature vector to obtain the first residual output feature;

[0047] Perform batch normalization on the first residual output feature, calculate the mean and variance of the feature data according to the batch dimension, and perform normalization transformation to obtain the normalized first residual feature;

[0048] The normalized first residual feature is input into the second residual block for deep feature abstraction processing. The second intermediate feature is calculated by combining multi-layer convolution and nonlinear activation function, and is skip-connected and added with the normalized first residual feature to obtain the second residual output feature.

[0049] The second residual output feature is subjected to global average pooling processing to compress the two-dimensional feature map into a one-dimensional feature vector, and the high-level semantic feature representation is obtained by calculating the average value of the spatial dimension.

[0050] Specifically, the nonlinear transformation processing of the first residual block passes the fused feature vector through two consecutive convolutional layers for feature learning. The first convolutional layer uses a three-by-three convolution kernel to perform a convolution operation on the fused feature vector, and performs a dot product summation operation with the weight parameter and the local area of ​​the input feature to generate a preliminary nonlinear feature representation. Then, the negative values ​​are set to zero and the positive values ​​are retained through the ReLU activation function processing to enhance the nonlinear expression ability of the feature. The second convolutional layer continues to convolve the activated features and uses the convolution kernel of the same size to further extract deeper feature information. The series processing of the two convolutional layers obtains the first intermediate feature, which contains the abstracted information after two nonlinear transformations. The jump connection addition operation adds the first intermediate feature to the original fused feature vector element by element. This direct addition mechanism ensures that the original information will not be lost during the deep network propagation process, and allows the network to learn residual mapping rather than complete mapping. After addition, the first residual output feature is obtained, which contains both new features that have undergone nonlinear transformation and retains the original feature information of the input.

[0051] Batch normalization performs statistical standardization on the first residual output feature. First, the mean of all samples in the current batch in each feature dimension is calculated. The mean is obtained by adding the eigenvalues ​​of all samples in this dimension and dividing it by the number of samples. Then, the variance of the eigenvalue of this dimension is calculated. The variance value is obtained by averaging the square of the difference between the eigenvalue of each sample and the mean. The standardization transformation subtracts the mean of the corresponding dimension from each eigenvalue and divides it by the standard deviation. The standard deviation is equal to the square root of the variance. This standardization operation makes the data distribution of each feature dimension have a mean of zero and a variance of one, eliminating the difference in the numerical range of different feature dimensions. Batch normalization also introduces learnable scaling parameters and translation parameters, allowing the network to further adjust the feature distribution on the basis of standardization to obtain the normalized first residual feature. This feature has stable statistical properties that facilitate the learning of subsequent network layers.

[0052] The deep feature abstraction processing of the second residual block uses a more complex network structure to further learn the features of the normalized first residual feature. The residual block contains three convolution layers to form a deeper feature extraction path. The first convolution layer uses a one-by-one convolution kernel to transform the dimension of the input feature. Its main function is to adjust the number of feature channels without changing the spatial size. The second convolution layer uses a three-by-three convolution kernel for spatial feature extraction to capture the spatial correlation between features. The third convolution layer uses a one-by-one convolution kernel again for dimensionality recovery. ReLU activation functions are inserted between the three convolution layers to enhance the nonlinear expression ability. The combined calculation of multi-layer convolution and activation functions enables the network to learn more abstract and complex feature representations, and obtain the second intermediate feature. This intermediate feature contains highly abstract semantic information after three nonlinear transformations. The jump connection addition operation adds the second intermediate feature to the normalized first residual feature element by element, maintaining the continuity of feature transfer, and obtaining the second residual output feature.

[0053] The global average pooling process converts the second residual output feature from a two-dimensional feature map into a one-dimensional feature vector. This processing process calculates the average value of the spatial dimension for each channel of the feature map. The specific operation is to add the values ​​of all positions in the two-dimensional feature map of each channel and divide it by the total number of positions to obtain the global average value of the channel. The global average values ​​of all channels are arranged in sequence to form a one-dimensional feature vector. This pooling operation effectively compresses the spatial information of the feature while retaining the semantic information of the channel dimension, eliminating the influence of position changes in the feature map on the classification results, and making the network invariant to the spatial transformation of the input. The high-level semantic feature representation obtained by calculating the spatial dimension average value is a compact vector containing the global information of all channels. This vector captures the final abstract representation of the plant stoichiometric characteristics after deep network learning.

[0054] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0055] Data on slope, aspect, altitude, and soil pH in karst vulnerable areas were collected, and the collected data were constructed into a microtopography parameter matrix. The sunny slope, shady slope, semi-shady slope, and flat land types were coded to obtain microhabitat type vectors.

[0056] The micro-topography parameter matrix and the microhabitat type vector are input into the long short-term memory network for temporal feature learning. The information flow is controlled by the forget gate, input gate and output gate to obtain the temporal feature sequence of environmental factors.

[0057] Based on the time series feature sequence of environmental factors, a bidirectional loop calculation process is performed, and the forward hidden state and the backward hidden state are weighted and summed according to the time step to obtain a bidirectional environmental impact feature representation;

[0058] The attention weight allocation mechanism is applied to the bidirectional environmental impact feature representation to perform importance evaluation processing. The attention score of each environmental factor is calculated through the softmax function to obtain the microtopography influence weight value and the microhabitat influence weight value.

[0059] Specifically, the data collection process uses topographic surveying equipment and soil testing instruments to obtain key environmental parameters of karst fragile areas. The slope parameter is obtained by measuring the surface inclination angle with a slope meter to obtain a value in degrees. The aspect parameter is obtained by determining the slope direction with a compass or GPS device to obtain the azimuth value. The altitude parameter is obtained by using an elevation measuring device to obtain the height above sea level. The soil acidity parameter is obtained by measuring the acidity and alkalinity of the soil solution with a pH meter to obtain the pH value. These collected values ​​are arranged in order according to the sampling point location and time to construct a micro-topography parameter matrix. The rows of the matrix represent different sampling points, and the columns represent the four parameters of slope, aspect, altitude, and soil acidity. The micro-habitat type coding process converts qualitative habitat descriptions into numerical identifiers. The sunny slope is coded as one, the shady slope is coded as two, the semi-shady slope is coded as three, and the flat land is coded as four. Each sampling point corresponds to a coded value, and the coded values ​​of all sampling points are arranged in order to form a micro-habitat type vector, which corresponds one-to-one with the micro-topography parameter matrix in terms of spatial position.

[0060] The long short-term memory network temporal feature learning process uses the micro-topography parameter matrix and the niche type vector as input sequences for temporal modeling. The network regards the environmental parameters of each sampling point as a time step in the time series. The forget gate calculates the forgetting probability through the sigmoid activation function. This probability determines which information in the cell state at the previous moment needs to be discarded. The input of the forget gate includes the current moment's environmental parameters and the previous moment's hidden state. After being transformed by the weight matrix, it outputs a forgetting coefficient between zero and one through the sigmoid function. The input gate also uses the sigmoid function to calculate the input probability and determines which part of the current input information needs to be stored in the cell state. The input gate also contains a tanh layer to create candidate update values, which performs a nonlinear transformation on the current input and the previous hidden state. The output gate controls which information in the cell state needs to be output to the hidden state. The output probability is calculated by the sigmoid function and multiplied by the tanh-activated cell state. The three gating mechanisms work together to enable the network to selectively remember and forget environmental information at different times, obtaining a temporal feature sequence of environmental factors, which contains a comprehensive representation of the environmental parameters at each time step.

[0061] The bidirectional loop calculation processing uses forward and backward recurrent neural networks to simultaneously process the time series feature sequence of environmental factors. The forward recurrent network calculates in the order of time steps from the first sampling point to the last sampling point. The hidden state of each time step is calculated based on the current input and the hidden state of the previous moment. The backward recurrent network calculates in the opposite time order from the last sampling point to the first sampling point. The hidden state of each time step is calculated based on the current input and the hidden state of the next moment. The forward hidden state and the backward hidden state capture the historical dependency and future dependency of the environmental factors respectively. The weighted summation operation is performed according to the time step to linearly combine the forward hidden state and the backward hidden state of the corresponding time step through a learnable weight coefficient. The weight coefficient is automatically learned through network training. After weighted summation, a bidirectional environmental impact feature representation is obtained, which integrates the bidirectional dependency information of environmental factors in time series.

[0062] The importance assessment process of the attention weight allocation mechanism calculates the importance weights of the influence of different environmental factors on plant adaptation strategies. First, the bidirectional environmental impact feature representation is input into the fully connected layer for linear transformation to obtain the importance score of each environmental factor. This score reflects the influence intensity of the corresponding environmental factor on plant growth and adaptation. The softmax function normalizes the importance scores of all environmental factors, and calculates the exponential value of each score through the exponential function. Then, each index value is divided by the sum of all index values ​​to obtain the attention score of each environmental factor. The numerical range of the attention score is between zero and one and the sum of all scores is equal to one. The microtopography influence weight value is obtained by taking the weighted average of the attention scores corresponding to the three microtopography parameters of slope, aspect, and altitude. The microhabitat influence weight value corresponds to the attention score of the microhabitat type parameter. The two weight values ​​quantitatively reflect the relative importance of the influence of microtopography and microhabitat on the stoichiometric characteristics of plants.

[0063] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0064] The nitrogen, phosphorus and potassium content in the plant stoichiometric characteristic matrix was subjected to restriction analysis, and the nutrient limiting factors were identified using a minimum value determination algorithm. The calcium and magnesium content were extracted as regulating variables to obtain a nutrient limiting factor dataset.

[0065] The nutrient limiting factor dataset is input into the generative adversarial network for ratio optimization. The optimal ratio relationship between calcium and magnesium and the limiting nutrient factors is calculated through the deep fully connected layer to obtain the initial nutrient ratio scheme.

[0066] Based on the results of plant adaptation strategy classification, the initial nutrient ratio scheme is subjected to strategy matching processing. The conservative strategy label and resource acquisition strategy label are used as constraints to adjust the ratio parameters and obtain a strategy matching nutrient ratio scheme.

[0067] The strategy-matching nutrient ratio scheme was spatially differentiated according to the microtopography influence weight value and the microhabitat influence weight value, and the special ratio coefficient was calculated according to different terrain conditions and habitat types to obtain the ecological restoration implementation plan for karst fragile areas.

[0068] Specifically, a restriction analysis process was performed on the nitrogen, phosphorus, and potassium content in the plant stoichiometric characteristic matrix. First, three columns of data (nitrogen, phosphorus, and potassium content) were extracted from the stoichiometric characteristic matrix. The restriction analysis process is based on the principle of Liebig's minimum factor law, which states that plant growth rate is limited by the most scarce nutrient element. The minimum value determination algorithm determines the limiting factor by comparing the relative contents of nitrogen, phosphorus, and potassium. The specific calculation process is to divide the nitrogen, phosphorus, and potassium content of each plant by the respective plant requirement threshold to obtain the relative sufficiency. The requirement threshold is a standard value determined based on the minimum nutrient concentration required for normal plant growth. The element with the lowest relative sufficiency is the limiting nutrient factor of the plant. At the same time, the calcium and magnesium contents are extracted from the stoichiometric characteristic matrix as regulating variables. Calcium and magnesium have unique regulatory effects in karst areas. Calcium is involved in cell wall construction and signal transduction processes, while magnesium is the core atom of chlorophyll molecules and participates in multiple enzymatic reactions. The identified limiting nutrient factor type, the corresponding element content value, and the calcium and magnesium content values ​​are combined to form a nutrient limiting factor dataset. This dataset contains the limiting factor identifier and related element concentration information for each plant.

[0069] The generative adversarial network (GAN) optimization process uses a dataset of nutrient limiting factors as input to calculate the optimal nutrient ratio. The GAN consists of two neural networks: a generator and a discriminator. The generator is responsible for generating nutrient ratios, while the discriminator is responsible for evaluating the rationality of the ratios. The generator first concatenates the nutrient limiting factor data with a random noise vector to expand the input dimension. It then performs a linear transformation through the first deep fully connected layer, which multiplies the input vector by a weight matrix and adds a bias vector. It then uses the ReLU activation function to introduce nonlinear characteristics. The second fully connected layer further transforms and abstracts the features. The third fully connected layer outputs the ratio coefficient of calcium and magnesium to the limiting nutrient factor. The ratio coefficient reflects the quantitative relationship between the amount of calcium and magnesium added and the amount of limiting factor missing. The discriminator network receives the generated nutrient ratios and evaluates their authenticity. Through adversarial training, the generator learns to produce more reasonable nutrient ratios, while the discriminator learns to distinguish between the true ratio and the generated ratio. The two networks interact until they reach equilibrium, resulting in an initial nutrient ratio that contains the optimal calcium and magnesium addition ratios for different limiting factors.

[0070] Strategy matching processing based on the results of plant adaptation strategy classification is to combine the initial nutrient ratio scheme with the plant survival strategy for personalized adjustment. The conservative strategy label corresponds to plants that adopt a cautious resource utilization method. Such plants usually grow slowly but have strong stress resistance. The resource acquisition strategy label corresponds to plants that actively acquire nutrient resources. Such plants grow fast but have high requirements for environmental conditions. Strategy matching processing adjusts the ratio parameters according to different strategy labels. For conservative strategy plants, the added concentration of calcium and magnesium elements is reduced to avoid physiological stress caused by excessive nutrients. For resource acquisition strategy plants, the added amount of calcium and magnesium elements is appropriately increased to meet their nutritional needs for rapid growth. The adjustment process is achieved through a preset strategy coefficient. A conservative strategy coefficient less than one is used to reduce the ratio concentration, and a resource acquisition strategy coefficient greater than one is used to increase the ratio concentration. The strategy coefficient is multiplied by the calcium and magnesium addition amount in the initial ratio scheme to obtain the adjusted ratio parameters to form a strategy-matched nutrient ratio scheme.

[0071] The spatial differentiation processing based on the microtopography influence weight value and the microhabitat influence weight value is to consider the impact of environmental heterogeneity on the nutrient ratio scheme. The microtopography influence weight value reflects the degree of influence of terrain factors such as slope, aspect and altitude on plant nutrient demand. The microhabitat influence weight value reflects the intensity of the effect of small-scale environmental factors such as light and water. The spatial differentiation processing calculates a special ratio coefficient according to different terrain conditions and habitat types. The uphill position needs to increase the amount of nutrient addition due to poor soil and water loss. The ratio coefficient is set to a value greater than one. The downhill position and depression are enriched in nutrients. If the soil and water content are sufficient, the addition amount can be appropriately reduced, and the ratio coefficient can be set to a value less than one. The sunny slope needs to enhance the calcium and magnesium regulation due to strong sunlight and large evaporation. The shady slope can reduce the regulation intensity due to the relatively mild environment. The microtopography weight value is multiplied by the terrain coefficient to obtain the terrain regulation factor, and the microhabitat weight value is multiplied by the habitat coefficient to obtain the habitat regulation factor. The final ratio scheme is calculated by multiplying the strategy matching ratio scheme by the terrain regulation factor and the habitat regulation factor, forming a differentiated nutrient regulation scheme for different spatial locations, which constitutes an implementation plan for ecological restoration in karst fragile areas.

[0072] In a specific embodiment, the step of inputting the nutrient limiting factor dataset into the generative adversarial network for ratio optimization processing may specifically include the following steps:

[0073] Perform random noise vector concatenation on the nutrient limiting factor dataset, merge the nitrogen, phosphorus and potassium limiting factor values ​​with the Gaussian noise vector by dimension, and obtain the generator input data;

[0074] The generator input data is input into the first fully connected layer for linear transformation processing, the first hidden layer output is obtained by weight matrix multiplication and bias vector addition, and the first nonlinear feature is obtained by activation function processing;

[0075] Based on the first nonlinear feature, a multi-layer fully connected network forward propagation process is performed, the feature is calculated in sequence through the second fully connected layer and the third fully connected layer, and batch normalization is applied to the output of each layer to obtain a candidate calcium-magnesium ratio scheme;

[0076] The discriminator authenticity evaluation process is performed on the candidate calcium-magnesium ratio schemes. The difference between the generated scheme and the true ratio is calculated through the adversarial loss function, and the network parameters are adjusted according to the gradient back propagation to obtain the initial nutrient ratio scheme.

[0077] Specifically, nitrogen, phosphorus and potassium limiting factor values ​​are extracted from the nutrient limiting factor dataset. These values ​​include the type identifier of the limiting factor and the corresponding element content value. The nitrogen limiting factor value indicates the relative degree of nitrogen deficiency, the phosphorus limiting factor value indicates the relative degree of phosphorus deficiency, and the potassium limiting factor value indicates the relative degree of potassium deficiency. The Gaussian noise vector is a numerical sequence randomly sampled from the standard normal distribution. The dimension of the noise vector is usually set to the same or higher dimension as the limiting factor data. The introduction of Gaussian noise enables the generator to produce diverse output results instead of a fixed mapping relationship. The nitrogen, phosphorus and potassium limiting factor values ​​are spliced ​​with the Gaussian noise vector on the feature dimension by the dimension merging operation. The specific operation is to arrange the limiting factor data vector and the noise vector in columns to form a longer vector. The length of the spliced ​​vector is equal to the original limiting factor data dimension plus the noise vector dimension. The generator input data contains the original nutrient deficiency information and random disturbance information, providing rich input features for the subsequent ratio generation process. The linear transformation processing of the first fully connected layer performs mathematical operations on the generator input data through the weight matrix and bias vector. The weight matrix is ​​the parameter matrix learned during the network training process. Its dimension is the input data length multiplied by the number of neurons in the first hidden layer. The weight matrix multiplication operation performs matrix multiplication on the input data vector and the weight matrix. Each input element is multiplied by the corresponding weight and the sum is obtained to obtain the pre-activation value of the hidden layer neuron. The bias vector addition operation adds the result of the weight matrix multiplication to the bias vector element by element. The dimension of the bias vector is equal to the number of neurons in the first hidden layer. Each bias value corresponds to a hidden layer neuron. The mathematical expression of the linear transformation is that the hidden layer output is equal to the input data multiplied by the weight matrix plus the bias vector. The output of the first hidden layer is the numerical vector after linear transformation. The activation function processing transforms the output of the first hidden layer through a nonlinear activation function. Common activation functions include ReLU function, Sigmoid function or Tanh function. The role of the activation function is to introduce nonlinear characteristics so that the network can learn complex nonlinear mapping relationships. The first nonlinear feature obtained contains the feature representation after linear transformation and nonlinear activation.

[0078] The forward propagation of the multi-layer fully connected network continues deep feature learning based on the first nonlinear feature. The second fully connected layer receives the first nonlinear feature as input, multiplies it by the second-layer weight matrix and adds the second-layer bias vector for linear transformation, and then processes it through the activation function to obtain the second-layer nonlinear output. The third fully connected layer continues to perform the same linear transformation and nonlinear activation processing on the second-layer output. The calculation process of each layer follows the same weight matrix multiplication plus bias vector addition and then activation function. Batch normalization is applied after each layer output. Batch normalization first calculates the mean and variance of the current batch data on each feature dimension, then normalizes each eigenvalue by subtracting the mean and dividing it by the standard deviation. The normalized result is then adjusted by learnable scaling and translation parameters. The role of batch normalization is to stabilize the training process and accelerate network convergence. The cascade processing of the multi-layer network enables feature abstraction and transformation layer by layer. Finally, the output of the third fully connected layer corresponds to the candidate value of the calcium-magnesium ratio. The resulting candidate calcium-magnesium ratio contains the recommended addition ratio of calcium and magnesium for the current input conditions.

[0079] Discriminator authenticity evaluation processing uses an independent discriminator network to judge the authenticity of the generated calcium-magnesium ratio candidate scheme. The discriminator network receives the ratio candidate scheme as input and calculates the authenticity probability score through a multi-layer fully connected network structure. The score reflects the rationality and credibility of the ratio scheme. The adversarial loss function calculates the degree of difference between the generated scheme and the real ratio data. The loss function usually takes the form of binary cross entropy or least squares error. The goal of the generator is to minimize the probability that the discriminator identifies the generated data as false, and the goal of the discriminator is to maximize the ability to distinguish between real data and generated data. Back propagation adjusts the network parameters by calculating the gradient of the loss function with respect to the network weights to update the parameter values. The gradient calculation uses the chain rule to propagate from the output layer to the input layer layer by layer. The weight update amount of each layer is equal to the learning rate multiplied by the corresponding gradient value. The generator and the discriminator update the parameters alternately. The generator adjusts its generation strategy according to the feedback of the discriminator, and the discriminator adjusts its discrimination ability according to the new generated data. After multiple rounds of adversarial training, the network reaches a balanced state. The generator can generate a ratio scheme close to the real distribution. The initial nutrient ratio scheme is the optimal addition ratio of calcium and magnesium elements after adversarial training optimization.

[0080] The above describes the karst fragile area ecological restoration method based on the measurement characteristics in the embodiment of the present application. The following describes the karst fragile area ecological restoration system based on the measurement characteristics in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the karst fragile area ecological restoration system based on metrological characteristics includes:

[0081] Detection module 201 is used to detect nutrient elements of plants in karst fragile areas using a multispectral sensor, obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and obtain a plant stoichiometric characteristic matrix through standardization processing;

[0082] Identification module 202, used to input the plant stoichiometric feature matrix into a deep convolutional neural network for pattern recognition processing, output conservative strategy labels and resource acquisition strategy labels, and obtain plant adaptation strategy classification results;

[0083] Quantification module 203, for performing weighted quantitative analysis on environmental impact factors based on micro-topography parameters and microhabitat data, and obtaining micro-topography impact weight values ​​and micro-habitat impact weight values ​​by calculation;

[0084] The ratio module 204 is used to optimize the ratio calculation of the limiting nutrient factors based on the calcium and magnesium element regulation mechanism, generate a targeted nutrient regulation plan based on the plant adaptation strategy classification results, and obtain an ecological restoration implementation plan for the karst fragile area.

[0085] above Figure 2 The karst fragile area ecological restoration system based on metering characteristics in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The karst fragile area ecological restoration equipment based on metering characteristics in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0086] Reference Figure 3 In the embodiment of the present invention, a karst fragile area ecological restoration device based on metering characteristics is also provided. The karst fragile area ecological restoration device based on metering characteristics can be a server, and its internal structure can be as follows: Figure 3 As shown. The karst fragile area ecological restoration equipment based on metrological characteristics includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the computer-designed processor is used to provide computing and control capabilities. The memory of the karst fragile area ecological restoration equipment based on metrological characteristics includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the karst fragile area ecological restoration equipment based on metrological characteristics is used to store the corresponding data in this embodiment. The network interface of the karst fragile area ecological restoration equipment based on metrological characteristics is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0087] Those skilled in the art will understand that Figure 3The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the karst fragile area ecological restoration equipment based on metrological characteristics to which the solution of the present invention is applied.

[0088] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of a method for ecological restoration of karst fragile areas based on metrological characteristics.

[0089] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0090] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a karst fragile area ecological restoration device based on metrological characteristics (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0091] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for ecological restoration of karst fragile areas based on metrological characteristics, characterized in that: The method comprises: The nutrient elements of plants in the karst fragile area were detected by multispectral sensors to obtain the carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and the plant stoichiometric characteristic matrix was obtained after standardization. The plant stoichiometric feature matrix is ​​input into a deep convolutional neural network for pattern recognition processing, and conservative strategy labels and resource acquisition strategy labels are output to obtain plant adaptation strategy classification results, including: inputting the plant stoichiometric feature matrix into a multi-scale convolution layer for feature extraction processing, extracting local stoichiometric relationship features through a three-by-three convolution kernel and extracting global nutrient distribution features through a seven-by-seven convolution kernel to obtain a multi-dimensional feature map; performing weighted fusion processing on the multi-dimensional feature map using an attention mechanism, merging local stoichiometric relationship features and global nutrient distribution features according to weight coefficients to obtain a fused feature vector; inputting the fused feature vector into a residual network structure for deep feature learning processing, optimizing the gradient propagation process through jump connections and batch normalization operations to obtain a high-level semantic feature representation; performing softmax classifier calculation processing based on the high-level semantic feature representation, mapping the feature representation into conservative strategy probability values ​​and resource acquisition strategy probability values, and obtaining plant adaptation strategy classification results; A weighted quantitative analysis of environmental influencing factors is conducted based on microtopography parameters and microhabitat data, and microtopography influence weight values ​​and microhabitat influence weight values ​​are obtained through calculation, including: data collection of slope, aspect, altitude and soil pH parameters in karst fragile areas, constructing the collected data into a microtopography parameter matrix, and encoding the types of sunny slopes, shady slopes, semi-shady slopes and flat land to obtain microhabitat type vectors; inputting the microtopography parameter matrix and microhabitat type vectors into a long short-term memory network for temporal feature learning processing, controlling the flow of information through forgetting gates, input gates and output gates to obtain a temporal feature sequence of environmental factors; performing bidirectional cyclic calculation processing based on the temporal feature sequence of environmental factors, performing weighted summation of the forward hidden state and the backward hidden state according to the time step, and obtaining a bidirectional environmental impact feature representation; applying an attention weight allocation mechanism to the bidirectional environmental impact feature representation for importance assessment processing, calculating the attention score of each environmental factor through the softmax function, and obtaining the microtopography influence weight value and microhabitat influence weight value; Based on the calcium and magnesium regulation mechanism, the limiting nutrient factors are optimized and calculated, and a targeted nutrient regulation plan is generated in combination with the plant adaptation strategy classification results to obtain an implementation plan for ecological restoration in karst fragile areas, including: restrictive analysis of the nitrogen, phosphorus and potassium content in the plant stoichiometric characteristic matrix, identification of nutrient limiting factors through a minimum value judgment algorithm, and extraction of calcium and magnesium content as regulation variables to obtain a nutrient limiting factor dataset; the nutrient limiting factor dataset is input into a generative adversarial network for ratio optimization, and the optimal ratio relationship between calcium and magnesium and the limiting nutrient factors is calculated through a deep fully connected layer to obtain an initial nutrient ratio plan; strategy matching is performed on the initial nutrient ratio plan based on the plant adaptation strategy classification results, and the ratio parameters are adjusted using conservative strategy labels and resource acquisition strategy labels as constraints to obtain a strategy-matched nutrient ratio plan; the strategy-matched nutrient ratio plan is spatially differentiated according to the microtopography influence weight value and the microhabitat influence weight value, and special ratio coefficients are calculated according to different terrain conditions and habitat types to obtain an implementation plan for ecological restoration in karst fragile areas.

2. The karst fragile area ecological restoration method based on metrological characteristics according to claim 1 is characterized in that: The multispectral sensor is used to detect nutrient elements in plants in the karst fragile area to obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and the plant stoichiometric characteristic matrix is ​​obtained through standardization, including: Spectral data of plants from different functional groups in karst fragile areas were collected, and the spectral reflectance data were input into the element content inversion algorithm for quantitative analysis to obtain the original content values ​​of carbon, nitrogen, phosphorus, potassium, calcium and magnesium for each plant. The stoichiometric ratios of the elements were calculated based on the original element content values ​​to obtain the plant individual stoichiometric data set; Outlier detection and missing value filling were performed on the plant individual stoichiometric data set, and the processed data were reconstructed into a matrix according to functional group categories to obtain a classified stoichiometric data matrix; The classified stoichiometric data matrix was subjected to zero-mean standardization and maximum-minimum value normalization, and the sample feature vector was constructed by combining the microtopography location label and microhabitat type label to obtain the plant stoichiometric feature matrix.

3. The method for ecological restoration of karst fragile areas based on metrological characteristics according to claim 1, characterized in that: The fused feature vector is input into the residual network structure for deep feature learning processing, and the gradient propagation process is optimized through skip connections and batch normalization operations to obtain high-level semantic feature representation, including: The fused feature vector is input into the first residual block for nonlinear transformation processing, the first intermediate feature is calculated through the convolution layer and activation function, and the first intermediate feature is skip-connected and added with the fused feature vector to obtain the first residual output feature; Perform batch normalization on the first residual output feature, calculate the mean and variance of the feature data according to the batch dimension, and perform normalization transformation to obtain the normalized first residual feature; The normalized first residual feature is input into the second residual block for deep feature abstraction processing. The second intermediate feature is calculated by combining multi-layer convolution and nonlinear activation function, and is skip-connected and added with the normalized first residual feature to obtain the second residual output feature. The second residual output feature is subjected to global average pooling processing to compress the two-dimensional feature map into a one-dimensional feature vector, and the high-level semantic feature representation is obtained by calculating the average value of the spatial dimension.

4. The method for ecological restoration of karst fragile areas based on metrological characteristics according to claim 1, characterized in that: The nutrient limiting factor dataset is input into the generative adversarial network for ratio optimization processing, and the optimal ratio relationship between calcium and magnesium elements and limiting nutrient factors is calculated through the deep fully connected layer to obtain the initial nutrient ratio scheme, including: Perform random noise vector concatenation on the nutrient limiting factor dataset, merge the nitrogen, phosphorus and potassium limiting factor values ​​with the Gaussian noise vector by dimension, and obtain the generator input data; The generator input data is input into the first fully connected layer for linear transformation processing, the first hidden layer output is obtained by weight matrix multiplication and bias vector addition, and the first nonlinear feature is obtained by activation function processing; Based on the first nonlinear feature, a multi-layer fully connected network forward propagation process is performed, the feature is calculated in sequence through the second fully connected layer and the third fully connected layer, and batch normalization is applied to the output of each layer to obtain a candidate calcium-magnesium ratio scheme; The discriminator authenticity evaluation process is performed on the candidate calcium-magnesium ratio schemes. The difference between the generated scheme and the true ratio is calculated through the adversarial loss function, and the network parameters are adjusted according to the gradient back propagation to obtain the initial nutrient ratio scheme.

5. A karst fragile area ecological restoration system based on metrological characteristics, characterized by: For implementing the karst fragile area ecological restoration method based on metrological characteristics according to any one of claims 1 to 4, the karst fragile area ecological restoration system based on metrological characteristics comprises: The detection module is used to detect nutrient elements in plants in karst fragile areas using a multispectral sensor, obtain carbon, nitrogen, phosphorus, potassium, calcium and magnesium content data, and obtain a plant stoichiometric characteristic matrix through standardization; The recognition module is used to input the plant stoichiometric feature matrix into a deep convolutional neural network for pattern recognition processing, output conservative strategy labels and resource acquisition strategy labels, and obtain plant adaptation strategy classification results, including: inputting the plant stoichiometric feature matrix into a multi-scale convolution layer for feature extraction processing, extracting local stoichiometric relationship features through a three-by-three convolution kernel and extracting global nutrient distribution features through a seven-by-seven convolution kernel to obtain a multi-dimensional feature map; performing attention mechanism weighted fusion processing on the multi-dimensional feature map, merging local stoichiometric relationship features and global nutrient distribution features according to weight coefficients to obtain a fused feature vector; inputting the fused feature vector into a residual network structure for deep feature learning processing, optimizing the gradient propagation process through jump connections and batch normalization operations to obtain a high-level semantic feature representation; performing softmax classifier calculation processing based on the high-level semantic feature representation, mapping the feature representation into conservative strategy probability values ​​and resource acquisition strategy probability values, and obtaining the plant adaptation strategy classification results; The quantification module is used to perform weighted quantitative analysis on environmental impact factors based on microtopography parameters and microhabitat data, and obtain microtopography impact weight values ​​and microhabitat impact weight values ​​through calculation, including: collecting data on slope, aspect, altitude and soil pH parameters in karst fragile areas, constructing the collected data into a microtopography parameter matrix, and encoding the sunny slope, shady slope, semi-shady slope and flat land types to obtain a microhabitat type vector; inputting the microtopography parameter matrix and microhabitat type vector into a long short-term memory network for temporal feature learning processing, controlling the information flow through forgetting gates, input gates and output gates to obtain an environmental factor temporal feature sequence; performing bidirectional cyclic calculation processing based on the environmental factor temporal feature sequence, performing weighted summation of the forward hidden state and the backward hidden state according to the time step, and obtaining a bidirectional environmental impact feature representation; applying the attention weight allocation mechanism to the bidirectional environmental impact feature representation for importance evaluation processing, calculating the attention score of each environmental factor through the softmax function, and obtaining the microtopography impact weight value and microhabitat impact weight value; The ratio module is used to optimize the ratio calculation of limiting nutrient factors based on the calcium and magnesium regulation mechanism, generate a targeted nutrient regulation plan based on the plant adaptation strategy classification results, and obtain an implementation plan for ecological restoration in karst fragile areas. The module includes: performing restrictive analysis on the nitrogen, phosphorus, and potassium content in the plant stoichiometric characteristic matrix, identifying nutrient limiting factors through a minimum value judgment algorithm, and extracting calcium and magnesium content as regulation variables to obtain a nutrient limiting factor data set; inputting the nutrient limiting factor data set into a generative adversarial network for ratio optimization processing, calculating the optimal ratio relationship between calcium and magnesium and limiting nutrient factors through a deep fully connected layer, and obtaining an initial nutrient ratio plan; performing strategy matching processing on the initial nutrient ratio plan based on the plant adaptation strategy classification results, adjusting the ratio parameters using conservative strategy labels and resource acquisition strategy labels as constraints, and obtaining a strategy-matched nutrient ratio plan; spatially differentiating the strategy-matched nutrient ratio plan based on the microtopography influence weight value and the microhabitat influence weight value, calculating special ratio coefficients according to different terrain conditions and habitat types, and obtaining an implementation plan for ecological restoration in karst fragile areas.

6. A karst fragile area ecological restoration device based on metrological characteristics, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the method for ecological restoration of karst fragile areas based on metrological characteristics as described in any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to execute the karst fragile area ecological restoration method based on metrological characteristics according to any one of claims 1 to 4.

Images